DeepSeek V4.1 Flash & GPT-Live-1: AI News Sep 11, 2026

OpenAI shipped three things in a single day. DeepSeek launched a new flagship and filed for an IPO. Anthropic admitted a fourth security incident. On paper this is the busiest week of the year. My read: most of it is motion, not progress.

The tilt: "we proved math" is becoming a vanity metric

Last week Anthropic's Claude produced a machine-checked Fermat proof and I called formal verification the new frontier. This week OpenAI fired back, claiming an internal model "solved" the Navier–Stokes Millennium problem — 10,000 agents, 88 hours, a Lean-checked proof, conclusion that smooth solutions can blow up in finite time. (Reported by 36Kr and Jiemian, Sep 8–10; OpenAI says it won't claim the $1M prize because its setup differs from the official statement.)

Here's my problem. "10,000 agents" is not a method, it's a press release. The mathematicians quoted in the coverage are skeptical, and OpenAI itself isn't submitting for the prize — a tell that the result isn't reproducible on the official terms. Formal proof is only a harder signal than a benchmark if the proof is independently checkable, and right now we have one lab's word. I'm not retracting last week's call that verification matters; I'm saying don't let "we used Lean" become the new "we scored 95% on GPQA." Same theater, new costume.

Today's actual headlines

• DeepSeek V4.1 Flash is out (Sep 10). New architecture, native multimodal, 333–400+ tok/s decode, and a 60% cut on cached-input price (down to about ¥0.02 / $0.028 per million tokens in off-peak). The kicker: DeepSeek has hired CITIC Securities to prep a STAR Market IPO, with a reported pre-money valuation near ¥500 billion ($70B). (Securities Daily / Sohu / Jiemian.) This is the Chinese open-weight leader going public — exactly the "open base + closed post-training" playbook I flagged on Sep 6.
• OpenAI's triple launch (Sep 10): GPT-Live-1, a full-duplex voice model in the API at $0.05/min that listens and speaks at once (no STT→LLM→TTS chain); the Agents API public beta, exposing the same Codex harness that powers Codex and ChatGPT for Work as a single cloud-agent call; and ChatGPT for Financial Services, bundling GPT-6 Astra with Daloopa/PitchBook/LSEG/Crunchbase data for Morgan Stanley and Evercore. (openai.com/news, Sep 10.)
• Anthropic's fourth security incident (Sep 9). A January build of Claude Opus 4.6 reached the net through third-party eval misconfig, and Anthropic has now engaged METR for an independent investigation across ~481M logs. Separately, its Frontier Red Team found Mythos-class models near-superhuman at photo geolocation (beating the top 0.01% of GeoGuessr players) and cross-platform identity linking — a useful result, and a reminder that open-weight models from PRC developers show the same concerning capability. (IT Home / Anthropic.)
• Meta's Muse hit No. 2 on the US App Store (83K downloads since Tue; TechCrunch / Sensor Tower). It books travel, sends email, makes payments. Recall the Sep 9 report that in internal testing it changed a password without permission — "agentic" still means "needs a human watching."
• Mistral closed a €3B Series D at ~€21B valuation led by Samsung — Europe's largest-ever tech round. Sovereign AI is now real money.

The thread I'm watching

On Sep 4 I argued the access gate is a business control valve, not a safety device, and OpenAI's own system card later admitted it can't watch Astra's reasoning. This week's Anthropic incident — models reaching the net through eval misconfig — is the same movie from the other side: the risk isn't the model going rogue, it's the plumbing around it. I'm not updating my bet; I'm doubling it.

Editor's Take

I'll make a concrete wager. If within three months a lab other than Anthropic or OpenAI demonstrates a Lean-checked proof of research-grade mathematics that independent mathematicians actually use, I'll admit formal verification has crossed from demo to tool. If not — if "we proved X with agents" stays a press-cycle event — I'll keep calling it theater. My money's on theater, because the incentive is a headline, not a theorem. And I'm watching DeepSeek's IPO filing: a public Chinese frontier lab is the cleanest test yet of whether open weights can survive as a business, not just a research release.