Between September 3 and October 7, 2026, OpenAI, Anthropic, xAI and the major open-model labs shipped 16 notable releases. OpenAI completed its GPT-6 family (Astra, Sol, Luna and GPT-6.1 Sol), Anthropic released all three Claude 5.5 models (Opus, Sonnet and Haiku), xAI shipped Grok 4.7 and a new transcription model, and DeepSeek, H Company, Google, Alibaba's Qwen team, Cohere, Ai2 and Mistral released or previewed open-weight and API models. This page lists each release with its date, type and license, and links to a dedicated guide for each.
| Date (2026) | Model | Lab | Type | Weights and license |
|---|---|---|---|---|
| Sep 3 | GPT-6 Astra | OpenAI | Frontier model | Closed |
| Sep 10 | DeepSeek-V4.1-Flash | DeepSeek | Open-weight model and API | MIT |
| Sep 17 | grok-voice-transcribe-2.0 | xAI | Speech-to-text model | Closed |
| Sep 18 | Qwen3.8-Omni-Flash | Alibaba Qwen | Omnimodal model via API | No weights release announced |
| Sep 20 | Qwen-Image-2.1 | Alibaba Qwen | Open-weight image model | Qwen Research License (non-commercial) |
| Sep 21 | Grok 4.7 | xAI | Frontier model | Closed |
| Sep 22 | Claude Opus 5.5 | Anthropic | Frontier model | Closed |
| Sep 22 | GPT-6 Sol and GPT-6 Luna | OpenAI | Mid and low-cost tiers | Closed |
| Sep 28 | Claude Sonnet 5.5 | Anthropic | Mid tier | Closed |
| Sep 28 | Holo4 and Holotron4 Nano | H Company | Open-weight computer-use models | Apache 2.0, CC BY-NC 4.0 or NVIDIA Open Model Agreement, by model |
| Sep 29 | GPT-6.1 Sol | OpenAI | Mid tier (newer successor to GPT-6 Sol) | Closed |
| Sep 30 | Embed 5 Pro and Fast | Cohere | Embedding models | Paid API; self-hosting supported |
| Oct 2 | AstaBrief 8B | Ai2 | Open-weight scientific report model | Apache 2.0 |
| Oct 6 | Mistral Large 4 | Mistral AI | Public preview; weights promised for end of October | License not yet named |
| Oct 6 | EmbeddingGemma 2 | Open-weight embedding model | Apache 2.0 | |
| Oct 7 | Claude Haiku 5.5 | Anthropic | Low-cost tier | Closed |
Dates and licenses are taken from each vendor's announcement, documentation or model card; each guide links its sources.
Three models now list at $2 per million input tokens: Claude Sonnet 5.5 and GPT-6.1 Sol in the middle tier, and xAI's frontier model Grok 4.7. Sonnet 5.5 and GPT-6.1 Sol also match at $10 output and $0.10 for cached input, while Grok 4.7 charges less for output ($6) and more for cached input ($0.50). At the low end, GPT-6 Luna and Claude Haiku 5.5 both list at $0.10 input and $0.50 output. Sources: Anthropic's pricing documentation, OpenAI's model pages for GPT-6.1 Sol and GPT-6 Luna, and xAI's release notes, all checked October 10, 2026. When headline prices match, the differences that decide the bill move into the fine print: caching rules, long-context surcharges and how many tokens each model uses per task.
The fine print varies widely:
A workload that routinely sends 150,000-token prompts would pay standard rates on Opus 5.5, Sonnet 5.5 and GPT-6.1 Sol, but surcharged rates on Haiku 5.5. The 272K threshold article and the Claude Fable 5.1 vs GPT-6 Astra comparison cover these structures in more depth.
All three Claude 5.5 models change behavior that existing code may rely on. Anthropic's migration guides (Opus 5.5, Sonnet 5.5, Haiku 5.5) list forced tool use returning 400 errors on Opus 5.5 and Sonnet 5.5 (Haiku 5.5 accepts a forced tool choice, but its response then starts with the tool call and has no thinking block), thinking that cannot be fully disabled (Opus 5.5) or is on by default (Sonnet and Haiku), thinking blocks tied to the conversation, and a newer computer-use toolset required on the Claude API and Google Cloud. Haiku 5.5 also uses a tokenizer that, per Anthropic, produces about 30% more tokens than Haiku 4.5 for the same text. Each Claude 5.5 guide includes a migration checklist.
The open-weight releases in this period span MIT (DeepSeek-V4.1-Flash), Apache 2.0 (EmbeddingGemma 2, AstaBrief 8B and Holo4-35B-A3B), non-commercial licenses (Holo4-27B under CC BY-NC 4.0 and Qwen-Image-2.1 under the Qwen Research License), a vendor agreement (Holotron4 Nano under NVIDIA's terms) and, for Mistral Large 4, a license not yet named for weights not yet released. Qwen-Image-2.1 is described by its makers as open source even though its license bars commercial use. Read the license of the exact model you plan to deploy. Self-hosting is not the same as open weights either: Cohere's announcement says Embed 5 supports self-hosting and private deployments, but it does not announce a public weights release.
Most announcements in this series include vendor-run benchmarks, and many compare against competitor models that had newer versions within days. OpenAI's GPT-6 Sol and Luna comparisons cite Claude Opus 5 and Claude Fable 5 or 5.1, depending on the benchmark; xAI's Grok 4.7 table uses GPT-5.6 Sol and Fable 5.1; Anthropic's Opus 5.5 announcement uses GPT-5.6 Sol. Treat cross-vendor charts as a reason to run your own evaluation, measured as success rate and cost per completed task, not as a ranking.
Two significant open-weight families shipped shortly before September and remain relevant reference points:
Product features, pricing changes and access expansions are covered inside the relevant guides rather than as separate entries. They include OpenAI's improved prompt caching (September 22) and GPT-6.1 Sol Ultrafast mode (October 8); Anthropic's Sonnet 5.5 cache-read price cut (October 7); and xAI's Grok 4.7 Fast variant. Quantized and repackaged versions of the open models listed here are not separate releases.
What were the biggest AI model releases in September and October 2026?
OpenAI's GPT-6 Astra and GPT-6.1 Sol, Anthropic's Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, xAI's Grok 4.7, Cohere's Embed 5, and open-weight releases including DeepSeek-V4.1-Flash, Holo4, EmbeddingGemma 2 and AstaBrief 8B, plus a preview of Mistral Large 4.
Which of these models have open weights?
DeepSeek-V4.1-Flash, Holo4 and Holotron4 Nano, EmbeddingGemma 2, AstaBrief 8B and Qwen-Image-2.1 have downloadable weights, under the licenses listed above. Mistral Large 4 weights are promised for the end of October 2026.
Which models cost $0.10 per million input tokens?
GPT-6 Luna and Claude Haiku 5.5 (for prompts up to 100,000 tokens), per their vendors' pricing pages as of October 10, 2026.
Discover more content:
