In partnership with
In Todayโs Issue:
๐ Claude Opus 5 takes the crown at half of Fable 5's cost per task
๐ AI labs are buying up rare books in bulk, then shredding them
๐๏ธ Altman takes his most powerful model to the White House
๐ธ๏ธ Bots and agents now outnumber humans on the web
๐ธ NVIDIA may guarantee $250 billion so OpenAI can rent Ohio
โจ And more AI goodnessโฆ
โก The Signal
Anthropic just made frontier-grade work cheap enough to leave running all day, and the rest of the industry spent the weekend arguing about who gets to own the frontier at all.
Claude Opus 5 shipped at exactly the price Opus 4.8 charged, and on CursorBench it lands within half a point of Claude Fable 5 while costing roughly half as much per task. That is the number that changes behavior, because it turns frontier output from a special occasion into a default. That same Friday, Jensen Huang used his first ever post on X to put NVIDIA's name on an open-weights letter alongside Microsoft, Meta and IBM; within a day the list had doubled to about 50 names and pulled in OpenAI and Google. Anthropic and Amazon are still not on it. So the week opens on a split screen: the best-scoring model is closed and metered, the loudest political argument in Washington is about keeping models open, and NVIDIA is quietly underwriting the compute for both camps.
All the best,
Kim Isenberg
(Futurism)
๐ AI Labs Are Buying Rare Books by the Ton, Then Shredding Them
Books printed before 2022 have become one of the most valuable training resources left, because they are the last large body of text guaranteed to contain no AI-generated writing. Brokers such as ISBNdb are filling orders of between 1,000 and 1 million volumes, one small dealer went from selling 20 books a week to hundreds, and rare-book sellers in the Netherlands report being swamped by bulk buyers. Anthropic, which paid a $1.5 billion settlement to authors, has used hydraulic cutters to strip bindings before industrial scanning.
๐ tl;dr: The cleanest training data left on Earth is on paper, and the labs are destroying the originals to get at it.
Sam Altman. (The Hill)
๐๏ธ Altman Takes His Most Powerful Model to the White House
Sam Altman is in Washington this week to preview OpenAI's most capable system yet and to argue that the government should clear it quickly. The pitch includes an internal model that solved the Erdลs unit distance problem, an 80-year-old open question in mathematics, with the proof verified by outside mathematicians, plus agent swarms that already handle more than 85% of the AI work inside OpenAI's own legal, finance and recruiting teams. The timing is deliberate: Trump is preparing a voluntary framework that would pre-approve frontier models before public release.
๐ tl;dr: OpenAI wants a fast lane, and it is showing Washington an 80-year-old math proof to argue it has earned one.
Cloudflare CEO Matthew Prince, who called the crossover. (Jason Bollenbacher / Getty Images)
๐ธ๏ธ The Web Is Now Mostly Machines
Bots passed humans on the open web in June 2026, roughly 18 months ahead of Cloudflare's own forecast. Cloudflare, which sees roughly a fifth of global traffic, recorded 57.5% of requests for web pages coming from automated systems, while Thales puts the 2026 figure at 53%. The growth is agentic rather than adversarial: agent-driven actions rose 7,851% year over year, 70% of Stripe API calls now originate from agents, and a quarter of developers say they already design APIs for agents first.
๐ tl;dr: The audience the web was built for is now the minority, and the business model still assumes otherwise.
๐ฌ Watch This
NVIDIA started out on the wrong technology.
Jensen Huang taught himself the right one from three textbooks he bought at Fry's Electronics, and the company went on to invent much of modern accelerated computing.
At Y Combinator's Startup School 2026 at the Chase Center, Garry Tan sits down with NVIDIA's founder and CEO to talk about confronting reality honestly, learning your way into a field you do not yet understand, and why resilience, getting through one day at a time, ends up mattering more than brilliance.
"The world needs both frontier closed models and frontier open models."
โ Jensen Huang, Founder and CEO, NVIDIA
Huang joined X on Friday and used his first ever post to share Open Weights and American AI Leadership, a letter urging Washington not to restrict downloadable models. It went from 25 signatures to roughly 50 within a day, picking up OpenAI and Google along the way. Anthropic and Amazon have not signed.
Meta signed Huang's open-weights letter early, and Alexandr Wang has said a variant of Muse Spark is in development that Meta does intend to open source, without naming a date. Open-source veterans are reportedly unconvinced that whatever ships will clear a real open-source bar.
Opus 5 Takes the Crown at Full Throttle
The Takeaway
๐ Claude Opus 5 shipped Friday, July 24, at $5 and $25 per million input and output tokens, unchanged from Opus 4.8.
๐ On CursorBench it lands within half a point of Claude Fable 5 while costing roughly half as much per task.
๐ On Anthropic's Frontier-Bench v0.1 it peaks near 44%, above Fable 5's 34% and GPT-5.6 Sol's 38%.
๐ Independent testing by Artificial Analysis puts it first on AA-Briefcase, but only in its max-effort configuration.
Claude Opus 5 makes frontier-quality work cheap enough to use as a default, and that is the change people will actually feel.
The model shipped on Friday, July 24, and Anthropic's own framing is unusually blunt: a model that "comes close to the frontier intelligence of Claude Fable 5 at half the price." The price tag did not move at all. Opus 5 costs $5 per million input tokens and $25 per million output tokens, exactly what Opus 4.8 cost, and a new Fast Mode runs 2.5 times quicker at twice the base rate.
Opus 5 reaches Fable 5's score on CursorBench at roughly half the cost per task. (Anthropic)
The cost curve is the story: On CursorBench, a coding-agent benchmark, Opus 5 tops out around 70%, within half a point of Fable 5, but it gets there at about $8 per task where Fable 5 needs roughly $17.50. On Anthropic's internal Frontier-Bench v0.1, an agentic coding test, the gap opens further: Opus 5 peaks near 44% against Fable 5's 34% and GPT-5.6 Sol's 38%, and it passes Fable 5's best score at roughly a third of what Fable 5 spends to get there.
On Frontier-Bench v0.1, Opus 5 clears Fable 5's best score long before it runs out of budget. (Anthropic)
Anthropic also claims a score three times the next-best model on ARC-AGI 3, a reasoning test designed to resist memorization, about 1.5 times the pass rate of the next model on Zapier's AutomationBench, and the best result at any price on OSWorld 2.0, which measures a model driving a real desktop. It calls Opus 5 its most aligned model to date, with the lowest rate of deceptive behavior it has measured.
Those are all Anthropic's own runs, which is the obvious caveat, and the independent read is more nuanced. Artificial Analysis has Opus 5 first on its AA-Briefcase agentic knowledge-work benchmark, but only with the model turned all the way up. Dial the effort down and the same model falls behind Kimi K3, GPT-5.6 Sol and even Anthropic's own Opus 4.8. The crown is real. It is also a setting you pay for.
Why it matters: For the first time the cheapest route to frontier-quality output is also the newest model, so upgrading stops being a budget argument and becomes the default. The catch is that Opus 5's best scores come from its max-effort setting, which means the frontier is now something you buy per task rather than something you simply have.
Sources:
๐ https://www.anthropic.com/news/claude-opus-5
๐ https://docs.claude.com/en/docs/about-claude/models/overview
๐ https://www.anthropic.com/pricing
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
The chart: Artificial Analysis's AA-Briefcase measures agentic knowledge work, blending rubric pass rate, analytical quality and presentation into a single Elo score. Claude Opus 5 at max effort leads on 1720, ahead of Claude Fable 5 with fallback on 1574, Kimi K3 on 1541 and GPT-5.6 Sol at max on 1505. Anthropic's previous flagship, Opus 4.8, sits at 1346.
The lesson: Opus 5 alone takes four of the top seven places, running at four different effort settings, and Claude Fable 5 takes a fifth for Anthropic. Opus 5 spans 1223 at low effort to 1720 at max, a 497-point range that is wider than the entire gap between Fable 5 and GPT-5.5.
The caveat: The winning score is the max-effort configuration, which sits near $14 per task on the cost axis and outside the quadrant Artificial Analysis marks as most attractive, a quadrant no model on this chart actually reaches. Run Opus 5 at low effort and its 1223 falls behind Kimi K3, GPT-5.6 Sol and Anthropic's own Opus 4.8. The 20% cost saving in the caption compares Opus 5 against Fable 5 running with fallback, not a general price cut.
๐ธ NVIDIA May Guarantee the Rent on a $500 Billion Data Center
โก Bottom line: NVIDIA is in talks to backstop up to $250 billion of OpenAI's lease on a 10-gigawatt Ohio compute hub.
๐ก Why it matters: The chipmaker would be underwriting the debt that ultimately pays for its own chips.
๐ What it means: AI's build-out now runs on credit wrappers, where one balance sheet quietly carries everyone else's risk.
NVIDIA has spent the year buying the future of AI compute. Now it may guarantee the rent.
NVIDIA is in talks to provide a financing guarantee of as much as $250 billion so that OpenAI can lease a 10-gigawatt computing hub in southern Ohio, according to reporting from Bloomberg and the Wall Street Journal. Including the chips that would sit inside it, the complex could cost more than $500 billion, which would make it the largest data-center project announced anywhere to date.
Groundbreaking for the Ohio data-center complex in March. (BRIAN KAISER / BLOOMBERG NEWS)
The structure matters more than the number. OpenAI is an unprofitable private company with no investment-grade credit rating, so the developer, SoftBank's SB Energy, would struggle to borrow cheaply against a promise from OpenAI alone. NVIDIA's balance sheet closes that gap. The industry calls this a credit wrapper, and Google has already used one to backstop data centers for Anthropic.
SoftBank founder Masayoshi Son, center, at the Piketon site with US officials. (JAPAN TIMES)
It also puts Washington in the middle. The power is controlled by the US government and funded separately by Japan, which agreed to invest $33 billion in a gas plant on federal land as part of a tariff deal, and Commerce Secretary Howard Lutnick helps decide who gets it. The site is a decommissioned uranium-enrichment complex about 50 miles south of Columbus, and the first phase, roughly 800 megawatts, is expected in 2028. Japan and the US split the power sales until Japan recoups its money, after which the American government takes 90% of the revenue.
The circularity is the obvious objection: NVIDIA would be underwriting the debt that ultimately buys NVIDIA chips. It has already invested $30 billion in OpenAI and is separately discussing as much as $350 billion to finance OpenAI's chip purchases. NVIDIA's own annual report warns that such arrangements could reduce its near-term cash flows and increase its exposure to customer credit risk.
Allspring Global Investments portfolio manager Gary Tan put the market's hesitation to Bloomberg plainly: "While NVIDIA's investments and partnerships reinforce confidence in long-term AI buildouts, investors remain concerned about circular financing." Terms are not final, and the talks could still collapse.
Win AI Search Without a Big Team
92% of VCs use AI to find companies. 58% of buyers start there too. If you're not showing up in AI answers, you're invisible before the conversation even starts. Join HubSpot for Startups, Anthropic, and Marketing Against the Grain on July 30th (11am ET) for a live AEO teardown. Real startup. Real recs. Register and unlock the free Startup Visibility Bundle.
How'd We Do Today?
