AI Model Landscape: Major Players and Model Families

The right AI provider is rarely "the one with the best model." It is the provider and model family that fits the job, data posture, integration surface, cost envelope, latency expectation, and operational controls of the system you are building.

This page maps major providers and their current model families as of October 12, 2026. Model IDs, availability, context limits, prices, and regional support change quickly; validate the linked first-party documentation before building or buying.

The selection lens

Start with the workload, then select a model tier. Do not begin with a brand preference.

At-a-glance reference

Provider	Current families to know	Common architectural fit	Deployment posture
OpenAI	GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna; GPT-5.6 Sol, Terra, Luna; GPT Image 2.5 Sunburst and Flare; GPT-Live 1, realtime and transcription models; specialist cyber and life-sciences models (gated)	Frontier reasoning, coding, agentic work, multimodal and voice applications	Hosted API and platform services; GPT-6.1 Sol also on Amazon Bedrock
Anthropic	Claude Fable 5.1; Claude Opus 5.5; Claude Sonnet 5.5; Claude Haiku 5.5; Claude Mythos 5.1 (invitation only)	Long-running agents, reasoning, coding, enterprise analysis, high-volume routing and extraction	Hosted API and cloud marketplace access
Google	Gemini 3 family (3.8 Flash current); Gemini 3.8 Live; Gemini 3.8 Flash TTS; Gemini 3.5 Transcribe; Gemini Omni Flash; Antigravity Agent; Nano Banana image models (2.1 current); Lyria 3.5; Gemma	Multimodal applications, real-time voice, managed agents, Google ecosystem, high-throughput work	Gemini API / Vertex AI; selected open models
xAI	Grok 4.7; Grok 4.6 / 4.5; Grok 4.3 and 4.20 long-context snapshots; Grok Build; Grok Imagine; Grok Voice	Reasoning, coding, real-time and creative modalities	xAI API; selected models through Amazon Bedrock and Google Enterprise Agent Platform
Meta	Muse Wonder (preview), Muse Spark, Muse Image, Muse Voice Transcribe; Muse Glimmer and SAM 3.1 (open weight); Llama 4 Scout and Maverick	Hosted agentic and coding work, custom hosting, controlled deployment, multimodal open-weight use cases	Meta Model API (hosted); open weights under Meta or Apache 2.0 licenses; multiple hosts
Mistral AI	Mistral Large 4 (preview); Mistral Large 3; Mistral Medium 3.5; Mistral Small/Ministral families	Efficient enterprise inference, European deployment options, OCR and specialist work	API plus open-weight / self-host options
Cohere	Command A+; Command A (incl. Reasoning, Vision, Translate); North; Transcribe; Embed/Rerank/Parse; Aya	Enterprise RAG, tool use, coding, document parsing, multilingual and speech workloads	API and private/VPC/on-prem options
Amazon	Nova 2 family; legacy Nova v1 generation	AWS-native multimodal, document and media workflows, governed cloud delivery	Amazon Bedrock
Microsoft	Phi; MAI and healthcare models; Model Router and partner catalog	Microsoft/Azure enterprise platform, governed multi-provider choice, smaller-model scenarios	Microsoft Foundry / Azure
DeepSeek	DeepSeek V4.1 Flash; DeepSeek V4 Pro; DeepSeek R1 family	Cost-conscious reasoning/coding and multimodal evaluation, compatible API integrations	Hosted API; open-weight ecosystem for select releases
Alibaba Cloud / Qwen	Qwen 3.8-Max and 3.8-Flash; Qwen3.7-Plus; Qwen code, audio and multimodal families	Multilingual, coding, hybrid-thinking and open-model evaluation	Qwen API / Alibaba Model Studio; open-weight ecosystem

Provider notes and model references

OpenAI

OpenAI now documents a generation above GPT-5.6. GPT-6 Astra (gpt-6-astra, released September 3, 2026) is listed as its most capable model, built for hard end-to-end work spanning reasoning, coding, computer use, research, and document creation. It carries a 1.05M-token context window with 128K max output at $10 / $50 per million input/output tokens, and input above 272K tokens bills at 2× the input rate and 1.5× the output rate — long-context work is a budget decision, not a free capability. Astra also adds agentic controls that change how an agent loop is written rather than only what it costs: async tool calling, mid-turn steering, and mid-conversation reasoning-effort adjustment.

On September 22, 2026, OpenAI filled in the rest of the GPT-6 generation beneath Astra. GPT-6 Sol (gpt-6-sol) is built for complex coding and agentic workflows at $2 / $10 per million input/output tokens, and GPT-6 Luna (gpt-6-luna) is its most efficient model for focused, high-volume tasks at $0.10 / $0.50. Both share Astra's 1.05M-token context window (922K max input), 128K max output, and the same 272K-token long-context surcharge, and both default to medium reasoning effort. There is no GPT-6 counterpart to Terra. One early operational note: on September 25, OpenAI fixed an image-encoding bug in GPT-6 Sol and Luna that had degraded image understanding, including computer use, so vision or computer-use evaluations run on either model during its first three days should be re-run.

The Sol tier then moved again a week later. On September 29, OpenAI released GPT-6.1 Sol (gpt-6.1-sol), aimed at complex coding and professional work at lower cost than Astra, at the same $2 / $10 per million input/output tokens, 1.05M-token context window, 128K max output, and 272K-token long-context surcharge as GPT-6 Sol, with multi-agent delegation in beta. The model catalog now features GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna; GPT-6 Sol has dropped out of the featured set, but its model page is still live with no deprecation notice and simply points to 6.1 Sol as the newer model. The GPT-5.6 Sol, Terra, and Luna pages also remain live with no deprecation notice and are still the named replacements for the snapshots shutting down this month. A model that leaves the catalog's front page without a deprecation date is a signal to plan an evaluation, not a deadline. GPT-6.1 Sol is also the first GPT-6-generation model on Amazon Bedrock (openai.gpt-6.1-sol, launched September 29), where it follows OpenAI's own deprecation terms rather than Bedrock's lifecycle policy.

OpenAI also added a speed tier that changes the cost model rather than the model. An Ultrafast service tier (service_tier: "ultrafast" on the Responses API) arrived for GPT-6 Astra on September 29 and for GPT-6.1 Sol on October 8, billed at six times standard rates. Its data-residency support differs by model: US only for Astra, US and EU for 6.1 Sol. Treat it as a per-request decision for latency-critical paths, not a deployment default. Other platform changes in the same window: the Agents API gained computer use in an OpenAI-hosted browser (September 29), with your application handling site-access approvals and sign-in; a Decisions API entered beta with gpt-6-luna for fast typed answers from text and image input (October 6); usage tiers were consolidated from five to three (October 6); and eligible organizations can now accept OpenAI's standard Business Associate Agreement and enable HIPAA support from organization settings (October 5).

The specialist and voice lines moved in the second week of September. On September 8, OpenAI released GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare — the most capable and the fast everyday image generation/editing options, adding xhigh and max quality settings at GPT Image 2 token rates — and made GPT-Rosalind Research generally available for life-sciences reasoning at $5 / $25 per million input/output tokens, with access limited to approved organizations and billing in effect since October 5, 2026. On September 10, GPT-Live 1 reached general availability as a full-duplex voice model on the v1/live/sessions endpoint, priced at $0.05 per minute billed per second, with backend model and tool usage charged separately — voice pricing is per-minute plus per-token, so model it as two cost lines, not one. The same day added the Agents API in public beta (a managed Codex harness with durable sessions, progress streaming, and MCP/custom tool support, where OpenAI handles session orchestration, context compaction, and recovery) and project API key expiration with organization-enforced maximum key lifetimes. On September 15 that credential work continued: organizations and projects can now restrict who may create API keys at all — service accounts only, user-owned project keys only, or no new keys — with organization restrictions taking precedence over project settings and existing keys unaffected. Prompt cache diagnostics went generally available on September 8 for GPT-5.6 and later on the Responses API, which makes cache-miss causes observable rather than inferred.

OpenAI also runs a gated cyber line that is easy to miss because it never appears in a general model list. GPT-5.6 Cyber is documented for authorized vulnerability research and security testing, with a 400K context window, and access is granted through the Daybreak program rather than by default. Its predecessor gpt-5.4-cyber shut down on October 1, 2026, so gpt-5.6-cyber is now the only documented model on this line. Pricing is $12.50 / $75 per million input/output tokens, so it is a deliberate, scoped-use model, not a general coding tier.

Platform-surface lifecycle continues to matter as much as model lifecycle. The legacy Assistants API reached its sunset on August 26, 2026 (its replacements are the Responses and Conversations APIs), and the Videos API and the Sora 2 model family were removed on September 24, 2026 with no named replacement — a useful reminder that a provider can exit a modality you built on. On August 26, 2026, OpenAI also deprecated its transcription line: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize shut down on February 26, 2027, with gpt-transcribe or gpt-live-transcribe as replacements. The gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 snapshots retired on September 28, 2026. Several dated waves follow for anyone still pinned to older IDs. A larger set of legacy snapshots shuts down October 23, 2026 — eleven days from this page's date — including gpt-3.5-turbo, the GPT-4 variants, o1, o3-mini, o4-mini, and gpt-image-1. gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest shut down December 1, 2026, with GPT Image 2.5 Sunburst or Flare as the replacements, and the GPT-5 and o3 snapshots follow on December 11, 2026. On October 1, 2026, OpenAI added two more waves. The tts-1 and tts-1-hd text-to-speech models and the dated gpt-4o-mini-tts snapshots shut down January 6, 2027, in favor of gpt-realtime-2.1-mini. gpt-5.1 and gpt-5.3-codex shut down April 1, 2027, in favor of gpt-6-sol, and gpt-5.4-nano the same day in favor of gpt-6-luna. Platform surfaces are on the clock too: reusable prompts (the v1/prompts API), the Evals dashboard and API, and Agent Builder are removed on November 30, 2026, and existing evals become read-only on October 31. If your evaluation history lives in OpenAI's Evals product, export it before then; a provider-hosted eval store is not a durable record.

Good starting point: Use a balanced model for normal product workflows; reserve the frontier tier for genuinely difficult reasoning, coding, or high-value decision support. A new top-of-range model is a reason to re-check routing, not to move everything up a tier — build model routing so a costly model is not used for simple extraction or classification.

Reference models: gpt-6-astra, gpt-6.1-sol, gpt-6-luna, gpt-6-sol (live, no longer featured), gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-image-2.5-sunburst, gpt-image-2.5-flare, gpt-live-1, gpt-transcribe and gpt-live-transcribe, GPT-Realtime family including gpt-realtime-2.1 and gpt-realtime-2.1-mini; gpt-rosalind-research and gpt-5.6-cyber where access applies. OpenAI model catalog · GPT-6 Astra model page · GPT-6.1 Sol model page · GPT-6 Sol model page · GPT-6 Luna model page · GPT-5.6 Cyber model page · OpenAI deprecations · OpenAI changelog

Anthropic

Anthropic's documented current lineup is four models, and the two lower tiers were both replaced in the past two weeks: Claude Fable 5.1 (released September 1, 2026) as the highest-capability model, for demanding reasoning and long-horizon agentic work; Claude Opus 5.5 (claude-opus-5-5, released September 22, 2026) for long-running agentic coding and knowledge work; Claude Sonnet 5.5 (claude-sonnet-5-5, released September 28, 2026) for the best combination of speed and intelligence; and Claude Haiku 5.5 (claude-haiku-5-5, released October 7, 2026) for high-volume, latency-sensitive work such as classification, extraction, and routing. All four now share a 1M-token context window and 128K max output. Opus 5.5 is priced at $4 / $20 per million input/output tokens, against Opus 5's $5 / $25, with cache reads at 5% of base input. Sonnet 5.5 keeps Sonnet 5's $2 / $10, and on October 7 its cache reads were cut to $0.10 per million tokens (5% of base input). Haiku 5.5 is priced by prompt length: $0.10 / $0.50 for prompts up to 100K tokens and $0.50 / $2.50 above that. The fivefold step at 100K tokens matters more than the headline rate if you plan to fill its context window. Anthropic's documented guidance is unchanged: start with Opus 5.5 for most workloads, and move up to Fable 5.1 only when evals on Opus 5.5 at higher effort still fall short. That is still a useful corrective to defaulting to the top of a provider's list. Fable 5, Opus 5, Opus 4.8, Sonnet 5, and Haiku 4.5 are now among the legacy-but-available models. Claude Mythos 5.1 shares Fable 5.1's capabilities, specifications, and pricing but is offered by invitation only through Project Glasswing.

Several lifecycle details are worth carrying into a design review. Fable 5.1 makes three breaking changes against Fable 5: forced tool use (tool_choice of type any or tool) returns a 400 error, thinking blocks are bound to the model that produced them so no earlier model can read Fable 5.1's, and editing anything earlier in a conversation invalidates the thinking blocks after it — enforced for accounts created on or after August 31, 2026, and opt-in for older ones. Treating conversation history as append-only is now an integration requirement rather than a style preference. Opus 5.5 brings the Opus tier onto the same rules, with four breaking changes against Opus 5. Thinking is always on, so thinking: {"type": "disabled"} or a manual budget returns a 400, and the effort parameter becomes the only control (its default is now medium, not high). Forced tool use returns a 400, as on Fable 5.1. Thinking blocks are bound to the producing model and to the conversation prefix. And on the Claude API and Google Cloud, computer use requires the computer_toolset_20260801 toolset, while the earlier computer_20251124 tool returns a 400 (Amazon Bedrock still accepts it). A fifth change fails no request but changes response shape: the short notes the model writes between tool calls now arrive as thinking blocks whose text is empty by default, so an interface that streams them as progress updates goes silent until it sets thinking.display. Anyone still moving up from Opus 4.8 hits the earlier version of this hazard: from Opus 5 onward a response can open with thinking blocks in code that assumed the first content block was text, and thinking tokens bill as output.

The new Sonnet and Haiku models carry the same integration rules down the lineup, so a model-ID swap is not a migration at any tier now. Sonnet 5.5 has five breaking changes against Sonnet 5. thinking: {"type": "disabled"} returns a 400; the lowest setting is now between_tools, accepted only at high effort or below. Forced tool use returns a 400. Thinking blocks are bound to the model and the conversation prefix. computer_20251124 is rejected on the Claude API and Google Cloud. And the advisor tool rejects Opus 4.8, Opus 4.7, Sonnet 5, and Haiku 5.5 as advisors. Haiku 5.5 is the bigger jump, because Haiku 4.5 predated most of these rules. Manual budget_tokens thinking, the temperature / top_p / top_k sampling parameters, assistant-message prefill, and computer_20250124 each return a 400. Editing earlier turns invalidates thinking blocks. On Amazon Bedrock, structured outputs are not available for it. Two quieter changes affect cost estimates on Haiku 5.5: the same text counts as roughly 30% more tokens than on Haiku 4.5, and thinking blocks from all earlier turns stay in context as input. Thinking blocks from both new models work only in the account that produced them (or a linked account), which matters if you store conversations and replay them elsewhere.

Two governance facts also carry forward. Fable 5.1 and Mythos 5.1 carry 30-day data retention, are not available under zero data retention absent express authorization from Anthropic, and include safety classifiers that can return a refusal as a successful HTTP 200 with stop_reason: "refusal" — so refusal handling and fallback belong in the integration, not in a runbook. Opus 5.5 also ships with classifiers (biology as well as cybersecurity, plus a reasoning_extraction category for requests that push it to reproduce its internal reasoning), so the same refusal handling now applies at the Opus tier. Sonnet 5.5 and Haiku 5.5 carry safety classifiers too. Sonnet 5.5 documents five stop_details categories (cyber, bio, frontier_llm, reasoning_extraction, and general_harms), and Haiku 5.5 has no server-side fallback, so every tier of the current lineup now needs client-side refusal handling. Refusals also became a cost line on September 24: Anthropic resumed billing for refusals that arrive before any output when stop_details.category is bio, frontier_llm, or reasoning_extraction, and refusals count against rate limits either way. Output from Fable 5.1 and Mythos 5.1 also carries Anthropic's statistical text watermark, with C2PA Content Credentials on generated media retrieved through the Files API, which is a provenance fact worth recording in a content policy.

On lifecycle, every current model now carries a 2027 retirement floor: Fable 5.1 not sooner than September 1, 2027, Opus 5.5 September 22, Sonnet 5.5 September 28, and Haiku 5.5 October 7. The nearer dates sit among legacy and deprecated models. On September 30, Anthropic deprecated Claude Sonnet 4.5 (claude-sonnet-4-5-20250929), with retirement on the Claude API on November 30, 2026 and claude-sonnet-5-5 as the named replacement. Claude Haiku 4.5, now legacy, has a floor of "not sooner than October 15, 2026," three days out. No retirement has been announced, and Anthropic commits to at least 60 days' notice, but Haiku 5.5 is now the obvious migration target. Amazon Bedrock runs its own schedule. Claude Sonnet 4 reaches end of life there on October 14, 2026, two days out. Sonnet 4.5 entered Legacy on Bedrock on October 8, with public extended access from January 8, 2027 and end of life on April 8, 2027. Claude Opus 4.1, retired on the Claude API on August 5, 2026, entered public extended access on Bedrock on October 8 ahead of its January 8, 2027 end of life, so expect provider-set pricing on it there.

Alongside the Opus 5.5 release, platform capabilities have continued to arrive, and they affect how a long-running agent is built. On September 22, tools became definable inside a mid-conversation system message (beta, inline-tools-2026-09-15), so an agent can add a tool or change its schema mid-session without editing tools and losing the prompt cache. On September 23, cache diagnostics came out of beta. On September 14, the Messages API added on-demand conversation compaction in beta (compact-2026-09-04): send a compaction parameter, get back a signed block summarizing earlier messages, and return that block in place of them on later requests, with recent turns preserved word-for-word and thinking still valid in those turns. That makes context budget a request-level control rather than something you hand-roll with your own summarizer. On September 10, Claude Managed Agents added an auto permission policy in which the server evaluates each agent or MCP tool call and runs it, denies it, or pauses for approval, reporting its decision in the event stream — an approval gate that lives in the platform rather than in your loop. On September 18, the Compliance API's local session endpoints began returning transcripts of Claude in Chrome sessions under the claude_in_chrome product surface, in beta for Enterprise organizations using the existing compliance access key and scope — worth knowing because browser-based assistant activity becomes retrievable evidence, which is a records-retention question as much as a capability. On September 24 the same endpoints came out of beta for Claude for Microsoft 365 sessions in Excel, PowerPoint, Word, and Outlook, while the Compliance API's Activity Feed stopped returning file names, project document names, and artifact titles. Any audit process that relied on those fields needs another source. Prompt cache reads on Fable 5.1 and Mythos 5.1 also cost 2.5% of base input price ($0.25 per MTok) rather than the usual 10%, which changes the arithmetic on cache-heavy agent designs at the top tier.

October brought more changes to Managed Agents, mostly to the controls around the agent loop. On October 7, the web_fetch tool began fetching only URLs that have already appeared in the session, for example in a user message or a search result, as a guard against data exfiltration. In environments with limited networking, allowed_hosts now also restricts web_search and web_fetch. An agent that used to fetch URLs it built itself, or URLs from its system prompt or tool output, will now get error results instead. On October 9, dynamic workflows entered beta: an agent can write a program that runs many agents in phases and combines their results, and the server runs it in the background. The Models API also now reports each model's line and whether it accepts disabled thinking and the web search and code execution tools (October 1–6), so a model picker or capability check can read those values instead of parsing model IDs.

Good starting point: Consider Claude when long-context analysis, coding workflows, tool-using agents, and safety/process documentation are central to the use case. Evaluate outputs against the work your people actually perform; do not treat a capability claim as a production acceptance test.

Reference models: claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5, claude-haiku-5-5; claude-mythos-5-1 where access and terms apply; claude-opus-5, claude-sonnet-5, and claude-haiku-4-5 as legacy-but-available. Claude models overview · Claude model selection · What's new in Claude Haiku 5.5 · What's new in Claude Sonnet 5.5 · What's new in Claude Opus 5.5 · What's new in Claude Fable 5.1 · Anthropic deprecations · Claude API release notes

Google

Google's portfolio spans Gemini for text, multimodal intelligence, and image generation (the Nano Banana models), Veo and Gemini Omni for video, Lyria for music, and Gemma for open models. The Gemini API's model guide now leads with Gemini 3.8 Flash (gemini-3.8-flash), generally available on September 2, 2026 and described as Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows. That stopped being only an option on October 8. Google deprecated Gemini 3.7 Flash (gemini-3.7-flash) and Gemini 3.5 Flash (gemini-3.5-flash) with no shutdown date. Requests to them are now automatically routed to gemini-3.8-flash and gemini-3.6-flash respectively. Nothing fails, but a pinned model string no longer pins the model serving the request, so output can change without any code change. Re-run evaluations for any workflow that was pinned to either ID. The stable Flash line is now 3.8 Flash and 3.6 Flash, with 3.5 Flash-Lite and 3.1 Flash-Lite below them. gemini-3.1-flash-lite is itself scheduled to shut down on May 7, 2027, with gemini-3.5-flash-lite as the replacement.

Real-time voice arrived in mid-September. On September 15, Gemini 3.8 Live (gemini-3.8-live) and Gemini 3.8 Live Extended Thinking (gemini-3.8-live-extended-thinking) became generally available as audio-to-audio models for real-time voice applications, the second supporting background reasoning during a conversation. A voice model that thinks mid-turn is a latency and interaction-design decision as much as a quality one; evaluate it against your own turn-taking expectations rather than a demo. Speech output followed on September 22: Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) and Gemini 3.8 Flash-Lite TTS (gemini-3.8-flash-lite-tts) became generally available with voice design, voice replication, and a library of 150+ voices. The older voice previews now have a date: gemini-3.1-flash-tts-preview, gemini-3.1-flash-live-preview, and the Gemini 2.5 native-audio and TTS previews shut down on November 17, 2026, with the 3.8 Live and 3.8 TTS models as replacements. Voice replication is also a consent and impersonation-policy question, not only a feature.

The Gemini 2.5 generation is also closing to new work. On September 18, Google began limiting access to Gemini 2.5 models to users who have actively used them in the past, and it recommends Gemini 3.5 Flash-Lite or 3.8 Flash for new projects. No shutdown date has been announced, but a new environment, account, or team may not be able to reach 2.5 at all, so a fallback that depends on it is not safe.

The newest change is on the managed-agent surface. On September 17, Google released Antigravity Agent 09-2026 (antigravity-preview-09-2026), a general-purpose agent reached in a single API call that plans, runs code, manages files, and browses the web inside a Google-hosted Linux sandbox. It is built on Gemini 3.8 Flash and defaults to it, with other Flash models selectable through agent_config. Gemini 3.7 Flash and 3.5 Flash are among them, and both are now deprecated and auto-routed, so check what that setting actually resolves to. The upgrade was not a drop-in: tool parameters moved from snake_case to PascalCase, file editing switched from full rewrites to line-range replacements, and find_by_name and grep_search were added. The version it replaced, antigravity-preview-05-2026, shut down on October 5, 2026. Managed research agents are also turning over: deep-research-pro-preview-12-2025 was deprecated on October 8 and shuts down October 23, 2026, with deep-research-preview-04-2026 and deep-research-max-preview-04-2026 as targets.

Two other changes landed earlier in the month. On September 1, agentic video understanding launched for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite across the Interactions and GenerateContent APIs, using up to 88% fewer tokens on long-form content than static processing — a cost lever for anyone already doing video analysis. On September 3, Lyria 3.5 (lyria-3.5) was released as generally available for full-length song generation with improved musical coherence and structural control; it is now Google's flagship music model. Generative media moved in early October. On October 6, Nano Banana 2.1 (gemini-nano-banana-2.1) became generally available as Google's latest high-efficiency image generation and conversational editing model, and gemini-3.1-flash-image (Nano Banana 2) was deprecated with no shutdown date announced. gemini-3.1-flash-lite-image and gemini-3-pro-image remain stable, and Imagen 4 has been shut down. The larger change is video. The Veo 3.1 preview models (veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview) are scheduled to shut down on October 22, 2026, with gemini-omni-1.1-flash as the named replacement. In the Gemini API, video generation is consolidating onto Gemini Omni rather than continuing as a separate Veo line. Anyone generating video through Veo 3.1 previews has ten days to migrate, and should treat it as a model change to re-evaluate, not an endpoint rename.

Late August moved two specialist lines from preview to general availability. On August 26, Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live became generally available as dedicated speech-to-text models, covering 85+ languages with speaker diarization, word-level timestamps, and custom vocabulary biasing — the streaming variant runs over the Live API. On August 27, Gemini Omni Flash reached GA as gemini-omni-1.1-flash, adding video extension, image-to-video interpolation, and explicit resolution control; the gemini-omni-flash-preview endpoint it replaces now has a shutdown date of October 22, 2026, later than the September 30 date previously listed. Preview shutdown dates can move in either direction, so read the deprecations page rather than a date you copied down earlier.

Good starting point: Gemini is a natural evaluation candidate for multimodal workflows — especially when a system must interpret combinations of text, image, video, or audio. Keep the API surface and the enterprise/Vertex deployment surface distinct when comparing governance and data controls.

Reference models: gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite; gemini-3.8-live and gemini-3.8-live-extended-thinking; gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts; gemini-3.5-transcribe and gemini-3.5-transcribe-live; gemini-omni-1.1-flash; antigravity-preview-09-2026; lyria-3.5; gemini-nano-banana-2.1, gemini-3.1-flash-lite-image, and gemini-3-pro-image; Gemma. Gemini 3.7 Flash and 3.5 Flash are deprecated and auto-routed; the Veo 3.1 previews shut down October 22, 2026. Gemini model guide · Antigravity Agent documentation · Gemini API deprecations · Gemini API release notes · Google AI developer platform

xAI

xAI's API exposes the Grok family across text, reasoning, coding, image/video generation, and voice. In September 2026 it released Grok 4.7 (grok-4.7), described as its frontier model for coding, agentic tasks, and knowledge work, and its API listing now recommends Grok 4.7 as the default for everything including code. It takes text and image input with text-only output, adds an xhigh reasoning effort level (the default is high), and is priced the same as Grok 4.6 at $2 / $6 per million input/output tokens below 200K, with the whole request billed at double rates above that. Its Responses API always returns reasoning.encrypted_content, even when it is not requested, which matters if you log or store responses. The Batch API is not yet available for it, and a faster Grok 4.7 Fast variant is offered only in Cursor and Grok Build, not on the public API. Grok 4.6 and Grok 4.5 remain listed alongside it, and all three carry a 500K context window. The listing also keeps Grok 4.3 and the Grok 4.20 reasoning, non-reasoning, and multi-agent snapshots, which have 1M-token context windows at lower per-token prices — so at xAI the longest context currently sits below the flagship, not at it. Grok Build is a coding model intended for agentic development tasks. Grok 4.6 became generally available through Amazon Bedrock on August 19, 2026, and available through Google Enterprise Agent Platform on August 21, expanding deployment choices beyond direct xAI access. xAI's release notes do not yet announce Grok 4.7 on either platform, and Microsoft Foundry's Model Router added grok-4.6, not 4.7, in October. A team consuming Grok through a cloud marketplace or router should expect the recommended model and the model it can actually reach to differ for now.

On the speech side, an earlier revision of this page said grok-voice-transcribe-1.0 remained the default after Grok Voice Transcribe 2.0 shipped in September 2026. xAI's release notes now list 2.0 as the default speech-to-text model. And grok-voice-transcribe-1.0 reached end of life on October 2, 2026, with requests routed to 2.0 at the same price. So the platform did move existing users, silently. Anyone who adopted xAI transcription earlier should re-run their accuracy evaluation against 2.0 now. xAI also added an optional safety_identifier request field in September, which ties policy violations to a specific end user rather than to your API key. For a multi-tenant product, that keeps one user's abuse from becoming the whole account's problem.

One deprecation is worth copying into a design review because it does not fail loudly: grok-imagine-image-quality retires on November 2, 2026, and from that date requests to it are routed to grok-imagine-image-2.0 with quality set to low at lower pricing, keeping the same request and response format. Code that keeps working while quietly producing different output is harder to catch than code that starts returning errors.

Good starting point: Evaluate Grok when your product benefits from a broad multimodal API or when coding-agent capabilities are central. As with any emerging platform, confirm data handling, regional access, retention, and model lifecycle before making it a core operational dependency.

Reference models: grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20-0309-reasoning and grok-4.20-multi-agent-0309, grok-build-0.1, grok-imagine-image-2.0, Grok Imagine video, grok-voice-transcribe-2.0 (default; 1.0 is past end of life), Grok Voice. xAI model documentation · xAI release notes · Grok 4.6 on Amazon Bedrock · Grok 4.6 on Google Enterprise Agent Platform

Mistral AI

Mistral offers both commercial API models and open-weight options, making it relevant when model portability and deployment choice matter. Mistral Medium 3.5 is its frontier-class multimodal model for agentic and coding work, while Mistral Large 3, Mistral Small 4, and the Ministral 3 sizes ship under Apache 2.0 — an unusually permissive license position for models at that capability level. On October 6, 2026, Mistral added Mistral Large 4 (v26.10) in public preview. It is an open-weight, multimodal mixture-of-experts model with 52B active and 1.05T total parameters and a 1M-token context window. List pricing is $1.36 / $4.18 per million input/output tokens, currently shown at a sale price of half that. Two cautions apply. It is a preview, and the model card labels it "Open" without naming a license, unlike the Apache 2.0 terms stated for Large 3. Confirm the license and the exact API ID before planning on it. The specialist catalog is where a lot of the practical value sits: OCR 4.1 for document extraction, Codestral for code, Voxtral TTS and Voxtral Mini Transcribe 2 for speech, and Shieldstral for moderation. OCR 4.0 retired on September 30, 2026, a day after it was deprecated, so OCR 4.1 is now the only current OCR 4 model. Mistral's catalog also hosts third-party Z.ai GLM models, and GLM 5.2 retires on October 31, 2026. Very short deprecation windows on specialist and third-party models are a reason to pin and monitor them.

Good starting point: Put Mistral on the shortlist when you need a strong general model with options beyond a single hosted API, or when OCR/document processing is part of the application. Check the current model card rather than assuming a prior model remains active.

Reference models: Mistral Large 4 (preview), mistral-large-2512 (Mistral Large 3), Mistral Medium 3.5, Mistral Small 4, Ministral 3 (14B / 8B / 3B), Mistral OCR 4.1, Codestral, Voxtral TTS and Voxtral Mini Transcribe 2, Shieldstral. Mistral model overview · Mistral Large 4 model card · Mistral Large 3 model card

Cohere

Cohere is oriented toward enterprise language workloads, retrieval-augmented generation (RAG), tool use, multilingual work, and flexible deployment. The current catalog includes Command A+, Command A Reasoning, Command A Translate, Command A Vision, and Command R7B; the North family is now North Mini Code for agentic coding and North Small Translate for translation. Cohere also maintains Transcribe (including an Arabic-optimized model), Embed, Rerank v4.0 in Pro and Fast variants, a Parse model for document extraction, and the Aya research and Tiny Aya multilingual families.

Good starting point: Cohere deserves careful evaluation for knowledge-grounded enterprise assistants, tool-using workflows, multilingual support, and systems that need a more controlled deployment story. Citation behavior and retrieval quality should be measured with your own knowledge corpus.

Reference models: command-a-plus-05-2026, command-a-reasoning-08-2025, command-a-translate-08-2025, command-a-vision-07-2025, command-r7b-12-2024, north-mini-code-1-0, north-small-translate-1-0, cohere-transcribe-03-2026, embed-v4.0, rerank-v4.0-pro and rerank-v4.0-fast, parse-v5.0, Aya Expanse and Tiny Aya. Cohere model list · Cohere model overview · Cohere release notes

Amazon

Amazon's Nova family is an AWS-native option accessed through Amazon Bedrock. Nova spans multimodal understanding, creative generation, and speech, with Nova 2 as the current generation. The earlier Nova v1 generation is now retired. Nova Premier v1 and Nova Sonic v1 reached end of life on September 14, 2026, and Nova Canvas v1 and Nova Reel v1/v1.1 followed on September 30, 2026, with no public extended access phase. Requests to those IDs now fail. Bedrock publishes its own dates, which can differ from the model provider's.

Bedrock also split its lifecycle policy in two, and which half applies depends on when a model launched. Models launched on or after September 7, 2026 follow a new policy: every model card carries an "EOL no sooner than" date plus a declared Legacy notice period of either six months or 45 days, and the EOL date is added to the card when the model enters Legacy. Models launched before that date stay under the previous policy, which adds a "public extended access" phase — after at least three months in Legacy, active users can keep calling the model until EOL, but at higher pricing set by the model provider. Under both policies, new customers cannot adopt a Legacy model and existing customers can lose access after 15 days of inactivity, so an occasionally used fallback model is not a safe fallback. The practical consequence is that a 45-day notice window is a real possibility now; read the model card, not the general policy page, before you depend on a model. The first exception arrived quickly: OpenAI GPT-6.1 Sol, launched on Bedrock on September 29, 2026, follows OpenAI's first-party deprecation terms (at least six months' notice for generally available models) and has no separate Bedrock minimum-availability date. On Bedrock, partner models increasingly carry their own lifecycle terms, which is one more reason the model card is the document to read.

Good starting point: Consider Nova when your data, identity, audit, networking, and deployment workflows already live in AWS. Bedrock can also be evaluated as a model-access and governance layer rather than as a commitment to a single provider.

Reference models: Nova 2 family. All Nova v1 models referenced in earlier revisions (Premier, Sonic, Canvas, Reel) are retired. Amazon Nova overview · Amazon Bedrock model lifecycle · Bedrock model lifecycle (legacy policy)

Microsoft

Microsoft's differentiator is the enterprise platform around models: Microsoft Foundry provides a catalog, build, evaluation, deployment, and governance surface for Microsoft, OpenAI, Anthropic, Meta, Mistral, Cohere, DeepSeek, xAI, and other models. Microsoft's own families include Phi, MAI, and healthcare-oriented models. Foundry Model Router expanded to 28 regions in August 2026, adding GPT-5.6 Sol/Terra/Luna and Claude Opus 4.8 while removing several retired model versions from its routing pool; in September 2026 it reached 32 regions (adding Canada Central, North Europe, Norway East, and UAE North) and added preview per-request routing metadata reporting the routing mode, ordered model attempts, and any fallback that occurred. That observability matters: a router is only governable if you can see which model actually served a request. September also added preview session affinity for Chat Completions — supply an opaque session ID and the router will try to keep related conversation turns on the same eligible model, reporting whether the association was initialized, retained, or switched. Routing that changes model mid-conversation is a consistency problem as much as a cost one, so this is the control to reach for before you pin a single model and give up routing entirely. In October 2026 the routing pool grew to 35 models, adding gpt-6-astra, claude-fable-5-1, claude-opus-5, claude-sonnet-5, grok-4.6, and three Fireworks-hosted open models (GLM-5.3, GLM-5.3-Flash, and Kimi-K3). The Anthropic additions must be deployed to your Foundry resource first. Note the lag: Opus 5 and Sonnet 5 entered the pool after Anthropic had already moved both to legacy behind Opus 5.5 and Sonnet 5.5, and neither GPT-6.1 Sol nor Grok 4.7 is in it yet. A router's pool trails the providers' current lineups, so check it against your own model policy.

Good starting point: Consider Foundry when Azure identity, networking, compliance controls, and multi-provider governance are more important than direct-to-provider access. Microsoft is often the platform choice, not a single-model answer — confirm the exact model, region, deployment type, and commercial terms.

Reference models: Phi-4 and current Phi family; MAI family; healthcare models; Azure OpenAI offerings; partner models in the catalog; Model Router where policy-based multi-model routing fits. Microsoft Foundry documentation · Model Router updates

DeepSeek

DeepSeek is relevant for its cost-conscious model ecosystem and API compatibility. On September 10, 2026, it released DeepSeek V4.1 Flash, described as the smallest model in a new architecture family, with native multimodal visual understanding, reached through the deepseek-flash ID and accompanied by API price reductions. Two lifecycle details matter more than the benchmark scores. First, the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names now route to V4.1 Flash for compatibility, so a pinned model string here does not pin a model version — if you need reproducibility, version your evaluations rather than trusting the ID. That also resolves the experimental vision model referenced in earlier revisions of this page: it is a redirect now, not a separate model. Second, DeepSeek extended API service for DeepSeek V4 Pro past its previously announced September 14, 2026 cutoff, keeping current billing, which is a reprieve rather than a commitment. The V3 and R1 families remain useful references for reasoning and open-weight ecosystem evaluation.

Good starting point: Use DeepSeek as an evaluation option for coding, reasoning, and compatible API integrations, but complete a deliberate legal, data-residency, security, and vendor-risk review before placing sensitive operational data in the path.

Reference models: deepseek-flash (DeepSeek V4.1 Flash), deepseek-v4-pro, DeepSeek R1 family. DeepSeek API updates · DeepSeek models and pricing · DeepSeek model listing

Alibaba Cloud / Qwen

Alibaba Cloud's Qwen portfolio includes hosted and open-model paths covering text, code, vision, audio, video, and research-oriented workloads. The Qwen3 family supports selectable thinking and non-thinking modes; Qwen 3.8-Max is described by Qwen as its most capable family model for coding and collaborative work, with Qwen 3.8-Flash as the economical tier. Model Studio's featured list now pairs those with Qwen3.7-Plus for vision-language and GUI-agent work and Qwen3.8-Omni-Flash, which appears across the vision, speech-recognition, and omni-modal categories at once. The speech line spans the qwen-audio-3.0 models (TTS, streaming and file transcription) plus qwen-audio-3.1-realtime-plus and qwen3.5-omni-plus for speech-to-speech, and retrieval work is served by text-embedding-v4, qwen3.7-text-embedding, tongyi-embedding-vision-plus, and qwen3-rerank alongside the qwen-image-3.0-pro and wan generation families.

Good starting point: Evaluate Qwen for multilingual applications, coding workloads, hybrid reasoning modes, or an open-model strategy. Treat deployment geography, model access method, and policy requirements as first-order architecture decisions.

Reference models: qwen3.8-max, qwen3.8-flash, qwen3.7-plus, qwen3.8-omni-flash, qwen3.5-omni-plus, Qwen code family, qwen-audio-3.0 speech models and qwen-audio-3.1-realtime-plus, qwen-image-3.0-pro, wan3.0-video, text-embedding-v4, qwen3.7-text-embedding, tongyi-embedding-vision-plus and qwen3-rerank. Qwen API platform · Qwen 3.8-Max release · Alibaba Model Studio model guide

How to evaluate a provider responsibly

Do not select a model from a marketing claim or a broad public benchmark alone. Use a small, representative evaluation pack drawn from the actual work you hope to improve:

Task quality: Does it solve the real task accurately and consistently?
Grounding: Does it cite or point to the correct internal evidence when required?
Failure behavior: Does it surface uncertainty, refuse unsafe actions, and avoid inventing facts?
Latency and cost: Does it meet the interaction and operating-cost target at realistic volume?
Integration: Does it support structured outputs, tool use, modalities, and authentication patterns you need?
Data and governance: Are retention, residency, audit, access, and vendor terms acceptable?
Operational control: Can you trace decisions, monitor quality, route exceptions, and replace the provider if needed?

For operational documentation, compliance, hiring, safety, finance, or other consequential workflows, keep a qualified human responsible for the decision. AI may prepare evidence, flag gaps, or draft recommendations; it should not silently become the accountability layer.

Quick selection patterns

Need	Sensible first evaluation set
High-stakes reasoning/coding with a hosted API	OpenAI, Anthropic, Google, xAI
High-volume classification, extraction, or summarization	Economical tiers from OpenAI, Google, Amazon, Cohere, Mistral, DeepSeek, Qwen
Document/image/video understanding	Google Gemini, OpenAI, Amazon Nova, Mistral, Cohere Vision, Meta Llama 4
Real-time voice, speech synthesis, and transcription	OpenAI GPT-Live and GPT-Transcribe, Google Gemini 3.8 Live, 3.8 Flash TTS, and 3.5 Transcribe, xAI Grok Voice and Voice Transcribe 2.0, Meta Muse Voice Transcribe, Cohere Transcribe, Mistral Voxtral, Qwen audio
Enterprise RAG and tool use	Cohere, Anthropic, OpenAI, Google, Amazon Bedrock, Microsoft Foundry
Self-hosting or deployment control	Meta Llama and Muse Glimmer, Mistral, Qwen, DeepSeek, Microsoft Phi
Azure-centered governance	Microsoft Foundry and the models available in its catalog
AWS-centered governance	Amazon Bedrock and its available provider catalog
Multilingual focus	Google, Cohere, Mistral, Qwen, Meta Llama, OpenAI, Anthropic — validated on your languages and tasks

The durable rule

Model selection is an architecture decision with a short shelf life. Keep a small abstraction boundary around the provider, version your prompts and evaluations, record the rationale in an architecture decision record, and revisit the choice as capabilities and terms evolve.

The strongest AI systems are not built around allegiance to a model brand. They are built around a clear task, reliable evidence, disciplined evaluation, and accountable operational ownership.

Change log

October 12, 2026

Added OpenAI's GPT-6.1 Sol (gpt-6.1-sol, September 29, 2026; $2 / $10, 1.05M context, multi-agent delegation in beta). It replaces GPT-6 Sol in the featured catalog; the GPT-6 Sol page is still live with no deprecation notice. Recorded its Amazon Bedrock availability under OpenAI's own lifecycle terms. Added the Ultrafast service tier (6× standard; US-only residency on Astra, US and EU on 6.1 Sol), Agents API computer use, the Decisions API beta, the three-tier usage structure, and the HIPAA BAA flow. Updated the at-a-glance row, reference list, and links.
Moved OpenAI's lifecycle dates forward. gpt-5.4-cyber shut down October 1, and the September 28 snapshot retirements are now past. Added gpt-image-1 to the October 23 wave. Added the October 1 deprecations: the TTS models shut down January 6, 2027, and gpt-5.1, gpt-5.3-codex, and gpt-5.4-nano shut down April 1, 2027. Added the November 30, 2026 removal of reusable prompts, the Evals dashboard and API (read-only from October 31), and Agent Builder, because eval history stored there needs exporting. Corrected the image-model replacement from gpt-image-2 to GPT Image 2.5 Sunburst or Flare, per the deprecations page.
Updated Anthropic's lineup with Claude Sonnet 5.5 (claude-sonnet-5-5, September 28) and Claude Haiku 5.5 (claude-haiku-5-5, October 7), moving Sonnet 5 and Haiku 4.5 to legacy. Added Haiku 5.5's prompt-length pricing step at 100K tokens and the Sonnet 5.5 cache-read price cut. Added each model's breaking changes (between_tools replacing disabled thinking on Sonnet 5.5; prefill, sampling parameters, and manual thinking budgets rejected on Haiku 5.5), plus the token-count and account-bound thinking changes. Extended the refusal-handling guidance to every current tier.
Rewrote Anthropic's lifecycle notes. All current models now have 2027 retirement floors. Claude Sonnet 4.5 was deprecated September 30 and retires November 30, 2026. The Haiku 4.5 floor is three days out. On Bedrock, Sonnet 4 reaches end of life October 14, Sonnet 4.5 entered Legacy October 8, and Opus 4.1 entered public extended access October 8. Added the October Managed Agents changes (web_fetch limited to URLs already in the session, allowed_hosts applying to web tools, dynamic workflows in beta) and the new Models API capability fields.
Added Google's October 8 deprecation of Gemini 3.7 Flash and Gemini 3.5 Flash: requests now auto-route to 3.8 Flash and 3.6 Flash, so pinned IDs no longer pin the serving model. Added Nano Banana 2.1 GA (October 6) and the deprecation of gemini-3.1-flash-image. Added the October 22, 2026 shutdown of the Veo 3.1 previews in favor of Gemini Omni, the Deep Research preview shutdown (October 23), and the November 17 voice-preview shutdowns. Corrected the gemini-omni-flash-preview shutdown from September 30 to October 22. Recorded that antigravity-preview-05-2026 shut down October 5.
Corrected xAI's transcription note. Grok Voice Transcribe 2.0 is now the default, and grok-voice-transcribe-1.0 reached end of life on October 2, with requests silently routed to 2.0. Added the safety_identifier request field. Clarified that Grok 4.7 Fast exists only in Cursor and Grok Build.
Rewrote the Meta section around the hosted Meta Model API and Muse family (Muse Wonder in preview, Muse Spark, Muse Image, Muse Voice Transcribe, open-weight Muse Glimmer and SAM 3.1), alongside Llama 4. Flagged Contributor mode, where discounted pricing comes in exchange for training use of your data, as a governance decision. Replaced the Meta links, which now redirect to dev.meta.ai. Updated the at-a-glance, voice, and self-hosting rows.
Added Mistral Large 4 (public preview, October 6, 2026; open-weight MoE, 1M context), with cautions that its license is not named and its API ID should be confirmed. Recorded the September 30 retirement of OCR 4.0 and the October 31 retirement of the hosted Z.ai GLM 5.2.
Moved the Amazon Nova Canvas v1 and Nova Reel v1/v1.1 end of life to past tense (September 30, 2026). Noted that GPT-6.1 Sol on Bedrock follows OpenAI's lifecycle terms rather than Bedrock's.
Added Microsoft Foundry Model Router's October expansion to 35 models (GPT-6 Astra, Claude Fable 5.1, Opus 5, Sonnet 5, Grok 4.6, and Fireworks-hosted GLM-5.3, GLM-5.3-Flash, and Kimi-K3). Noted that the router's pool lags the providers' current lineups.
Confirmed no material change this period for Cohere, DeepSeek, and Qwen; those sections stand as written.

September 28, 2026

Added OpenAI's GPT-6 Sol (gpt-6-sol, $2 / $10) and GPT-6 Luna (gpt-6-luna, $0.10 / $0.50), released September 22, 2026. Both have a 1.05M-token context window and Astra's 272K long-context surcharge, and there is no GPT-6 Terra. Recorded that the GPT-5.6 pages remain live with no deprecation notice, so GPT-6 is an option rather than a forced migration. Added the September 25 image-encoding fix, which means vision and computer-use evaluations run on Sol or Luna before that date should be re-run. Added both models to the at-a-glance row, the reference list, and the links.
Moved OpenAI's lifecycle dates forward. The Videos API and Sora 2 were removed on September 24. The gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106 snapshots retire today, with gpt-5.6-terra as the replacement. The gpt-5.4-cyber shutdown is now three days out.
Replaced Claude Opus 5 with Claude Opus 5.5 (claude-opus-5-5, September 22, 2026) in Anthropic's current lineup, the at-a-glance row, and the reference list, and moved Opus 5 to legacy-but-available. Opus 5.5 costs $4 / $20 against Opus 5's $5 / $25 and has a retirement floor of not sooner than September 22, 2027. Updated Anthropic's documented starting-model guidance from Opus 5 to Opus 5.5.
Added Opus 5.5's four breaking changes against Opus 5: always-on thinking, forced tool use returning a 400, thinking blocks bound to the model and conversation, and computer_20251124 rejected on the Claude API and Google Cloud. Also added the silent response-shape change for text between tool calls, and its biology and reasoning_extraction refusal categories.
Added Anthropic's September 24 decision to resume billing for pre-output refusals in the bio, frontier_llm, and reasoning_extraction categories, which turns refusal handling into a cost line. Added the other platform changes from the same period: tools definable in mid-conversation system messages (beta), cache diagnostics out of beta, Compliance API GA for Claude for Microsoft 365 sessions, and the Activity Feed no longer returning file names or artifact titles. Tightened the Claude Haiku 4.5 retirement floor to two and a half weeks out.
Added Google's Gemini 3.8 Flash TTS and Flash-Lite TTS GA (September 22), including the Flash-Lite model's role as the replacement for gemini-3.1-flash-tts-preview. Added the September 18 restriction of Gemini 2.5 access to users who have actively used it, because a fallback built on 2.5 may not be reachable. Tightened the Antigravity antigravity-preview-05-2026 shutdown to one week out and the gemini-omni-flash-preview deprecation to two days out. Broadened the voice quick-selection row to include speech synthesis.
Added xAI's Grok 4.7 (grok-4.7, September 2026) as the listing's new recommended default, priced the same as Grok 4.6. Noted that its Responses API always returns encrypted reasoning content, and that Batch API support and Bedrock or Google availability are not yet announced.
Tightened the Amazon Nova Canvas v1 and Nova Reel v1/v1.1 end of life to two days out (September 30, 2026).
Confirmed no material change this period for Meta, Mistral, Cohere, Microsoft, DeepSeek, and Qwen; those sections stand as written.

September 21, 2026

Added Google's Antigravity Agent 09-2026 (antigravity-preview-09-2026, September 17) — a managed agent running in a Google-hosted Linux sandbox, built on and defaulting to Gemini 3.8 Flash — and recorded that it is a breaking upgrade: tool parameters moved to PascalCase, file editing moved to line-range replacements, find_by_name and grep_search were added, and the previous antigravity-preview-05-2026 shuts down October 5, 2026. Added Antigravity to the at-a-glance row and the reference model list.
Added OpenAI's gated GPT-5.6 Cyber line, which does not appear in a general model listing: access runs through the Daybreak program, the model carries a 400K context window at $12.50 / $75 per million input/output tokens, and its predecessor gpt-5.4-cyber was deprecated September 11, 2026 with an October 1, 2026 shutdown. Changed the at-a-glance description from "specialist research models" to name the cyber and life-sciences lines as gated.
Added OpenAI's September 15 API key creation governance — organizations and projects can restrict key creation to service accounts, to user-owned project keys, or disable it entirely, with organization settings taking precedence — as a companion to the key expiration control added September 10.
Moved the Videos API and Sora 2 removal from "nine days out" to three days out (September 24, 2026), and listed the migration waves that follow: legacy snapshots including gpt-3.5-turbo, the GPT-4 variants, o1, o3-mini, and o4-mini on October 23, 2026; gpt-image-1-mini, gpt-image-1.5, and chatgpt-image-latest on December 1, 2026; and the GPT-5 and o3 snapshots on December 11, 2026.
Recorded no change to Anthropic's four-model lineup, and added two things that do change planning. Claude Haiku 4.5 carries the lineup's nearest retirement floor — not sooner than October 15, 2026 — against 2027 dates for Fable 5.1, Opus 5, and Sonnet 5, so an economical tier built on it needs a tested fallback. And on September 18 the Compliance API began returning Claude in Chrome session transcripts under the claude_in_chrome product surface (beta, Enterprise), which makes browser assistant activity retrievable evidence and therefore a records-retention question.
Added xAI's Grok Voice Transcribe 2.0, noting that grok-voice-transcribe-1.0 remains the default — the inverse of the usual lifecycle hazard, since nothing changes underneath you and nothing improves until you select the model explicitly.
Sharpened the Amazon Nova Canvas v1 and Nova Reel v1/v1.1 end of life from "this month" to nine days out (September 30, 2026), and clarified that neither was given a public extended access phase, so that date ends access rather than raising the price.
Refreshed the Qwen references against Model Studio's current featured list, adding qwen3.8-omni-flash (which spans the vision, speech-recognition, and omni categories), qwen-audio-3.1-realtime-plus for speech-to-speech, and the qwen3.7-text-embedding and tongyi-embedding-vision-plus retrieval models.
Confirmed no material change this period for Meta, Mistral, Cohere, Microsoft, and DeepSeek; those sections stand as written.

September 15, 2026

Added OpenAI's September 8 and September 10 releases: GPT-Image-2.5 Sunburst and Flare with the new xhigh and max quality settings, GPT-Rosalind Research GA for life sciences (with billing starting October 5, 2026 and access limited to approved organizations), GPT-Live 1 GA for full-duplex voice at $0.05 per minute plus separate backend model and tool charges, the Agents API public beta, prompt cache diagnostics GA, and organization-enforced API key expiration. Updated the reference list from GPT Image 2 to the current image, voice, and transcription IDs, and flagged the Videos API / Sora 2 removal as nine days out rather than merely scheduled.
Recorded no change to Anthropic's four-model lineup, and added the two platform changes that do affect agent design: on-demand Messages API compaction (September 14, beta compact-2026-09-04), which makes context budget a request-level control, and the Managed Agents auto permission policy (September 10), which moves the tool-call approval gate into the platform. Noted the 2.5% prompt-cache read rate on Fable 5.1 and Mythos 5.1, and Amazon Bedrock's nearer Claude date — Sonnet 4 end of life on October 14, 2026.
Added Google's Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking GA (September 15) for real-time audio-to-audio voice work, and corrected Lyria 3.5 from public preview to generally available. Noted that generative-media lifecycle labels differ per model — Nano Banana image models GA, Imagen 4 deprecated, Veo 3.1 still preview.
Corrected the xAI section to match its current model listing: Grok 4.6 is the recommended default at 500K context, while Grok 4.3 and the Grok 4.20 snapshots carry 1M-token context at lower prices, so the longest context sits below the flagship. Added the grok-imagine-image-quality retirement on November 2, 2026, where requests are silently rerouted to grok-imagine-image-2.0 at quality: low instead of failing.
Rewrote the Amazon section around Bedrock's split lifecycle policy: models launched on or after September 7, 2026 carry an "EOL no sooner than" date with a six-month or 45-day Legacy notice period, while earlier models keep the public extended access phase at provider-set higher pricing. Moved Nova Premier v1 and Nova Sonic v1 to retired (September 14, 2026) and left Nova Canvas v1 and Nova Reel v1/v1.1 as September 30 migrations. Added the 15-day inactivity rule, because it means a rarely exercised fallback model is not a real fallback.
Updated DeepSeek to V4.1 Flash (September 10, 2026), reached via deepseek-flash with price reductions, and recorded that the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names now route to it — a pinned ID no longer pins a version. Also noted DeepSeek V4 Pro service continuing past its September 14, 2026 cutoff.
Added Microsoft Foundry Model Router's preview session affinity for Chat Completions, which addresses mid-conversation model switching without giving up routing.
Refreshed the Cohere references against the current model list, replacing North Micro Vision — no longer listed — with North Small Translate, and adding Command A Reasoning, Parse, Rerank v4.0 Pro/Fast, the Arabic transcription model, and Tiny Aya. Refreshed Mistral with its Apache 2.0 licensing position and named specialist models (OCR 4.1, Codestral, Voxtral, Shieldstral), and Qwen with its current image, video, embedding, and reranking IDs.
Added a real-time voice and transcription row to the quick selection patterns, since three providers shipped voice or transcription changes in this period.
Confirmed no material change this period for Meta; that section stands as written.

September 7, 2026

Added OpenAI's GPT-6 Astra (gpt-6-astra, released September 3, 2026) as a generation above GPT-5.6, with its context window, pricing, the 272K-token long-context surcharge, and the async tool calling / mid-turn steering / mid-conversation effort controls that change how an agent loop is written. GPT-5.6 Sol, Terra, and Luna are now described as current tiers beneath it rather than as the frontier family.
Added OpenAI's August 26, 2026 transcription deprecation (whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize, shutting down February 26, 2027) and reconfirmed that the Videos API and Sora 2 removal is still set for September 24, 2026.
Updated Anthropic to Claude Fable 5.1 (released September 1, 2026) as the current top model, with Fable 5 moving to legacy-but-available and Claude Mythos 5.1 replacing Mythos 5 as the invitation-only Project Glasswing offering. Recorded Anthropic's documented guidance to start with Opus 5 for most workloads, because it argues against defaulting to the top of a provider's list.
Replaced the Anthropic integration notes with Fable 5.1's three breaking changes — forced tool use returning a 400, thinking blocks bound to the producing model, and history edits invalidating later thinking blocks (enforced for accounts created on or after August 31, 2026) — and added the text watermark and C2PA Content Credentials as a provenance fact. Kept the Opus 4.8 → Opus 5 adaptive-thinking hazard and the 30-day retention, no-zero-data-retention, and refusal-as-HTTP-200 terms, which now apply to Fable 5.1 and Mythos 5.1.
Added Google's Gemini 3.8 Flash GA (September 2, 2026), agentic video understanding for Gemini 3.7/3.6 Flash and 3.5 Flash-Lite (September 1), and Lyria 3.5 in public preview (September 3), while noting that the earlier Flash and Flash-Lite models remain stable — a new top option, not a forced migration.
Updated Microsoft Foundry Model Router to 32 regions and added its preview per-request routing metadata, since routing is only governable when you can see which model served a request.
Sharpened the Amazon Nova v1 end-of-life language from future planning to immediate action, with Nova Premier v1 and Nova Sonic v1 retiring September 14, 2026.
Confirmed no material change this period for xAI, Meta, Mistral, Cohere, and DeepSeek, and no change to the featured Qwen lineup; those sections stand as written.

August 31, 2026

Corrected the Anthropic section to the four-model lineup Anthropic currently documents — Claude Fable 5, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5 — after verifying that Claude Opus 5 superseded Opus 4.8 at the Opus tier and that Fable 5 is a widely released model rather than a controlled-access one. Mythos 5 is now identified as the only access-gated offering.
Added the Claude Opus 5 thinking-on-by-default breaking change and the Fable 5 refusal/data-retention terms, because both change integration and governance work rather than just model choice.
Noted OpenAI's August 26, 2026 Assistants API sunset and the September 24, 2026 removal of the Videos API and Sora 2 family, which is a provider exiting a modality and not only a model swap.
Added Google's August 26 GA of Gemini 3.5 Transcribe and Transcribe Live and the August 27 GA of Gemini Omni Flash, with the September 30, 2026 deprecation of the preview omni endpoint.
Refreshed the Qwen references to the Model Studio featured list, replacing Qwen3.6 with Qwen3.7-Plus and adding Qwen 3.8-Flash and the audio families.
Confirmed no material change this period for xAI, Meta, Mistral, Cohere, Amazon, Microsoft, and DeepSeek; those sections stand as written.

August 24, 2026

Updated OpenAI GPT-5.6 Sol pricing/lifecycle guidance and corrected the primary model-catalog link.
Reframed Anthropic around its broadly available Sonnet 5, Opus 4.8, and Haiku 4.5 lineup; retained Fable/Mythos as specialized-access references.
Added Grok 4.6 availability through Amazon Bedrock and Google Enterprise Agent Platform.
Expanded Cohere references to include North, Transcribe, Embed, and Rerank families.
Marked affected Amazon Nova v1 models as legacy with September 2026 end-of-life dates.
Updated Microsoft Foundry Model Router coverage and current routing pool.
Added DeepSeek V4 Flash Vision Experimental with an explicit production-lifecycle caution.