AI in September 2026: 7 new models at a glance

September was one of the busiest months of the year in AI: five major labs shipped new flagships or cheaper “fast” models within a few weeks of each other. The common theme is clear: models that run longer, more complex tasks on their own, at lower prices.

In brief

• Claude Opus 5.5 (Anthropic, Sep 22): Anthropic’s new flagship, now $4 input / $20 output per million tokens.
• GPT-6 Astra (OpenAI, Sep 3): OpenAI’s new-generation top model, strongest in computer use and software engineering.
• GPT-6 Sol and Luna (OpenAI, Sep 22): lighter, cheaper siblings of Astra.
• Gemini 3.8 Flash (Google, Sep 2): a new version of Google’s fast, low-cost workhorse.
• Muse Spark 1.3 (Meta, Sep 2): Meta’s agent- and coding-focused update.
• Grok 4.7 (SpaceXAI, Sep 21): a bigger base model at the same price.
• Open models: DeepSeek V4.1 Flash, Kimi K3 and Qwen3.8-Max keep the open-weight front busy.

In independent tests Claude Opus 5.5 currently leads, followed by Claude Fable 5.1 and GPT-6 Astra. See the live ranking on our AI Models page.

Model by model

Claude Opus 5.5. Anthropic says it performs at the level of Claude Fable 5.1 on most work and, at default settings, costs about 40% less than Opus 5 on typical workloads, with output speed up more than 30%. Prices dropped 20% ($4 / $20) and cache reads fell to $0.20. It is also on AWS, Google Cloud and Microsoft Azure; Sonnet 5.5 and Haiku 5.5 are expected within weeks.

GPT-6 Astra. OpenAI calls it its “most intelligent and aligned” model yet and says it is state of the art in computer use, browsing, software engineering, cybersecurity and science. It is rolling out in ChatGPT and Codex.

GPT-6 Sol and Luna. On OpenAI’s internal factuality evaluation, Sol makes roughly half as many mistakes as its predecessor. API prices are 50% below GPT-5.6’s promotional pricing, and Luna is available to free users too.

Gemini 3.8 Flash. Google calls it its “most intelligent workhorse model”, with clear gains over 3.7 Flash in software engineering and multi-step reasoning. It is in the Gemini app, AI Studio and the Gemini API. Google also announced 3.8 Flash Cyber for security researchers; the Chrome Security team reported 2.6x more correct vulnerability patches than much larger commercial models.

Muse Spark 1.3. Meta says it sustains long, multi-step work better, asks clarifying questions when needed, and uses about 20% fewer tool calls and 25% fewer tokens than version 1.2. Pricing is unchanged ($1.25 / $4.25) with a 1M-token context window.

Grok 4.7. SpaceXAI’s new model sits on a 2.1-trillion-parameter base model, 40% larger than Grok 4.6 (1.5 trillion), at the same price ($2 / $6). The company says it checks its own work more carefully on long tasks and ships a rebuilt safeguard stack.

Open-weight models. DeepSeek published V4.1 Flash, which adds image understanding, in early September. The cloud version of Alibaba’s roughly 2.4-trillion-parameter Qwen3.8-Max launched in early August, with open weights promised. Moonshot AI’s Kimi K3 is another large open model from the same period.

Our take

Competition is now as much about what a task costs as about raw intelligence: price cuts or speed gains headline this month’s announcements from Anthropic, OpenAI and Google. For crypto, cheaper capable agents make on-chain automation and analysis more practical, which makes it easier for “AI x crypto” projects to show real usage.