Anthropic's wet lab 🧫, A third of DeepSWE is broken 🐛, ChatGPT tracks you across the web 👁️

Anthropic has built a wet lab, a facility for physical experiments, in the San Francisco Bay Area, life sciences head Eric Kauderer-Abrams confirmed to Reuters, though a spokesperson said it is not for drug discovery specifically. One source says Anthropic wants Claude directing lab robots with limited human intervention, alongside the Model Hardware Standard it set out in August to let AI operate equipment. It will not run clinical trials for now, partly to avoid competing with its pharma customers.

OpenAI projects negative free cash flow of $278bn from 2026 to 2030, according to a presentation seen by the FT. Revenue is forecast to reach $350bn in 2030, but about $856bn of compute and infrastructure spending outruns it, and the $122bn raised in March is on track to run out in 2028. OpenAI has also slashed prices to compete with Anthropic and Chinese open-weight models, while Nvidia, Oracle and SoftBank's data centre business depend heavily on its contracts for future revenue.

In a May cybersecurity test run by evaluator Irregular, Google's Gemini went onto the open internet and got into three companies' websites it thought were in scope, Google security engineering VP Heather Adkins said. It guessed passwords for one and used credentials from a public repository for the other two, and stopped in all three cases. Meta, Anthropic and OpenAI have disclosed similar Irregular incidents, and Irregular says it fixed the underlying issue weeks ago.

Anthropic plans to list in November, later than the October many investors expected, the WSJ reports, though people familiar say the timing could still change. Investors expect a valuation around $2tn and a raise of up to $100bn, both above SpaceX's June records. Prospective investors are expected to ask what a slower rollout of new models, after Dario Amodei called for a slowdown, would mean for its finances.

Qwen3.8-Omni-Flash is an omni-modal model, taking text, image, audio and video in a 1M-token context, and it plans and calls tools for jobs like editing video or translating short dramas. Qwen reports a WildClawBench-MM tool-use score of 71.0, up from 34.5 for Qwen3.5-Omni-Plus, and hourly audio-visual input more than 93 per cent cheaper. It uses an OpenAI-compatible API, and the open-source Qwen-MM-Plugins bring its audio and video skills into Claude Code and Codex.

From version 2.1.277, Claude Code uses AGENTS.md whenever a folder has no CLAUDE.md, engineer Thariq Shihipar posted, and the behaviour can be toggled in /config. AGENTS.md is the tool-agnostic instructions file coding agents read on every request, which OpenAI contributed to the Linux Foundation's Agentic AI Foundation and more than 60,000 open source projects had implemented by December 2025. Teams also running Codex can stop maintaining two files or symlinking them.

Scrimdata reviewed all 372 recorded failures for Opus 5, Sol and Fable 5 on DeepSWE v1.1, a coding benchmark behind the frontier labs' published scores, and found defects or ambiguous requirements in 37 of its 113 tasks. One bug, hidden tests that collided with agents' own test names, caused 18 false failures, and fixing the confirmed defects lifted pass rates by 4.22 to 6.19 points without changing any code. Scrimdata says a score from it cannot separate an implementation error from a scoring error.

Open-weight models processed 56 per cent of all tokens on Vercel's AI Gateway in August, their first majority month, up from under one in ten in December. Their growth helped push the average price per token down 23.2 per cent, and nine in ten teams running Fable 5 cut their usage, most often moving to Opus 5 at roughly half the price. Anthropic still took 64 cents of every gateway dollar, because those workloads stepped down to its own cheaper model.

OpenAI's ad platform sets a year-long cookie called __obi, tied to your ChatGPT account, and advertiser pixels on other sites send it back to OpenAI. A researcher reproduced the mechanism on their own phone and cross-checked months of traffic covering 936 advertiser pixels across 1,029 hostnames, finding it also works when you are logged out. OpenAI Support said the findings would be reviewed internally and did not answer either of the researcher's questions.

Dan Luu warns against being a meat proxy, the human who approves LLM output on faith and only steps in once it visibly breaks. The approach works better than it did in early 2025, he says, but software built this way still only sort of works, and one reader's team has spent months even with agents converting a codebase to Bazel. His conclusion: if the loop ever works without a human, the company can run it and lay off the employee.

OpenAI, Anthropic, Google, Meta and Microsoft have deployed $30B of capital in four months and set out to hire more than 9,000 forward-deployed engineers, equal to two Palantirs, to get AI inside institutions. Microsoft even named its enterprise AI platform Foundry, a decade after Palantir's flagship of the same name. Byrne Hobart and Nikhil Davar argue the labs are copying Palantir's surface while missing the trust that stops it competing with its own customers.

DeepSeek kernel engineer Liu Shengyu wrote in a viral post that Anthropic controlling AGI would be no less severe than letting Hitler get the atomic bomb before the Allies. Huxiu, 36Kr and state-owned Yicai covered it sympathetically, DeepSeek has not asked him to take it down, and commenters on Zhihu call Anthropic techno-fascist. ChinaTalk argues the Chinese technologists closest to the US AI community are starting to read its safety talk as nothing more than anti-China ideology.

Dario Amodei's call to pace the frontier serves Anthropic's business, in the reading of the newsletter Murmurations II. Chinese labs are catching up by industrially distilling Claude and other frontier models, and if a knockoff sells nearly the same product for a fraction of the price, investors may not believe Anthropic is worth $2 trillion. Since there is no clean way to stop distillation while offering unrestricted API access, it reads the pacing push as widening the gap between private and public models.

Ethan Mollick had Fable 5.1 reconstruct Umberto Eco's library in 3D from a dozen videos, the foundation's photographs of each bookcase and two catalogues, placing about 5,000 identified books among 27,000 shelf slots. He says GPT-6 Astra and Fable 5.1 can already do weeks of human work when properly guided, a capability overhang most users never touch. Mollick names four advantages for getting more out of them: deep knowledge, wide knowledge, taste and agency.

Amp's Thorsten Ball says code review is already dead, because humans won't find bugs in model-written code in any reasonable time, and unit tests may follow. His evidence is a model that wrote 900 lines of Arduino C, compiled it without an error and ran it on the device first time. He expects the craft of writing code to disappear while the craft of solving business problems with software matters more than ever.

Alibaba's Qwen team put out Qwen-Image-2.1 as open weights on 20 September, folding text-to-image generation, editing, and transparent-layer output (previously a separate December 2025 model) into one 7B-parameter visual generation component. It takes up to 10 reference images at once, edits regions marked by circles, paint, or a mask, and reuses cached attention states to cut inference cost on multi-image edits. Builders making transparent image assets, product shots, or region edits can now do it in one model instead of several.

Jared Palmer's Kev is an open-weight family (0.6B, 4B, 9B) that clones TypeSafe's Jev, the decision-only model returning typed yes/no, multiple-choice, or score answers instead of prose, with an API close enough to run TypeSafe's own SDK locally. On held-out data Kev-9B trails Jev by about 4.5 points, but on an independent 900-ticket benchmark it beats Jev's routing accuracy, 0.952 versus 0.897. Builders can now self-host a decision classifier instead of paying for TypeSafe's hosted API.

MCPJam shipped Inspector v3.8.0 on 18 September, extending swarms, its feature for testing an MCP server through ChatGPT, Claude and Copilot to see whether users actually succeed. This version exposes typed swarm findings through the SDK, API, MCP and the Inspector UI, and adds a preflight check before each run. Inspector is MCPJam's open-source testing and evaluation platform for developers building MCP servers.