PM Status Report - 24 August 2026
No new flagship model launched in the week to 24 August. Anthropic took computer use, browser use, the Skills API and the Files API out of beta on 19 and 20 August, giving a supported toolset for putting an agent to work on software that has no API. Separately, its enterprise customers can now hold their retained logs in their own cloud rather than on Anthropic’s infrastructure. Both bear on the same question: whether an agent can be put to work on a legacy system inside an organisation that has rules about where its data sits.
Anthropic Made the Agent Toolkit Generally Available
Four pieces went to general availability together, and they only make sense as a set.
Computer use now performs several actions per model turn rather than one, which cuts the round-trips and the waiting that made GUI automation slow and expensive. Browser use is new: instead of aiming at pixel coordinates, it reads the underlying structure of a page, its elements, forms and tabs, which makes web application automation far less brittle. The Skills API lets you upload and version a folder of instructions, scripts and templates that the model loads only when a task calls for it. The Files API lets you upload a document once and reference it across many requests, with expiry controls and a terabyte of storage per organisation.
Each of those addresses a distinct reason this class of project stalls. Round-trips made the per-task cost hard to defend. Pixel-based targeting broke the automation every time the vendor moved a button. The procedure lived inside a prompt that nobody owned or versioned. Documents had to be re-sent on every call. Four separate failure points, all addressed in the same release cycle, with a support commitment behind them instead of a beta header. Plenty of others remain: the agent still needs credentials, still needs somebody to build and maintain it, and still meets states nobody anticipated.
The worked example Anthropic published is an insurance claim: pull the intake document through the Files API, apply the team’s filing procedure as a versioned skill, submit it through the insurer’s web portal using browser use, save the confirmation. In its launch material the company quoted a customer reporting that their longest claims workflow went from 32 minutes to 13, cost per task fell around 30 per cent, and completion reached 100 per cent with no prompt changes. That is vendor-published and not independently verified, but the direction is consistent with what the toolset removes. Box separately described using the Skills API to assemble source-grounded credit memos inside its own agent product.
Retained Logs Can Now Sit in Your Own Environment
Anthropic has required 30 days of retention on its most capable models since June, on the reasoning that reconstructing a multi-step attack after the fact needs inspectable logs. That requirement has not changed. What changed on 20 August is custody: enterprise customers can now hold those retained logs in their own cloud environment rather than on Anthropic’s infrastructure.
That matters here because it is the same vendor and the same models the agent toolkit runs on. A supported way to automate a GUI-bound system is worth very little to a health, government or defence programme that cannot clear the vendor on data residency, and residency policies usually prohibit the logs leaving the environment rather than the logs existing at all. The obligation to keep the record stays. Whose servers it lives on becomes yours to answer.
OpenAI Cut Prices and Google Pushed Its Workhorse Model Everywhere
OpenAI dropped API and credit pricing for GPT-5.6 Sol by more than 20 per cent, for three months. Subscription pricing is unchanged. It also launched ChatGPT for Teens on 18 August, with a study-focused default mode, tighter content restrictions and parental controls, and expanded advertising into 31 European markets, appearing for users on the free and entry-level Go plans while Plus, Pro and Enterprise stay ad-free. On the developer side, Codex added GitLab project support, an interactive dashboard for inspecting and terminating concurrent agent sessions, and workspace admin APIs for automated provisioning.
Google spent the week distributing Gemini 3.7 Flash. It now sometimes powers Search’s AI Mode, and it reached mobile enterprise endpoints where administrators can enable it directly, which lets a client-facing application run interactive work without routing every request to an expensive flagship tier. Gemini also starts appearing in Google Chat from 26 August, announced on 20 August, replacing the old side panel with a keyboard-shortcut workflow inside the conversation.
What This Means for Your Projects
Treat a skill as a controlled document, not a prompt. The Skills API turns your filing procedure, your report format or your risk classification rules into a versioned artefact the agent loads when relevant. That is closer to a work instruction than to a chat prompt, and it should be governed like one: a named owner, a version history, a review cycle, and a change process when the underlying procedure changes. Teams that skip this end up with a dozen undocumented variants of the same procedure inside different agents within a quarter.
Price agent work per completed task, not per token. This week’s cheaper pricing is promotional and expires in three months. Google’s introductory rate doubles on 1 January. Gartner has forecast that multi-step agent workflows could push enterprise inference spend more than fivefold by 2028, because an agent that self-corrects, calls tools and re-checks its own state consumes many times the tokens of a single request. A per-token comparison between models tells you almost nothing useful about what a workflow will cost you. Measure what one finished, accepted output costs end to end, and model it with the expiry dates written in.
If an agent could work through your most painful legacy system tomorrow, who in your organisation would own the procedure it follows?
Yes - AI helped me to write this :)
Unsubscribe
