Fourteen AI models shipped in September 2026. Cache pricing matters more.
Key takeaways
- September 2026 saw 14 new AI models from 10 providers, a cadence faster than most teams can finish evaluating one.
- Anthropic shipped Claude Fable 5.1 on 1 September with cache reads at a quarter of the previous price, which matters more than any benchmark for agent workloads that re-read context on every step.
- OpenAI moved ChatGPT to the GPT-6 Astra engine on 14 September: 1.05M context, 128K output, 10 dollars per million input, 50 per million output, cached reads at 1 dollar.
- The comparison worth making is cost per task at your own context length, not a leaderboard position.
Fourteen new AI models from ten providers in one month. The count is the least interesting thing about it. The most consequential change in September was a price on cached tokens.
Anthropic shipped Claude Fable 5.1 on 1 September, positioned for coding, knowledge work and long-running agent tasks, with cache reads at a quarter of the previous price. For agent workloads that number does more than any benchmark in the announcement. Agents re-read the same context on every step, so cached input is where the bill actually accumulates.
What the three frontier releases are each for
Google shipped Gemini 3.8 Flash on 2 September, about three weeks after 3.7 Flash, with improvements to software engineering and multi-step reasoning at below frontier pricing. It also shipped Gemini 3.8 Flash Cyber, aimed at autonomous vulnerability discovery.
OpenAI moved ChatGPT to the GPT-6 Astra engine on 14 September, with a 1.05 million token context window and 128K output, priced at 10 dollars per million input tokens and 50 per million output, with cached reads at 1 dollar.
Lined up, those are three different products rather than three entries on a ranking. One is priced for long agent runs, one for volume at lower cost, one for very large single contexts. The comparison that answers a real question is cost per task at your context length.
The cadence has outrun the announcements
Fourteen releases in thirty days is past the point where reading launch posts keeps anyone current. Gemini 3.8 Flash arrived three weeks after 3.7 Flash, a shorter gap than most teams take to finish evaluating a model they have already adopted.
The practical response is to stop tracking releases and start tracking your own numbers against them. Open weight options belong in that calculation, with releases like StepFun's Step 5 preview changing what is worth paying for at all, and local inference through vLLM, Ollama and similar runtimes removing per token cost entirely for workloads that fit in available memory.
What to watch
Whether cached input pricing keeps falling. It is the least discussed number on a model announcement and increasingly the one that decides whether an agent product has a workable margin. OpenAI's projected 278 billion dollar cash burn through 2030 is a reminder that current pricing reflects a strategic choice rather than the cost of serving the request.
