| |
|
Best-AI.org
The Briefing
|
|
● INSIDE A MIGRATION CHECKLIST for GPT-6, the new Anthropic price ladder, and the model that retires on November 30.
|
|
The cold open
Seven days between two flagships. Read the migration notes.
On September 22 OpenAI released GPT-6 Sol and GPT-6 Luna. On September 29 it released GPT-6.1 Sol. Anthropic, in the same month, shipped Fable 5.1 on the 1st, Opus 5.5 on the 22nd and Sonnet 5.5 on the 28th. Google added Gemini 3.8 Flash on the 2nd.
It is tempting to read that as a leaderboard. It is a calendar. Each of those names is a string in someone's config file, and each one comes with a changelog line that says what stopped working.
GPT-6 Astra, released on September 3, removed the none reasoning effort and custom sampling controls. That line decides whether your swap takes one minute or one sprint.
|
The signals
|
| 01 |
GPT-6.1 Sol came seven days after GPT-6 Sol. GPT-6 Sol and Luna launched on September 22, GPT-6.1 Sol with multi-agent support in beta on September 29. |
|
| 02 |
Anthropic made a $4 and $20 Opus. Claude Opus 5.5 launched on September 22 at $4 and $20 per million tokens. Opus 5 is $5 and $25. |
|
| 03 |
Sonnet 4.5 has a retirement date. Anthropic announced on September 30 that it retires on November 30. |
|
|
IF YOU DO ONE THING →
Open the changelog for every model you call and read its breaking changes or migration notes. It takes ten minutes.
|
|
|
|
| 01 |
OpenAI
GPT-6 arrives in three steps, and the migration notes are the product
GPT-6 Astra was released on September 3. Its migration notes: no none reasoning effort, no custom temperature, top_p or logprobs, and tool calling requires the Responses API. On September 22 came GPT-6 Sol at $2 input and $10 output and GPT-6 Luna at $0.10 input and $0.50 output per million tokens, for prompts up to 272K input tokens. On September 29 GPT-6.1 Sol followed at $2 and $10, with multi-agent support in beta.
Around the models: GPT-Live-1 became generally available on September 10 at $0.05 per minute, billed per second, plus backend usage. On September 2 OpenAI split its errors: 429 with slow_down for traffic spikes, 503 with server_is_overloaded for temporary overload.
THE TAKE
Read the migration notes before the price. A model that rejects a parameter you set in 40 places is more expensive than any per-token difference. The split between 429 and 503 is small and worth an afternoon: your retry logic should treat them differently.
SOURCE OpenAI, API changelog
|
|
|
| 02 |
Anthropic
A new price ladder: Fable at the top, Opus 5.5 in the middle
Claude Fable 5.1 and Claude Mythos 5.1 launched on September 1 at $10 input and $50 output per million tokens, with cache reads at $0.25. Mythos 5.1 is for Project Glasswing participants. On September 22 Claude Opus 5.5 launched at $4 and $20, below Opus 5 at $5 and $25, with fast mode as a research preview.
Claude Sonnet 5.5 followed on September 28. Its release note does not print a price, so check the model page. It comes with breaking changes from Sonnet 5: thinking: "between_tools" replaces "disabled", and tool_choice types any and tool return 400 errors. On September 30 Anthropic deprecated Claude Sonnet 4.5, retiring on November 30.
THE TAKE
The interesting line is Opus 5.5 at $4 and $20: the newer model costs less than the older one. If your routing table still lists Opus 5 as the safe default, you may be paying a premium for caution.
SOURCE Anthropic, API release notes
|
|
|
| 03 |
Google
Gemini 3.8 lands in pieces, and the 2.5 models go quiet
Gemini 3.8 Flash reached general availability on September 2. gemini-3.8-live and an extended-thinking variant followed on September 15, and the 3.8 Flash text-to-speech models on September 22, with a Voices endpoint and more than 150 voices.
On September 18 Google limited access to the Gemini 2.5 models to users who had used them recently. The page says they are not deprecated and remain available until further notice, and recommends 3.5 Flash-Lite or 3.8 Flash for new projects. antigravity-preview-05-2026 shuts down on October 5.
THE TAKE
"Not deprecated, limited to recent users" is a deprecation notice that avoids the word. If you depend on 2.5 and have not called it lately, test it this week.
SOURCE Google, Gemini API changelog
|
|
Hype check |
OUR OPINION, LOUDLY |
|
| ▲ |
Underrated
Opus 5.5 priced below Opus 5
A newer model that costs less than its predecessor is worth a second look. Check your routing table before you assume the old default is the safe one.
|
| ▼ |
Overrated
GPT-6.1 Sol, seven days after GPT-6 Sol
Fast iteration is not the same as progress you can build on. OpenAI fixed an image-encoding bug in GPT-6 Sol and Luna on September 25, three days after launch. Let the first bug list settle before you migrate.
|
| ● |
Rated correctly
The 429 versus 503 split
Dull and correct. A client that treats a spike and an overload alike will back off wrongly in one of them.
|
|
Receipts |
GRADING OUR OWN HOMEWORK |
"Do not roadmap ultrafast mode; it is in limited preview."
US, STORY 02, THE BRIEFING AUGUST 2026
On September 29 OpenAI added ultrafast mode for GPT-6 Astra in the Responses API, with US data residency only; the entry does not call it a preview. Verdict: overtaken for GPT-6 Astra, still open for GPT-5.6 Sol. If you need EU residency, the advice holds.
|
|
The number
7 days
Between GPT-6 Sol and Luna on September 22 and GPT-6.1 Sol on September 29. Both list the same $2 input and $10 output price for prompts up to 272K tokens.
SOURCE OpenAI, API changelog
|
|
One thing to try
The migration check, 20 minutes
| 1 |
List every place your code sets temperature, top_p, logprobs or a reasoning effort of none. GPT-6 Astra does not accept them. |
| 2 |
Check which of your tool calls still use an API other than the Responses API. |
| 3 |
Search for Claude Sonnet 4.5 model strings. They stop working on November 30. |
| 4 |
If you plan to move to Sonnet 5.5, check for tool_choice of type any or tool, and for thinking set to "disabled". |
| 5 |
Write the new model names into one config file, with an owner and a date to re-check. |
|
|
If you read one thing
OpenAI changelog, September 3
The GPT-6 Astra entry lists what the model no longer accepts. Read the migration notes before you read the benchmarks.
READ THE CHANGELOG →
|
|
|
Built something worth finding?
Submit it for free review. Add our badge with a normal followed link and we gift you Fast-Track: a decision within 48 hours after we confirm the link.
|
|
|
That is September. Go and read the migration notes for the model you use most.
Think we called one wrong? Excellent. Reply and tell us which, with a source. We will grade ourselves in public.
THE BRIEFING / SEPTEMBER 2026 / BEST-AI.ORG
|