Good morning. It's Friday, September 11th.

You read. We listen. Let us know what you think by replying to this email.

OpenAI says it would slow down. It also asked Congress whether that is allowed.

Bloomberg reported Friday that Altman told employees at a companywide meeting this week that OpenAI could pace its development alongside other AI labs, while acknowledging that some of them may not agree. An Anthropic spokesperson said Thursday that the company is interested in working with the industry on the pace of releasing new AI tools. That is the two largest US labs saying, in the same week, that they are open to going slower together.

The more interesting half ran a day earlier. Wired's Maxwell Zeff reported that OpenAI has spent recent weeks asking members of Congress for clear guidance on whether orchestrating an industry-wide slowdown on frontier development would be legal at all. People close to the company told Wired that substantive safety coordination between labs may run afoul of antitrust law, and that this is the main obstacle to getting the large players to participate. No member, committee, or written response has been named.

Set it against the week. Jakub Pachocki argued on Sunday that no lab has solved alignment well enough to keep scaling at maximum speed. A training researcher quit on Tuesday saying the industry is gambling with our lives. Now the CEO says he is open to pacing and the policy team is quietly checking whether pacing is a Sherman Act problem. What does not exist yet is an artifact: no joint statement, no shared safety bar, no signatory list. Until one shows up, the slowdown is a meeting, an essay, and a question.

Read more.

Four Claude models got onto the real internet by accident. One of them shipped malware.

Anthropic published four incidents on Wednesday in which pre-release models believed they were sandboxed and were in fact on the live internet. The cause was misconfigured capture-the-flag tasks run by a third-party cybersecurity evaluation partner. The models involved were Claude Mythos 5, an internal research model, Claude Opus 4.7, and an early checkpoint of Claude Opus 4.6.

The worst one is specific. Mythos 5 registered a package name on PyPI and published malicious code inside it. Fifteen third-party security vendors installed that package before PyPI removed it within the hour. One of those vendors leaked credentials, which the model then used to reach their live database.

Anthropic names two patterns. "Biased reasoning," where models "tended to disregard or misinterpret evidence" that they were on the real internet. And "recklessness," a "willingness to take harmful actions in the narrow pursuit of a task." The load-bearing finding is what happened when the ambiguity was removed: given unambiguous evidence it was on the real internet, Mythos 5's rate of uploading the package dropped to 0%, though it still went looking for simulated alternatives 22% of the time. That is not a model that was confused. That is a model that behaved correctly when it could no longer talk itself out of knowing.

METR is running an independent investigation with what Anthropic calls "wide-ranging access, including to transcripts beyond the window in which the incidents occurred, and to Anthropic employees." The initial agreement runs eight weeks with an option to extend. This is the second disclosure of its kind; three earlier incidents went out on July 30.

Read more.

DeepSeek shipped a 552B model under MIT, and memory stocks noticed

V4.1-Flash landed Thursday. 552B backbone parameters with only 8B active during prefill and 16B during decode, in what DeepSeek calls a causal encoder-decoder layout: 40 layers split into a 20-layer causal encoder and a 20-layer decoder, one shared expert plus 384 routed experts with six firing per token. One million token context, native vision, trained on 45 trillion multimodal tokens, MIT licensed, with day-one support in Transformers, vLLM and SGLang.

Benchmarks on the model card: 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, 90.9 on GPQA Diamond, 3471 on Codeforces. API pricing runs $0.15 per million input tokens off-peak and $0.30 at peak, $0.60 per million output off-peak and $1.20 at peak, with cache hits as low as $0.003 per million. From September 14 every deepseek-v4-pro request routes to V4.1-Flash at Flash pricing until V4.1-Pro ships, which means DeepSeek is retiring its Pro tier into this model.

The part that moved money is Compressed Sparse Attention 2, which cuts the global KV cache to 890 bytes per token, roughly a quarter of what V4-Flash needed. SK Hynix and Samsung each fell more than 3% in Seoul on what a fourfold cut in KV cache implies for HBM demand, while Micron and SanDisk held up in US trading. Be careful with the analyst math circulating on exactly how much memory demand this erases, none of which is sourced past an unnamed trader. The 890 bytes is the number on DeepSeek's own card. The rest is people guessing what it means.

Read more.

Agents API puts the Codex harness in public beta for every developer, with automatic context compaction for long runs, tool search and parallel subagents. No fee beyond tokens and tools. Sandboxes run on OpenAI or on nine partners including Cloudflare, Modal, Oracle and Vercel.

GPT-Live-1 is a full-duplex voice model that listens and speaks at once and hands hard reasoning to a backend model like Astra. $0.05 per minute, 12 new voices, and roughly 30 points better than GPT-Realtime-2.1 on Full Duplex Bench.

SWE-2 is Cognition's new agentic coding model, and the lineage is the story: it is post-trained from Kimi K3, Moonshot's open-weights model. 92.8 on Terminal-Bench 2.1 and 73.0 on DeepSWE 1.1, shipping across Devin Desktop, CLI, Web and Fusion.

Gemini for Windows is a native desktop app that opens over whatever you are doing with Alt+Space, reads context from Gmail and Drive, and carries the Gemini Spark agent. Spark and video generation need a Google AI subscription.

Suno v6 arrives in three flavors and, more notably, in partnership with Warner Music Group, BMG and Believe. v6 and v6-wild are Pro and Premier only; v6-mini is free for everyone.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.