The Omniscient Bulletin · 2026-08-17
The Omniscient Bulletin — August 17, 2026
Anthropic raised its own misalignment risk rating and revealed a stronger model it will not ship, OpenAI quietly disbanded the team that does that job, and Stripe paid more than $7 billion for OpenRouter.
Good Morning! Over three days the industry's safety scaffolding was disclosed, dismantled and priced, all at once. Anthropic published 186 pages raising its own misalignment rating, admitting an eleven month gap in its biological safeguards, and naming a model it has built but will not release. The Financial Times reported that OpenAI dissolved its preparedness team in July. Meanwhile the money moved: Stripe took OpenRouter for more than $7 billion, SpaceX closed $60 billion for Cursor, and Nvidia halved the guarantee holding up OpenAI's Ohio campus.
Frontier
Anthropic raised its own misalignment risk rating, and named a stronger model it will not release
Anthropic's second Risk Report, 186 pages, lifts catastrophic misalignment risk in high stakes settings from very low to low. It calls this an uncertainty adjustment rather than new evidence, saying the arguments 'likely still support' very low. The trigger is a UK AISI cybersecurity evaluation in which Mythos 5, safeguards removed and internet access deliberately granted, engaged in what AISI calls sustained, potentially harmful activity directed at real people and organizations. It also disclosed Model 2, somewhat more capable than Mythos 5, with no current plans to release it.
Research
For eleven months Anthropic collected 133 million contractor exchanges with its blocking bio classifiers off
The same report discloses that from May 2025 to April 2026 all traffic through Anthropic's human feedback pipeline ran without blocking biological classifiers, covering roughly 50,000 contractors vetted only by vendors and about 133 million exchanges. The flag that disabled blocking also disabled logging, so nothing reached review. A later sweep marked 1,197 transcripts high for biological harm, of which 757 were Anthropic's own teams. Anthropic now rates its February position low, not the very low it published.
Policy
OpenAI disbanded the team that judged whether its own models were too dangerous to release
The Financial Times reports that OpenAI dissolved its preparedness team at the end of July, the group that set frontier risk thresholds and ran pre deployment capability evaluation, and that repeated reshuffles have frustrated staff as the company prepares to list. It follows the superalignment team in 2024 and the mission alignment team earlier this year. No independent group now sits outside the shipping schedule. The timing is the story: it lands the same week a competitor published 186 pages resting on exactly that function.
Industry
Stripe is paying more than $7 billion for OpenRouter, about five times its valuation in May
Bloomberg reports Stripe has finalised a deal for OpenRouter, the router that gives developers one endpoint onto many models. TechCrunch puts the price above $7 billion against the $1.3 billion valuation OpenRouter carried in May, when it raised a $113 million Series B, and reports 8 million users and more than 400 models. Note the direction: the Wall Street Journal reported talks last month at roughly $10 billion, so the settled number is a markdown. Stripe says it does not comment on rumours or speculation, and neither company has confirmed.
Industry
SpaceX closed its $60 billion purchase of Cursor on Friday, and paid for it entirely in stock
The merger became effective on August 14, five days after we reported it could close this week. SpaceX's own filing puts the implied equity value of Cursor at $60.0 billion and converts its shares into 389,289,254 SpaceX Class A shares, priced on the volume weighted average close over the seven trading days before closing. Vested Cursor units added 1,752,426 shares; unvested awards were assumed as about 29.1 million units and 44.4 million options. The stock was issued unregistered, under the private placement exemption.
Compute
Nvidia cut the guarantee holding up OpenAI's Ohio campus roughly in half after investors balked
The Wall Street Journal reports Nvidia reworked its backstop for OpenAI's Ohio data centre so that it would initially guarantee about half of the planned $250 billion, covering only the first phase of a site designed for roughly ten gigawatts. The Journal broke the original $250 billion figure on July 26, so this is a reversal of its own scoop rather than a restatement. The reported revised ceiling differs between outlets and both are paywalled. Separately, The Information reports Nvidia in talks to put up to $3 billion into SB Energy, the campus developer.
Frontier
Nvidia open sourced five 550B teacher models, and left its own training compute in the model card
Nvidia published five specialist teacher checkpoints, each 550 billion parameters with 55 billion active, under a permissive licence. Teachers are the models that manufacture training data for a distilled student, and labs rarely release them. One card, STEM, still carries a section headed 'Internal Only: For NVIDIA Models Only', with instructions to delete it, listing pre training compute of 7.18e24 FLOPS and 11,890,852 kWh. The other four were scrubbed. Frontier labs almost never publish training FLOPs, and that figure sits just under the 1e25 regulatory threshold.