Good morning. The money story is today’s story: OpenAI closed $11B at a $1.25T valuation — short of what it wanted — and Anthropic’s upcoming IPO is being whispered about at as low as $1.1T, down from a $2T target. Ramp’s spending data suggests why investors are getting cold feet, and Wired has a timely reminder that open Chinese models are now nearly as good at hacking as the restricted frontier ones.
OpenAI raises at $1.25T, well short of the $1.4T it wanted. The round, reported by TechCrunch and The Verge, is expected to close around $20B total. The Verge attributes the gap to falling tech stocks and voter backlash after the midterms, with Sam Altman saying OpenAI’s own IPO is on hold until it can make stronger safety claims. Given yesterday’s Astra sandbagging disclosure, that bar may have just moved.
Anthropic’s IPO sees bids as low as $1.1T. Bloomberg reports buyer interest has softened considerably from the $2T figure floated earlier. HN commenters focused on the S-1 lockup schedule, which unlocks just as a new Congress is likely to be writing AI legislation, and on anecdotal evidence that local models are eating into subscription usage — one user reported dropping to 18% of their Claude Pro weekly quota after getting Qwen3.8-Flash-Next running on a Strix Halo box.
Ramp data shows AI spending flatlining. TechCrunch covers Ramp’s finding that corporate AI spending grew just 1.5% in August and September, down from 42% monthly growth in July. Token volumes are up, but price cuts from OpenAI and Anthropic are keeping dollars flat. The more worrying detail: growth has been concentrated in the top 1% of AI-heavy firms, whose per-employee spend is plateauing around $10K — a concentration risk right as Anthropic heads to market.
Free Chinese models catch up on offensive security. Vals tested open-weight models against OpenAI’s restricted Daybreak on mock corporate networks: DeepSeek completed 73% of hacking tasks versus Daybreak’s 81%, at roughly a tenth of the cost ($2–3 vs $17–46 per attack). This more or less kills the “restrict frontier access to slow capability spread” argument Anthropic was making seven months ago. Researchers note defenders will benefit from the same economics, for whatever that’s worth.
Claude Code email leak, continued. Following yesterday’s note on Claude Code stuffing user emails into User-Agent headers, Anthropic has responded in the GitHub thread: the account email is being removed from model context, with a fix expected early next week. A cc-plugin-privacy mod is already shipping in 2.1.317+ as an interim block. Part of the behavior traces to the SEC’s API, which requires an email in User-Agent, but subagents were also stuffing identifiers into arbitrary curl requests far beyond that case.
Google ships Gemini Omni 1.2 Flash. The new video model does 15-second native clips (up from 10), supports conversational multi-turn editing, and drops 720p pricing 30% to $0.07/second. HN reaction was almost entirely about the naming and rollout: one commenter noted their friend on Google AI Plus is still stuck on Gemini 3.6 Flash with no path to anything newer, and the version numbering restarting at 1.2 while the main line is at 4 drew the usual exasperation. The model is live in Google Vids, Flow, and the Gemini API.
METR says AI progress roughly doubled, citing redacted evidence. A METR analysis referenced in Anthropic’s S-1 claims capability progress from October 2025–2026 roughly doubled year over year, meeting Anthropic’s own “dramatic acceleration” threshold in its Responsible Scaling Policy. HN was unimpressed: the key supporting evidence is redacted, and commenters pointed out that benchmark aggregates conflate genuine capability gains with compute scaling, capital deployment, and training-set contamination. One ML researcher put it plainly: LLMs speed up the parts of ML research that happen inside a compute environment, but the algorithmic ideas aren’t obviously getting faster.
An “escalate to Opus” pattern claims 4x cost reduction. Hightouch writes up a hybrid where a small decision model (Jev) handles identity-resolution cases itself and escalates uncertain ones to Claude Opus, matching Opus accuracy at a quarter the cost. The HN thread liked the pattern but noted the result is specific to a well-defined task with calibratable confidence. One commenter made the uncomfortable observation that the application here — mapping fragmented records back to real individuals — is essentially cheaper, faster privacy violation.
China reportedly tried to recruit a spy close to Jensen Huang. The Verge reports that China’s MSS recruited Megaspeed exec Alice Huang — no relation, but close to Jensen — to gather intelligence on Nvidia’s chip roadmap and lobby for access to its top-end parts. Three House committees want Jensen for a closed-door briefing. Trump was reportedly briefed last year and did nothing.
That’s the week. Next week brings Anthropic’s roadshow in earnest, and we’ll see whether the $1.1T whisper holds or moves.