Last week the cheap seats came for the frontier's price. Nvidia bought the means of mass-producing free models and AT&T showed it could cut costs by more than half routing easy work to open weights. This week the labs answered, and the answer was not a price cut.
On Wednesday Nvidia reported $96.2 billion in a single quarter, and within a day had agreed to buy Hugging Face, the place those free models actually live. On Friday night OpenAI told Cursor it will stop serving it on November 12, because SpaceX owns it now and OpenAI does not trust Musk to honor a contract. On Monday Anthropic signed a $35 billion compute deal with a cloud provider most people have never heard of. On Tuesday it shipped Fable 5.1 and Mythos 5.1, held the sticker price flat, kept the stronger one behind a vetted-customer gate, and quietly locked new accounts out of the trick rivals use to copy its reasoning.
Four moves, one shape. The question this week stopped being what a model costs and became who is allowed to buy it. Supply is the lever now.
Anthropic released Claude Fable 5.1 and Mythos 5.1 on Tuesday, calling them the "world's most advanced models for coding and knowledge work." It's the first refresh of the Mythos-class line since Fable 5 landed on June 9, which is a reminder of how compressed this all is. Three months ago that model was new. In between it got switched off by the Commerce Department for eighteen days, then undercut on price by Anthropic's own Opus 5.
The capability story is real. Fable 5.1 matches or beats its predecessor at low and medium effort and pulls further ahead when you let it think longer, and Anthropic says it tops Fable 5, Opus 5, and OpenAI's GPT-5.6 Sol across multiple benchmarks. The demo everyone repeated came from the hedge fund Millennium, where the model traced a rare crash that had defeated the firm's own engineers and other models for years. The headline pricing did not move, still $10 per million input tokens and $50 per million output. What moved is cache reads, cut by 75% to $0.25 per million, which Anthropic says works out to roughly 25% cheaper for typical workloads and up to 45% for heavy agent runs.
Worth sitting with the skeptical version of that number. Artificial Analysis, which did pre-release testing, found that at maximum effort Fable 5.1 actually costs about 20% more per task than Fable 5, because it burns roughly 1.7 times as many output tokens getting there. On its index, a max-effort Fable 5.1 task ran $3.76 against $2.34 for Opus 5, which scored three points lower. So the cheaper-per-workload claim is true for the cached, agentic pattern most teams run, and false if you turn the dial all the way up. Both things are in the same launch.
Now the part that fits this week's shape. Fable 5.1 and Mythos 5.1 are the same underlying model with different safety filters. Fable is generally available. Mythos is restricted to American companies and individuals inside Anthropic's trusted access programs, with a new Life Sciences Verification Program for biology work. Fable's cybersecurity filter now fires roughly 60% less often per session in Claude Code, and the model is cleared to find software vulnerabilities but not to write exploits for them. That is the exact capability that got the June model recalled, loosened deliberately, with the release valve pointed at customers the government can identify.
Two quieter changes matter more than the benchmarks. New API accounts opened on or after August 31 can no longer edit a conversation's earlier turns while keeping Claude's stored reasoning intact, which is a known path into distilling a rival model. Older accounts are exempt for now, and Anthropic says every account will face the check on future releases. And every Fable 5.1 output carries a statistical watermark with no opt-out, with a detection API in private preview for regulators, newsrooms, and researchers. Both are defensive. One protects the training data from competitors, the other proves the text came from Claude.
The day before all this, Anthropic signed a $35 billion cloud deal with Lambda, first reported by the Wall Street Journal and confirmed by Reuters through a person familiar with it. The capacity is about 350 megawatts at a Nueces County, Texas campus that Hut 8 is building, a former bitcoin miner turned data-center landlord. Lambda is backed by Nvidia. Nvidia supplies the chips. And according to the Financial Times, Nvidia holds the 15-year lease under the whole thing. That comes a week after $45 billion committed to Nscale in West Virginia, and sits alongside reported deals near $50 billion with Fluidstack and $45 billion with Musk's SpaceX. Forbes put the recent total around $180 billion.
My take: I write this newsletter with Claude, so read me with that in mind. What strikes me is the gap between the two announcements. On Monday a company that showed investors one profitable quarter committed another $35 billion to compute it will pay for over years, underwritten by a chip vendor signing the lease on the building its own chips will sit in. On Tuesday that company shipped a model whose most interesting features are a lock on distillation, a watermark you cannot turn off, and a capability gate keyed to whether you are American and vetted. That is not the behavior of a business that thinks its edge is the benchmark. It's the behavior of one that thinks its edge is the supply, and is spending accordingly to hold it.
Late Friday, OpenAI said it will end Cursor's access to its models on November 12. The trigger was ownership. SpaceX closed its $60 billion all-stock purchase of Anysphere, Cursor's parent, on August 14, and OpenAI invoked a change-of-control clause with the longest notice its contract allows. The stated reason is not competitive. OpenAI says it cannot be confident SpaceX will stay inside its terms of service, citing what happened at X after Musk took over and Musk's own courtroom admission that xAI trained on OpenAI outputs. An OpenAI executive reduced it to four words, that "it boils down to trust."
Musk replied on X that he "couldn't care less," then called Altman and Brockman thieves, which is roughly the temperature that feud has run at since 2018. The practical damage is smaller than the headline. Cursor CEO Michael Truell said OpenAI models serve about 5% of user traffic and that the two sides are still talking. Developers can bring their own OpenAI key or route through Azure or Bedrock. Most Cursor work already runs on Claude or Gemini, and Cursor has been training its own Composer models on cheaper open checkpoints precisely so a moment like this stings less.
About two hours after the announcement, Anthropic co-founder Tom Brown posted that Cursor had been a trusted partner since Sonnet 3.5 and that his company would "continue to increase compute to support Claude models in Cursor." Higher usage limits followed. It was a clean piece of opportunism and it did not go unchallenged. Replit CEO Amjad Masad pointed out that Anthropic cut Windsurf off from Claude last June when Windsurf was heading toward OpenAI, and Anthropic also pulled OpenAI's own Claude access over benchmark use. Nobody in this fight has clean hands. The loyal-supplier posture lasts exactly as long as the loyalty is convenient.
What makes it matter is the precedent. This is the highest-profile case yet of a frontier lab using contract terms as a competitive instrument rather than a legal formality. Model access has been sold like a utility, metered and neutral, available to anyone with a credit card. It turns out to be closer to a distribution agreement, revocable when your cap table changes. Cursor is the most valuable application company ever built on somebody else's models, and its supply got cut because its owner changed, not because it did anything wrong.
My take: Every founder building on an API should read the Cursor cutoff as a product spec, not gossip. The lesson is not that OpenAI is petty or that Musk is difficult. It's that your model supply is a contract with a termination clause, and the events that trigger it may have nothing to do with you. An acquisition, an export directive, a rival's phone call to Washington. We have now watched all three inside one quarter. The teams that shrugged this week were the ones already running more than one model behind an abstraction, and Cursor was one of them, which is the only reason a 5% number was the whole story instead of an outage.
For three years the pitch on frontier models was that they were a utility. You paid per token, the meter ran, and the only real question was which one performed best for the money. Last week that story reached its logical end, with open weights good enough to take 40% of a phone company's traffic and Nvidia funding more of them. If intelligence is a commodity, price is the whole game.
This week the labs declined to play that game. OpenAI stopped selling to a customer over who owns it. Anthropic left its headline price untouched, gated its strongest model to vetted Americans, and shut the door on copying its reasoning. Nvidia moved to buy the hub where the free alternative is distributed. None of that is competing on price. All of it is competing on who gets served, on what terms, and with what strings. A commodity does not care who buys it. This does.
If you build on top of these systems, that reframes the risk. Your exposure was never really that a better model would arrive and make yours look slow. It's that the thing you depend on is a relationship, and relationships end for reasons that have nothing to do with your product. Your vendor buys your competitor. Your parent company gets acquired. A federal letter arrives on a Friday evening. The right response is unglamorous and I keep landing on the same one, which is to make the model a swappable part rather than a foundation. The teams that did that had a shrug of a week. Everyone else had to read a contract.
Buried in that pile of California bills are two that should be familiar to anyone who followed the Ratepayer Protection Act here. AB 2383 and SB 886 both cleared the legislature Monday night and both direct the state utilities commission to set a separate rate structure for large data centers, with new transmission and generation tariffs for big loads starting in 2027 plus new permitting and disclosure rules. That is close to what Senate Bill 730 tried to do in North Carolina before it died without a floor vote. California passed its version in the last hours of a session. Newsom has until September 30 to sign.
Our version of that fight moved this week too, just in a quieter room. The large-load tariff proceeding that came out of Duke's rate settlement is now live at the Utilities Commission, with competing proposals from Duke and from the Public Staff on the table. Both would cover customers drawing more than 50 megawatts, roughly the appetite of 40,000 households, and both include minimum billing, contract commitments in the ten-to-fifteen-year range, and termination terms making a developer pay for grid upgrades even if the project never materializes or never uses the power it reserved. The Carolinas rate decision is still expected in November, with new rates on January 1.
Meanwhile the local pauses keep stacking. Frank Muraca, an analyst tracking this across the state, now counts 46 North Carolina local governments that have adopted a data-center moratorium, up from the 28 the North State Journal tallied in early August. Durham County's nine-month version, passed 4 to 1 just before midnight on August 24, drew a line from Commissioner Nida Allam worth keeping, that "a delay for a company of nine months is monetary," while the effects on residents last a lifetime.
Indy Week reported this week on the wrinkle specific to us, which is that most Durham data centers sit outside the city limits, and most of those are in Research Triangle Park. The research economy that is supposed to be the Triangle's advantage in AI now runs through the same buildings residents are voting to pause. That tension is going to define the ordinance county staff drafts by December. My read is that the moratoriums are the visible fight and the tariff docket is the real one. A county can decide where a building goes. Only the Utilities Commission decides whose bill pays for the substation feeding it.
That's the week access became the product. See you next Wednesday.
Daniel
BullCity AI · Durham, NC
P.S. How many models could your stack actually swap to tomorrow, without a rewrite? Hit reply with the honest number. I asked for model lock-in stories back in June and I'm still collecting them, and this week gave everyone a fresh reason to check.
P.P.S. Forward this to whoever on your team owns the vendor contracts. They had a more interesting week than they realize.