Leading AI Model Providers (2026)
Snapshot as of 2026-10-02 (re-verified; first written 2026-09-30). The 2025 edition (GPT-4o, Gemini 1.5, Claude 3, Llama 3, Grok-1) is obsolete and was replaced. Model names change monthly: confirm on the vendor page before relying on a row. “Open” = weights downloadable (licence varies; check it).
1. Current flagships
| Provider | Current flagship(s) | Open / closed | Where to get it | Vault note |
|---|---|---|---|---|
| OpenAI | GPT-6 Astra (2026-09-03), GPT-6.1 Sol (09-29), GPT-6 Sol / Luna (09-22). Codex now runs on Astra / 6.1 Sol / 6 Luna | Closed (gpt-oss is a separate open line) | OpenAI API, ChatGPT (Work/Codex), Azure | gpt-5-and-gpt-6-family |
| Anthropic | Fable 5.1 (09-01), Opus 5.5 (09-22), Sonnet 5.5 (09-28), Haiku 5.5 (10-07); Mythos 5.1 invite-only (Glasswing) | Closed | Claude API, Bedrock, Vertex, Foundry | Claude hub |
Stable (API docs): Gemini 3.8 Flash (GA 09-02), 3.7 / 3.6 / 3.5 Flash. Pro tier: only gemini-3.1-pro-preview (preview). Gemini 4 Argon announced 09-30: restricted to trusted cyber defenders (Fairwind Program), not generally available, not on the API models page; Gemini 3.5 Pro was never released (see note below) | Closed (open: Gemma 4) | Gemini API / AI Studio, Vertex | gemini-3-x-family, gemma-4 | |
| SpaceXAI | Grok 4.7 (2026-09-21, grok-4.7, 500k ctx); Grok 5 not released. xAI is a SpaceX subsidiary since 2026-02-02, branded SpaceXAI | Closed | xAI API, Grok app, Cursor, OpenRouter | grok |
| Meta | Muse Spark (closed): 1.1 (07-09, Meta blog per Meta Muse and Llama), 1.2 (08-05) and 1.3 (09-02) per Wikipedia (secondary), Muse Code agent (08-05); Muse Glimmer 30B (open, Apache 2.0, 2026-08-10, distilled from Muse Spark; Meta research blog). Newest Llama on Hugging Face is still Llama 4 (2025) | Mixed | Meta AI; Hugging Face (Muse Glimmer) | Meta Muse and Llama |
| Mistral | Mistral Medium 3.5 (2026-05-22, Modified MIT), Small 4 (2026-03-16, Apache 2.0), Large 3 (675B MoE, Apache 2.0, Dec 2025); Magistral / Devstral 2 / Codestral (Premier licence) | Mostly open | La Plateforme, Hugging Face | mistral-family |
| DeepSeek | DeepSeek-V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B active), 2026-04-24, 1M context, open weights; V4.1-Flash (09-10, MIT). deepseek-chat / deepseek-reasoner retired 2026-07-24. No R2 | Open | DeepSeek API, Hugging Face | deepseek-family |
| Alibaba Qwen | Qwen3.8-Max (open weights as Qwen3.8-2.4T-A95B, 08-12, restricted licence); Qwen3.8-27B (Apache); Qwen 4 announced 09-22, unreleased | Mixed | Alibaba Cloud, Hugging Face | qwen-family |
| Moonshot Kimi | Kimi K3 (hosted 07-16; weights 07-26/27; 2.8T MoE, 104B active, 1M ctx; custom licence). K2.6 and K2.7 Code still served | Open (custom licence) | kimi.com, Kimi API, Hugging Face | kimi-k3 |
| Z.ai GLM | GLM-5.3 (API 08-14; weights late Aug, 753B, GLM licence) | Open | Z.ai API, Hugging Face | GLM-5 series |
| MiniMax | MiniMax M3 (API 06-01, weights 06-02, per techtimes.com 2026-06-01; 428B MoE / ~23B active, 1M ctx, community licence) | Open | MiniMax API, Hugging Face | minimax-m3 |
| Nvidia | Nemotron 3 Ultra (06-04; 550B / 55B active hybrid Mamba-Transformer MoE; open weights; licence reported as NVIDIA Open Model License by Wikipedia and OpenMDW-1.1 by the vault’s nvidia-nemotron note: check the model card); Nemotron 3 Nano (Dec 2025), Super (Mar 2026) | Open | Hugging Face, build.nvidia.com (NIM) | nvidia-nemotron |
| Microsoft | MAI-Thinking-1 (microsoft.ai: ~1T total / 35B active MoE, public preview in Microsoft Foundry and the MAI Playground; context window and pricing not stated on the model page); hosts OpenAI and Claude via Foundry | Closed (MAI) | Microsoft Foundry, Copilot | (none) |
| amazon | On Bedrock, Nova Premier and Nova Sonic v1 reached end-of-life 2026-09-14 and Nova Canvas and Nova Reel 2026-09-30 (AWS Bedrock legacy lifecycle page) | Closed | Amazon Bedrock | (none) |
| IBM | Granite 4.2 (3B / 8B / 30B dense reasoning models, 2026-08-25, Apache 2.0, 128K ctx), Granite Speech 5.0 Turbo | Open | Hugging Face, watsonx.ai, Ollama | granite |
Gemini 3.5 Pro vs Gemini 4 Argon (conflict resolved)
A web search summary said Argon “replaces Gemini 3.5 Pro”. Google’s own Argon post (2026-09-30) does not mention 3.5 Pro. Google’s API models page lists no 3.5 Pro, only gemini-3.1-pro-preview as Pro tier, and no Gemini 4. 9to5Google’s launch report refers to the “cancellation of Gemini 3.5 Pro” and a shift of focus to 3.8 Flash. Reading: 3.5 Pro was announced at I/O (2026-05-19), repeatedly delayed, never shipped, and Argon is now the frontier model. The “cancelled” wording is press, not Google. Appendix A “3.5 Pro not released” stands.
Verification status
Verified from primary pages: Google (blog + API docs), xAI docs, Mistral docs/news, DeepSeek API docs, Meta research blog (Muse Glimmer), IBM / Hugging Face (Granite), Kimi platform pricing, Microsoft AI model page (MAI-Thinking-1 parameters, preview status), AWS Bedrock lifecycle page (Nova EOL dates), Mistral news (Series D EUR 3B, valuation above EUR 21B, 2026-09-08). OpenAI and Anthropic rows follow the vault’s already-verified hubs (gpt-5-and-gpt-6-family, Claude hub). Attributed to secondary sources: Muse Spark 1.2/1.3 dates (Wikipedia), MiniMax M3 dates (techtimes.com), the Nemotron licence (Wikipedia says NVIDIA Open Model License; the vault’s nvidia-nemotron note says OpenMDW-1.1).
2. How to choose
- Task and budget: frontier agentic coding → Astra / Fable / Gemini 4 Argon when it opens up; cost-efficient high volume → GPT-6 Luna, Flash-tier Gemini, Haiku.
- Open weights / self-hosting: DeepSeek V4.x, Kimi K3, GLM-5.3, Qwen3.8, Nemotron 3 Ultra, Muse Glimmer, Granite 4.2, Gemma 4, gpt-oss. Read each licence (several are custom, with revenue thresholds).
- Context: 1M tokens is now common (GPT-6, Claude, Gemini 3.x, DeepSeek V4, Kimi K3, MiniMax M3).
- Access restrictions: some top models are gated (Claude Mythos via Glasswing; Gemini 4 Argon via Fairwind).
- Compliance: EU AI Act GPAI obligations, data residency, hosting on Bedrock/Vertex/Foundry.
- Benchmarks: use live boards, not notes: llm-comparisons-artificialanalysis.ai, artificial-analysis-llm-api-provider-leaderboard.
3. Related notes in ai-models/
ai-models - language-models - granite (IBM) - minimax-m1 - kimi-k2-dot-6 - gemma-4 - image: nano-banana-pro, flux-dot-1 - video: google-veo, openai-sora (discontinued)
The 2025 edition also listed Apple, Baidu, Tencent, Stability, Cohere, AI21, Databricks, Snowflake and others; they are omitted here because they were not re-checked.
Limits
Rows for Qwen, GLM, Kimi K3, MiniMax and Nemotron lean on Hugging Face cards and secondary reports; the Muse Spark 1.2/1.3 dates are per Wikipedia; Samsung’s reported stake in Mistral is not included. Confirm any row on the vendor page before relying on it.
Sources (accessed 2026-10-02 unless noted)
- https://microsoft.ai/models/mai-thinking-1/ ; https://microsoft.ai/news/
- https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html (Nova EOL)
- https://en.wikipedia.org/wiki/Muse_Spark ; https://en.wikipedia.org/wiki/Nemotron ; https://en.wikipedia.org/wiki/Mistral_AI
- https://mistral.ai/news (Series D)
- https://developers.openai.com/api/docs/models , /changelog , /deprecations
- https://platform.claude.com/docs/en/models/overview
- https://ai.google.dev/gemini-api/docs/models ; https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/
- https://docs.x.ai/developers/grok-4-7 ; https://docs.x.ai/developers/release-notes
- https://docs.mistral.ai/getting-started/models/models_overview/
- https://api-docs.deepseek.com/news/news260424 ; https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
- https://huggingface.co/moonshotai/Kimi-K3 ; https://huggingface.co/Qwen ; https://huggingface.co/zai-org/GLM-5.3
- https://en.wikipedia.org/wiki/Qwen ; https://en.wikipedia.org/wiki/Muse_Spark
- https://9to5google.com/2026/09/30/gemini-4-argon-announcement/ (3.5 Pro “cancellation” wording)
- https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model
- https://mistral.ai/news/ ; https://docs.mistral.ai/getting-started/models/models_overview/
- https://api-docs.deepseek.com/news/news260424
- https://platform.kimi.ai/docs/pricing/chat
- https://www.ibm.com/granite ; https://huggingface.co/ibm-granite
- https://x.ai/news/xai-joins-spacex ; https://en.wikipedia.org/wiki/SpaceXAI
- https://learn.chatgpt.com/docs/models (Codex models)
- Secondary: MiniMax M3 and Nemotron 3 Ultra (techtimes.com 2026-06-01; Wikipedia)