Two flagship models launched this week at the identical price, and neither of them changes a single decision on your Amazon account. That is the loud half. The quiet half is a default that moved on Thursday, a video model string that stops working in 24 days, a desktop agent that now ships with the software to open your flat file, and a seller conference in 16 days where Amazon historically announces the AI defaults it flips next.
Last week the theme was tools growing memory, keys and logins. This week the models got the headlines and the plumbing got the fuses. Same rule as always: the announcement is not the event, the default is.
1. GPT-6 Astra shipped at $10/$50, and Codex quietly made it your default
OpenAI released GPT-6 Astra on September 3 as a limited preview for trusted partners, then pushed a restricted version to paid ChatGPT users the following day, with API, Azure and Bedrock access rolling over the coming days (OpenAI release notes, Simon Willison, Sept 3). API pricing is $10 per million input and $50 per million output, roughly 2.5x GPT-5.6 Sol's $4/$20. The model string is gpt-6-astra.
Three operator facts, none of which are the benchmark scores.
First, the public version is restricted. It refuses certain prompt classes, cybersecurity being the named one. That will not touch a product bullet. It does mean a refusal is now a normal output on a paid tier, and a batch job that treats a refusal as a blank field will write blanks into a catalog without erroring.
Second, the Codex CLI at 0.153.4 made Astra the bundled default when no model is explicitly configured. If you built Codex skills this summer and never wrote a model string into them, they changed model this week and got 2.5x more expensive per call without anyone touching them. That is the floating-default problem again, and it is the third time this year I have written that sentence about a different vendor.
Third, on the same day Astra rolled out, OpenAI opened an incident for elevated errors across ChatGPT and Codex. If a catalog job depends on somebody else's launch day going smoothly, launch day is the wrong day to run it.
The 2.5x number is not a reason to panic. Catalog-scale work does not run on the flagship tier. It runs on Sonnet 5 at $2/$10 and on Flash-class models, and nothing about that tier moved. Astra is for the judgment call at the top of the stack, if it is for you at all.
2. Claude Fable 5.1 launched at the same price, and the Claude Code default moved too
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1 (Anthropic news). Fable is the generally available model and carries additional safety measures. Mythos is the same underlying model available only to approved organizations. Astra's $10/$50 was reported as matching Fable's pricing, so the two flagships now sit on an identical rate card, which tells you what the flagship tier costs in September 2026 and nothing about what your bullet-writing job costs.
Two things from the Claude Code changelog matter more than the model (changelog).
On September 3, version 2.1.260 changed the default model for seat-based Enterprise subscriptions to Opus 5. A default moved, on an Enterprise plan, in a point release. If your agency runs client work on Enterprise seats without a pinned string, the model under that work changed on Wednesday.
The same release added managedMcpServers, letting an organization push HTTP MCP servers to every user. September 4 added an "Organization policy" line to /status that says why a policy could not be loaded. Read those together: the org-level controls I said last week live on the wrong plan are getting more specific, and there is now a visible line that tells you when the policy is not being applied at all. A shop that says "we have an org policy" should be able to show you that line reading clean.
One more thing nobody will mention. Anthropic's text watermark applies to models launching on or after August 2. Fable 5.1 launched September 1. Every word it writes into your A+ carries a mark you cannot see and cannot strip, held by a company that is not you. Same as three weeks ago, more so now.
3. Gemini 3.8 Flash went GA, and a video string you may be running dies September 30
Google shipped three things (Gemini API changelog). Agentic video understanding on September 1, claiming up to 88% fewer tokens for long-form video. Gemini 3.8 Flash generally available on September 2, positioned for long-running agents. And the line that actually matters: the gemini-omni-flash-preview endpoint is deprecated on September 30, with migration to gemini-omni-1.1-flash.
If you built a listing-video pipeline in May or June, when Omni's price collapse made it worth trying, there is a good chance it is on the preview string. That string has 24 days left. The failure is loud, which is the good version, but it lands in the first week of October with Prime Big Deal Days live and a creative freeze in place. Migrate it this week. Then re-run your six-clip golden set on the 1.1 endpoint, because a new model string on a video pipeline is a new pipeline until you have frozen a frame at second 27 and read the label.
Same column, different vendor: xAI retires grok-imagine-image-quality on November 2 and redirects requests to grok-imagine-image-2.0 with quality set to low. A redirect that silently lowers output quality is worse than a shutdown, because nothing errors. If any image step runs on that string, the output degrades on a Monday and nobody gets an email.
4. Codex now ships with LibreOffice, and the flat file is the write path
Simon Willison found, while clearing cache, that the Codex desktop app bundles a 1.7GB runtime with LibreOffice, Python, Node, Poppler and git (Sept 1). That is a developer footnote everywhere except in this business.
The single most consequential file in an Amazon operation is an .xlsx. Inventory loader, price and quantity file, category listing template. A desktop agent that can natively open, edit and save a spreadsheet is a desktop agent one drag-and-drop away from producing a flat file that uploads. Combine that with last week's ChatGPT Work scheduled tasks and a browser login, and the path from "summarize my inventory report" to "here is the corrected file, uploaded" no longer requires anyone to write code.
I am not saying agents will run your catalog. I am saying the tool to edit the file that runs your catalog now arrives bundled, on a personal seat, with no admin control. The rule from last week holds: no scheduled task logs into Seller Central, and no flat file goes up without a human reading the diff.
5. Amazon Accelerate is in 16 days, inside your freeze
Amazon Accelerate runs September 22 to 24 in Seattle, 150-plus sessions across 11 breakout topics, speakers now announced on the Seller Forums. Not an AI story on its face. Include it because the last two years of Amazon's seller-facing AI defaults were announced there: Project Amelia in 2024, the agentic Seller Assistant in 2025. Amazon also confirmed this week that the 75-character title policy applies to its own retail listings, with more than 984 million titles updated since June. That rewrite was an AI default too.
Whatever gets announced on the 22nd will be on by default for some share of sellers before Black Friday. It will arrive as "no action required." Diary the 22nd, read the session titles now, and decide before the keynote what you are willing to have auto-published on your catalog during peak.
What I would do this week
- Grep every skill, script and vendor config for a missing model string. Anything without one changed model this week on at least one platform. Pin it.
- Find the Omni preview string if you have one. Migrate before the 30th, re-run the golden set, write the new string in the inventory with a date.
- Add a "redirects to" column next to "retirement date." A shutdown fails loudly. A redirect fails quietly.
- Ask your agency one question: which plan are your seats on, and does
/statusshow an organization policy loaded. A screenshot answers it. - Read the Accelerate agenda on the Seller Forums and put the 22nd on the calendar of whoever owns your catalog freeze.
What I would ignore
The benchmark war between Astra and Fable. Perfect scores on ARC-AGI say nothing about whether a model writes a bullet that survives a compliance check. Still the only benchmark that bills you when it is wrong, still built from twenty of your own SKUs.
The "GPT-6 will replace your agency" cycle and its mirror image. What shipped this week was a price and a default. Neither one reads a return comment.
Any vendor quoting Astra's or Fable's capabilities as a reason to raise your retainer. The flagship tier got no cheaper and no more expensive relative to each other, and the person checking the output did not get cheaper either.
The urge to migrate your stack to whichever flagship won the week. Migration is a config string. Re-validating against your own SKUs is the expensive part, and nine weeks from peak is the wrong time to spend it.
Two flagships at one price, a default that moved in a point release, a fuse set for the 30th, and a spreadsheet editor bundled into a chat app. Nothing shipped that you can put on a listing. Plenty shipped that can touch one while you are not looking.