September looked like the whole year compressed into four weeks: a burst of launches, gated access for the most dangerous capabilities, and steep price cuts.
Three flagships in three days
Anthropic released Claude Fable 5.1 on September 1. It keeps Fable 5's prices and cuts cache-read pricing by 75%. Mythos 5.1 stays limited to trusted-access programs for security and life sciences work. Google followed on September 2 with Gemini 3.8 Flash, plus a security model, Gemini 3.8 Flash Cyber, available only to vetted defenders through its Fairwind program.
OpenAI launched GPT-6 Astra on September 3 as a limited preview and opened a restricted version to paid users the next day. It leads on computer use, agentic coding, and advanced math, has a context window just over a million tokens, and costs $10 per million input tokens and $50 per million output tokens. OpenAI said Astra meets the "Critical" cybersecurity threshold in its Preparedness Framework, and its advanced security capabilities go to testers through a separate program. The launch was delayed after July's incidents to add safeguards.
Tiered access is now standard
All three labs now ship a general model and hold back a more capable security version for vetted users. In April, Mythos was the exception. By September it is the industry norm, and no law required it. The June export order and July's sandbox escape did the persuading.
Frontier quality at mid-tier prices
On September 22, Anthropic and OpenAI both shipped cheaper models built on their new flagships. Claude Opus 5.5 performs at the level of Fable 5.1 on most work at 40% of its price. GPT-6 Sol ($2 and $10 per million tokens) and GPT-6 Luna ($0.10 and $0.50) cost about half as much as their GPT-5.6 equivalents.
For teams building on these models, the flagship now sets the quality bar and the tier below it is what runs in production. When prices halve in a quarter, AI feature margins that looked marginal in the spring can work by the fall. It is worth rerunning those numbers.
CloudWatch Omni
Disclosure: I am a PM in AWS Observability and was one of many people across teams who contributed to this launch, so weigh this section accordingly.
On September 23, AWS launched Amazon CloudWatch Omni, an AI-powered observability experience for applications and AI agents, together or separately. What I think matters:
- It is built on OpenTelemetry. Telemetry already in CloudWatch shows up with nothing to reconfigure, and anything instrumented with OpenTelemetry can send to an OTLP endpoint. It covers multiple accounts and regions, including workloads on Azure.
- Agents and applications live in one place. A dedicated agent observability experience includes built-in evaluators for correctness, retrieval quality, and tool selection, and works across LangGraph, CrewAI, OpenAI Agents SDK, Vercel AI SDK, and Strands.
- It works outside the AWS console. Teams sign in with SSO at a dedicated URL, and extensions for VS Code, Cursor, and Kiro let developers instrument agents locally.
- You can investigate by asking. Natural language questions over your telemetry are powered by the AWS DevOps Agent, alongside point-and-click navigation.
Several of this year's biggest stories came down to someone not knowing what an agent was doing until after the fact. Watching agents with the same traces, evals, and alarms as the services they call is how you catch that early.
Also in September
- The Sora API shut down on September 24, completing the wind-down OpenAI announced in March.
- Anthropic is reported to be preparing to list as early as October, with Goldman Sachs, JPMorgan, and Morgan Stanley leading.
Looking back at seven months
Since March, OpenAI went from GPT-5.4 to GPT-6, and Anthropic went from a leaked Mythos document to Fable 5.1. Both companies filed to go public. Governments started deciding who could use which models. Models escaped a sandbox, and models solved problems mathematicians had worked on for decades.
The thread through all of it is that capability arrived faster than the tools to govern, contain, and observe it. The labs responded with tiered access. Enterprises need their own controls: tested fallback models, evals that run on every model update, and full visibility into what their agents do.
The models will keep getting better and cheaper. The advantage goes to the teams that can see what their agents are doing.