This is issue #1 of a monthly digest covering what actually changed across OpenAI, Anthropic and Google's model APIs: new releases, deprecations, pricing moves, and breaking API changes. No speculation, no roadmap rumors, just what shipped and what you need to do about it if you're building on these APIs. Future issues will follow this same format: one section per vendor, newest and most consequential changes first.
A note on scope: this issue leans heavier on late September than October, because that's where the verifiable material is as of this writing (September 23). Where I couldn't confirm a claim against an official source, I left it out rather than guess.
OpenAI
GPT-6 Astra shipped September 3–4. OpenAI's newest flagship rolled out to approved users on September 3 with general availability the next day, after the company delayed the release earlier in the summer to add safeguards following a string of unsanctioned agent-driven cyberattacks in July. Astra is positioned as state of the art on computer use, browsing, software engineering, and cybersecurity tasks: this is the model to be targeting for anything agentic if you're on OpenAI.
The GPT-6 family expanded September 22 with GPT-6 Sol and GPT-6 Luna joining Astra, giving OpenAI a tiered GPT-6 lineup rather than a single flagship release. If you were building against gpt-5.6-* model IDs, budget time to evaluate the GPT-6 tier that matches your latency/cost/capability needs: Sol and Luna are presumably the mid- and lower-cost tiers, mirroring how the 5.x family was structured, though check current pricing directly before assuming the mapping holds exactly.
Evals platform is being retired: mark your calendar. OpenAI notified developers back on June 3 that the Evals platform, Agent Builder, and reusable Prompt objects are being deprecated. The dashboard and Evals API go read-only on October 31, 2026 and shut down entirely on November 30, 2026. If you have evals you rely on, this is the month to export them, we wrote up the full migration path to Promptfoo in a dedicated post.
A batch of older models retire October 23, 2026: o1, o1-pro, o3-mini, and gpt-3.5-turbo are scheduled for shutdown. If anything in your stack still references these IDs directly, this is the deprecation to act on this month: OpenAI's guidance points o1/o1-pro/o3-mini traffic toward gpt-5.6-sol (with reasoning.mode: pro for the o1-pro case) and gpt-3.5-turbo toward gpt-5.6-terra.
Anthropic
Claude Opus 5.5 launched September 22, the first model in a new 5.5 line. The headline claim is that it matches Claude Fable 5.1's performance on most work at roughly 40% lower cost to run, with output speed up over 30%, current self-serve pricing is $4/$20 per million input/output tokens (a 20% cut from Opus 5), with cached reads down to $0.20/MTok. On Anthropic's own benchmarks, Opus 5.5 edges out Fable 5.1 on agentic coding, computer use, and chart/visual reasoning specifically.
Claude Fable 5.1 (and Mythos 5.1) shipped earlier the same month, around September 1, ahead of Opus 5.5, so if you're only now catching up, you've actually got two new model generations to evaluate, not one.
Breaking change: forced tool choice is gone on both new models. Claude Opus 5.5 and Claude Fable 5.1 both reject tool_choice: {"type": "any"} and named-tool forcing with a 400 error: only auto (default) and none are supported. This broke several frameworks that force tool choice to guarantee structured output (multiple LangChain and crewAI integrations needed patches). If you're upgrading from Opus 5 or an earlier Claude generation and rely on forced tool choice anywhere, this will surface as a hard failure, not a quiet behavior change, check our tool-calling comparison across providers for the strict: true + prompt-instruction pattern that replaces it.
A scheduled price increase for Claude Sonnet 5 did not happen. Sonnet 5 launched at introductory $2/$10 per million input/output tokens through August 31, with a bump to $3/$15 slated for September 1. Anthropic held the introductory pricing instead: worth knowing if you'd budgeted for the increase.
Gemini 3.8 Flash released September 2, Google's third Flash-tier release in about six weeks, which tells you something about their release cadence right now. It posts a meaningful jump on agentic coding and long-horizon benchmarks (90.8% on Terminal-Bench 2.1, up from 81.6% on 3.7 Flash) and beats several larger frontier models on long-horizon coding evals. Pricing holds at $0.75/$3.75 per million input/output tokens through the end of 2026, then steps up to $1.50/$7.50.
Gemini 3.5 Flash-Lite (the fastest, cheapest tier in the current lineup) shipped July 21 and remains the recommended default for high-volume, latency-sensitive workloads where you don't need Flash-tier reasoning depth.
Two smaller cleanups worth knowing about if you're on Gemini specifically: gemini-omni-flash-preview is deprecated as of September 30, 2026, and the antigravity-preview-05-2026 agent model shuts down October 5, replaced by a new Antigravity Agent release that also changed its local tool-handling parameter conventions (PascalCase instead of snake_case, line-range file edits instead of full rewrites), a genuine breaking change if you built directly against the preview's tool interface rather than going through a wrapper.
What to actually do this month
If you take one action item from this issue: check whether anything you have in production references o1, o1-pro, o3-mini, gpt-3.5-turbo, gemini-omni-flash-preview, or antigravity-preview-05-2026 by exact model ID, and start the migration now, October has two separate shutdown dates in it (the 5th and the 23rd) and a third looming right after (Evals going read-only on the 31st). Everything else this issue covers is either a new option worth evaluating (GPT-6 Sol/Luna, Opus 5.5, Gemini 3.8 Flash) or a syntax detail to fix before it breaks your next deploy (Claude's forced tool-choice removal).
We'll be back with issue #2 in early November, if there's a specific vendor or category of change (pricing, safety policy, SDK breaking changes) you want more depth on, that's useful signal for what to prioritize next time.
