OCTOBER 2026 AT A GLANCE
RELEASED — 40+ models and tools you can use today, from Claude Sonnet 5.5 on 1 October to DeepSeek V4.1-Flash-1017 on the 17th.
PREVIEW ONLY — Mistral Large 4, Reka Rho-1, Tavus Griffin-Lite and Helix World exist but are gated. Do not plan around them.
NOT RELEASED — NVIDIA Bolt E4B (due 23 Oct), Mistral Large 4 open weights (promised end of month), Microsoft Maia 300, the OpenAI humanoid robot. Announcements, not products.
Dates below are cross-checked against three independent release trackers. Where they disagreed, we say so rather than picking one.
Read this before you trust any October 2026 release list
We checked these dates against three separate trackers and found a systematic error worth knowing about.
Several widely-shared October release lists date everything to 1, 8 or 15 October. Those are Thursdays. They are the publication dates of a weekly AI newsletter, not the dates the models came out. Lists built by copying that newsletter inherit its calendar.
Three examples where the real date is different:
- Claude Haiku 5.5 — listed as 8 October in weekly roundups. Two per-day trackers both say 7 October.
- Mistral Large 4 — listed as 8 October. Both per-day trackers say 6 October, and both call it a preview, not a release.
- Grok Voice 3.0 — listed as 15 October. Both say 14 October.
Where sources agreed on a day, we use it. Where only a weekly roundup carried an item, the table says wk of rather than inventing a date we cannot support.
RELEASED — available now
Frontier and flagship models
| Date | Company | Model | What it is |
| 1 Oct | Anthropic | Claude Sonnet 5.5 | $2 / $10 per million tokens; 70.6 on Terminal-Bench 4.0 |
| 1 Oct | OpenAI | GPT-6.1 Sol | Update to GPT-6 Sol; 95% cheaper cached input |
| 7 Oct | Anthropic | Claude Haiku 5.5 | $0.10 per million input tokens |
| 8 Oct | OpenAI | GPT-6 in ChatGPT | Became the default model for all users, with Intelligent UI |
| 13 Oct | Google DeepMind | Gemini 4 Argon | Paid Gemini API and Android Studio. Limited release to government and cyber-defence first, on 1 Oct |
| 15 Oct | Meta | Muse Spark Open | 816B total parameters, open weights |
| 17 Oct | DeepSeek | V4.1-Flash-1017 | 552B MoE, MIT licence; $0.30 / $1.20 per million at peak |
Open-weight releases
| Date | Company | Model | Licence / size |
| 3 Oct | Aleph Alpha | Kolibri-1 | Apache 2.0, 78B MoE, German and English |
| 6 Oct | Google DeepMind | EmbeddingGemma 2 | Apache 2.0, multimodal embedding |
| 6 Oct | Perplexity | pplx-decider-v1.1-27b | Open weights |
| 7 Oct | Liquid AI | d1-3B, d1-omni-600M | Open weights; omni adds audio |
| 8 Oct | JetBrains | Mellum2.1 | Apache 2.0 |
| 9 Oct | Cloudflare | Clef-omni | Open weights on Hugging Face |
| 15 Oct | NVIDIA | Nemotron 3.5 Super VL | 120B vision-language, 12B active |
| 15 Oct | Scira | Amos 1 | Four open models, 0.5B to 27B |
| 15 Oct | OpenBMB | MiniCPM-V 4.7 | 35B-A3B vision, plus a 1B variant |
| 15 Oct | Fish Audio | Drama 3 | Open-weights text-to-speech |
| 16 Oct | Reflection AI | Beam | 501B, Apache 2.0 |
| wk of 1 Oct | H Company | Holo4 | Open-weight computer-use, 27B and 35B-A3B |
Voice, image and video
| Date | Company | Model | Detail |
| 5 Oct | Amazon | Nova 2.5 Sonic | Generally available |
| 6 Oct | Google DeepMind | Nano Banana 2.1 | $0.0336 per 1,000 images |
| 14 Oct | xAI | Grok Voice 3.0 | Speech-to-speech, via Voice Agent API |
| 15 Oct | OpenAI | GPT-Live-1.1 | Live voice update; regional rollout |
| wk of 1 Oct | ElevenLabs | v4 and v4 Turbo | 90 languages |
| wk of 8 Oct | Black Forest Labs | FLUX 3 Image | 4K output |
| wk of 8 Oct | Tavus | Griffin | Conversational video |
Tools, platforms and features
| Date | Company | Release |
| 1 Oct | OpenAI | Dots — always-on ChatGPT agents on GPT-6 Astra, Pro only |
| 1 Oct | OpenAI | UltraFast mode — served on Cerebras, $500/month Pro tier |
| 1 Oct | OpenAI | Sign in with ChatGPT — third-party apps use a user's existing plan |
| 1 Oct | OpenAI | Codex cloud — the Codex harness running fully in the cloud |
| 1 Oct | Cloudflare | cf CLI — agentic command line over 3,000+ API operations |
| 1 Oct | CoreWeave | Forge, Agent Lens (public preview), Sandboxes GA, serverless GPUs |
| 8 Oct | OpenAI | Decisions API — public beta, after a 1 Oct waitlisted preview |
| 8 Oct | Anthropic | Claude for Google Workspace — Docs, Sheets and Slides sidebar, beta on paid plans |
| 8 Oct | Anthropic | Sonnet 5.5 cache-read price halved |
| 9 Oct | Microsoft | Microsoft-Decision-1 — 9B, via Foundry and OpenRouter |
| 12 Oct | Meta | Muse Wonder and Muse Wonder Full (the Full version is Muse Code subscribers only) |
| 15 Oct | Anthropic | OpenTelemetry agent traces — public beta |
| 16 Oct | Anthropic | Claude Code 2.1.302 — 63 changes; fork subagents on by default |
PREVIEW — exists, but gated
These are real and some press coverage calls them released. They are not generally available, and building a plan around them is how a launch slips.
| Date | Company | Model | Status |
| 6 Oct | Mistral AI | Mistral Large 4 | Public preview. 1T parameters, multimodal. Open weights are a separate promise. |
| 5 Oct | Reka | Rho-1 | 19B omni, research preview, limited access |
| 1 Oct | Tavus | Griffin-Lite | Research preview, selected testers only |
| 12 Oct | Noiz AI / HKUST | Helix World | Preview checkpoint on Hugging Face |
| 14 Oct | Sqwish Labs | kettei fat | Preview, not yet available |
NOT RELEASED — announced only
Everything below has a name and a press mention and nothing you can use. Two have dates. The rest do not.
| Company | What | When |
| NVIDIA | Bolt E4B 4GB | 23 October 2026 — the only firm date in this table |
| Mistral AI | Mistral Large 4 open weights | Promised for end of October |
| Microsoft | Maia 300 inference chip | Announced 15 Oct, no ship date |
| Microsoft | MAI-Transcribe-2-Streaming, MAI-Voice-2.1, MAI-Voice-2.1-Flash | Announced, availability not stated |
| Meta + Sierra | Personal Agent Protocol | Announced as an open standard |
| CoreWeave | NVIDIA Vera CPU | Described as coming |
| Decagon | Chord | Used inside Voice 3; standalone access not stated |
| OpenAI | Humanoid robot | Altman: a demo in 2027, street deployment years later |
| OpenAI | aurora-red / green / blue-alpha | Three model IDs reported in the API on 18 Oct. No documentation, no changelog, no company comment, and the original sighting is unverified. Treat as rumour. |
One retirement
Not a release, but it lands in the same month and will break things: Laguna S 2.1 retires on 31 October 2026. If anything you run still points at it, that is a deadline rather than a roadmap item.
What October 2026 actually tells you
The cheap tier got serious. Claude Haiku 5.5 at $0.10 per million input, Sonnet 5.5 at $2/$10, DeepSeek V4.1-Flash at $0.30/$1.20, GPT-6.1 Sol cutting cached input by 95%. The interesting pricing movement this month was at the bottom of the stack, not the top.
Open weights kept pace with closed. Beam at 501B under Apache 2.0, Muse Spark Open at 816B, DeepSeek under MIT, Nemotron 3.5 at 120B. Four genuinely large open releases inside three weeks is not a trickle.
Small decision models were the quiet theme. Microsoft-Decision-1 at 9B, Liquid d1-3B and d1-omni-600M, Scira Amos 1 from 0.5B, Cloudflare Clef, Fastino GLiDE, Perplexity pplx-decider. A whole category appeared in one month, and almost none of it made headlines.
And the gap between announced and usable widened. This is the reason the three sections above are split the way they are. Mistral Large 4 was reported as a release and is a preview. Beam's weights were promised for the month before they landed. If you planned October around press coverage rather than availability, you planned around four things that did not exist yet.
FAQ
What AI models were released in October 2026?
Over forty, including Claude Sonnet 5.5 and Haiku 5.5, GPT-6.1 Sol and GPT-6 as the ChatGPT default, Gemini 4 Argon, Grok Voice 3.0, Meta Muse Spark Open and Muse Wonder, Reflection Beam, DeepSeek V4.1-Flash-1017, Amazon Nova 2.5 Sonic and NVIDIA Nemotron 3.5 Super VL. The full dated list is above, split by whether you can actually use each one.
What is still coming in October 2026?
NVIDIA Bolt E4B 4GB on 23 October is the only item with a firm date. Mistral has promised open weights for Mistral Large 4 by the end of the month. Everything else announced in October — Microsoft Maia 300, the MAI voice models, the Meta and Sierra Personal Agent Protocol — carries no ship date.
Was Mistral Large 4 released?
Not as a general release. It entered public preview on 6 October with 1T parameters and multimodal support. Several roundups list it as a release and date it 8 October; two per-day trackers both say 6 October and both label it a preview. The open weights are a separate promise for end of month.
Which October 2026 models have open weights?
Reflection Beam (501B, Apache 2.0), Meta Muse Spark Open (816B), DeepSeek V4.1-Flash-1017 (MIT), NVIDIA Nemotron 3.5 Super VL (120B), Aleph Alpha Kolibri-1 (Apache 2.0, 78B), Google EmbeddingGemma 2 (Apache 2.0), JetBrains Mellum2.1 (Apache 2.0), Cloudflare Clef-omni, Liquid d1-3B and d1-omni-600M, Scira Amos 1, Fish Audio Drama 3, Perplexity pplx-decider and H Company Holo4.
Why do October 2026 release dates differ between sites?
Because many lists are assembled from a weekly newsletter published on Thursdays, so every item inherits the issue date of 1, 8 or 15 October instead of the day it shipped. Claude Haiku 5.5 shows as the 8th in those lists and the 7th in per-day trackers; Grok Voice 3.0 as the 15th rather than the 14th.
Did OpenAI release new models in October 2026?
Yes: GPT-6.1 Sol on 1 October, GPT-6 becoming the ChatGPT default on the 8th, and GPT-Live-1.1 on the 15th, plus Dots, UltraFast mode, Codex cloud, Sign in with ChatGPT and the Decisions API. The three aurora-alpha model IDs reported on 18 October are not a release — there is no documentation or confirmation of any kind.