Three of these are safe to ignore

Welcome to the Debrief, the Tuesday half of The AI Debrief. This is what happened: the few AI moves that change what you build, and the ones you can safely skip. Friday is the Toolkit, where we hand you something to use.

Here's what we're getting into this week. A stack of things you may already depend on stopped being experimental on August 19 and 20, and exactly one of them will not run on your existing request shape.

Anthropic's agent stack went GA on two days, not one

On August 19 the Files API and Agent Skills left beta. On August 20 computer use followed, along with a new browser tool and a major version of the Python SDK. Most of the coverage treated it as one launch. It reads better as four separate migrations with four separate risk levels.

Computer use is now computer_toolset_20260801, and it replaces the beta computer_20251124. This one is not a header swap. Anthropic's own release notes say upgrading “changes the request shape and tool handling,” and they publish a dedicated migration guide for it. What you get in exchange is the thing that actually matters for cost: batch actions, so the model can take several actions per turn instead of one per model call. zoom is on by default now, and there is per-member configuration through configs.

The other three are gentler. The Files API dropped the files-api-2025-04-14 header, added file expiry through expires_in_seconds and expires_at, added pagination, and raised the media ceiling from 100 to 600 images or PDF pages per request on the 1M-token context window. Agent Skills dropped skills-2025-10-02 and picked up a /v1/skills endpoint that loads through the container parameter. Both keep working unchanged if you still send the old header, which is the detail worth holding onto.

The Python SDK is the one that will surprise people, because nothing announces it. v1.0 moved from httpx to httpx2, requires Python 3.10 or later, removed the legacy Text Completions API, and removed temperature, top_p and top_k from Messages methods. Async .with_raw_response results now need await response.parse(). And AnthropicBedrock raises an error when no AWS region is configured instead of quietly defaulting to us-east-1, which is a better behaviour and also a behaviour change in something that used to work.

All of it runs on Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8.

Read the release notes →

A browser tool that reads the page instead of looking at it.
browser_toolset_20260801 is new, and it is the most interesting thing in the whole release. It drives a hosted browser by reading the accessibility tree - elements, forms, tabs - rather than hunting for pixels. It adds element references, form input, tab management and download reporting. The difference is not cosmetic. Pixel-targeted automation breaks when a button moves four pixels or a font renders differently; structure-targeted automation does not. That is roughly the line between a demo that works in a video and something you can leave running unattended. Friday's Toolkit has a way to try this idea for nothing.

Data retention quietly turned into a contract question.
OpenAI published Private Safety Processing on August 19. The mechanism: your content stays either on infrastructure you control, or on OpenAI servers encrypted with keys you hold, and OpenAI states its own personnel do not have a copy. When something is flagged, what comes back is “a narrowly defined signal indicating the type of activity involved,” not the prompt and not the response. Named design partners are Glean, Databricks, Abridge and Microsoft. Full rollout is September, with a technical white paper. Pricing is not disclosed. One caveat, because a lot of coverage is running it the other way: several outlets frame this as a direct shot at a competitor's retention policy, but OpenAI's own page says nothing about competitors at all. Treat the rivalry angle as reported, not stated. What is solid is the direction. If you sell into health, legal or finance, “which model” is drifting from a benchmark question toward a procurement one.

A DeepSeek update, not a new story.
We covered the V4-Pro pricing restructure on August 18, so this is one line rather than a section: since August 23, all day Saturday and Sunday count as off-peak. If you already moved batch work to dodge the peak windows, you can now park the heaviest weekly job on a weekend at half rate. That is the entire change.

Unitree's IPO, and the “the robots are here” week it produced. The stock debuted on Shanghai's STAR Market on August 18, popped 629% intraday and closed up 487%, landing near $53B. The underlying business earned roughly ¥278M on ¥1.7B of 2025 revenue and shipped about 5,500 humanoids. That is somewhere near 1,300 times earnings and 210 times sales. It was priced at about $9B on August 6, which means the entire re-rating happened in retail order books over twelve days, on an offer oversubscribed thousands of times over. Nothing about what the robots can do changed in that window. It is a genuinely interesting market story and there is nothing in it you can build on.

The one thing worth doing this week takes about thirty seconds: search your codebase for computer_20251124. That is the only item here whose request shape has already changed. Everything else is a header you can drop when convenient, or a version pin that will introduce itself later. Honest worst case if you ignore all of it - an agent that used to work starts throwing shape errors on a Tuesday, and you spend an hour finding out why.

The 30-day challenge we opened on Friday is on Day 5, and it is built to be caught up on.

Free, no card. Day 1 takes 20 minutes.

That's the Debrief.

See you Friday for the Toolkit.

From Drew and The AI Debrief team

P.S. What's still pinned to a beta header inside something you own? Hit reply and tell us. We read every one.

P.P.S. New here, or skipped one? Every past edition is in the archive.