Price:

Claude’s new model can build bigger, but not judge better

Nov 13, 2026Summary from 3 podcasts.
  • Anthropic’s Fable 5.2 takes on longer, more complex work, from coding to scientific projects.
  • Its ability to finish a task still doesn’t mean it can judge whether the result is good.
  • The biggest gains may come from trying ambitious projects, not automating routine work.

Anthropic has released Claude Fable 5.2 for broad use and Mythos 5.2 for vetted partners, presenting a new pair of models built for demanding work. Fable led several coding and research benchmarks, while Anthropic said Mythos set a new bar in cybersecurity. The release comes with a practical question: how much can people delegate to a model that can work for hours, but still needs a human to set direction and assess what it makes?

On a test of contractor tasks, Fable completed 41.3% of assignments, nearly twice Opus 5.5’s 21%, according to figures discussed by Nathaniel Whittemore on The AI Daily Brief. Garrison Lovely called the jump alarming, while noting that these tasks are more clearly defined than typical office work. The result points to progress on discrete work, not proof that AI can take over a whole job.

Anthropic’s examples stretch beyond routine coding. The company reported that about 1,000 agents helped formalize a proof of Fermat’s last theorem in 30 hours; that is a formalization of a known proof, not the discovery of a new theorem. It also reported work in physics and protein modeling.

In a robotics demonstration, Fable controlled two arms on tasks it had not explicitly been trained to perform. Alex Finn, on This Week in AI, was more impressed by robotics, product design and code-generated media than by the physics claims.

The models are also being used to build large projects. Finn said Fable worked on a complex video game for two weeks without his input. The Every team said it could assemble a production-grade video editor from editing, effects and music tools, with less supervision than Opus 5.5 required. Those examples suggest a shift from asking AI to speed up familiar tasks toward attempting projects people might not otherwise build.

But output is not the same as judgment. Ethan Mollick called Fable “almost AGI” while describing work that was accurate but dull, repetitive or overcomplicated. Whittemore likewise found its sense of what would interest people unreliable. In one test, obvious mistakes and overlapping slide text escaped the model’s own review. People still have to decide what is worth making, and whether the result is actually good.

Cost may also shape where the model makes sense. Anthropic cut Fable’s token prices by 20%, but Whittemore said running it at maximum effort cost $26.54 per task in one benchmark suite, 2.6 times as much as Opus 5.5. Anthropic recommends medium effort by default, with higher settings for more demanding work.

Alex Finn called Fable the best model he had used; Sarah Hooker saw the releases mainly as a response to cost pressure. Either way, Fable’s promise is clearest on work ambitious enough to justify the expense, and important enough for a person to check.

Source Intelligence

- Deep dive into what was said in the episodes

Alex Finn: Anthropic has the best model, Grok has the best harness | TWiAI 39 • Nov 12

  • Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, while OpenAI released GPT-6 Cyber; the labs restricted their highest-risk cyber capabilities to vetted organizations, and Fable remained broadly available with cyber and biology safeguards.
  • Alex Finn called Fable 5.2 the best model he had used, citing faster, cheaper, stronger recent Anthropic releases and a complex video game built over two weeks without his input. Sarah Hooker argued the changes mainly reflect cost pressure, while Manu Sharma credited broad scaling across reinforcement learning, multimodal capability, and model behavior.
  • Manu Sharma said Anthropic generates more than 80% gross margins and that Claude writes 90% of its own training code, giving the lab an internal advantage. He said improved compute efficiency supports those economics, while lower prices could drive usage growth above the price reduction.
Also discussed on this episode: (14)

Safety (2)

  • GPT-6 Cyber was rated critical, OpenAI's highest preparedness risk tier, and reportedly stayed within its stated scope 99.93% of the time, up from 94.49% for GPT-5.6. The host framed the release as a response to rising cyber threats and companies' demand for stronger defenses.
  • Fireworks AI said attackers accessed customer datasets and unfinished models after using internal credentials, then attacked computers holding research data. The company blamed model-access theft and resale; CEO Lynn Xiao said the attackers' behavior suggested AI agents were involved.

Models (1)

  • Alex Finn said Grokbot's harness delivers more value for everyday knowledge work than Claude for his team, while Sarah Hooker said Anthropic deliberately keeps its model and harness tightly coupled. Manu Sharma distinguished routine work suited to smaller models from long-horizon science and engineering tasks that still benefit from frontier models.

Enterprise (1)

  • Manu Sharma said enterprise integrations, IT controls, memory, and accumulated context create substantial application-level lock-in, even as small teams switch tools freely. Sarah Hooker warned that open-weight models and automated self-improvement are making API dependencies less sticky.

Markets (1)

  • Anthropic's updated IPO filing set a $70 to $78 share range, implying up to $1.405 trillion fully diluted value; strategic investors agreed to buy $26 billion of stock. Alex Finn questioned how a model company could grow tenfold without a durable moat, while Sarah Hooker argued Anthropic must move into vertical products such as legal, tax, healthcare, and finance.

China (2)

  • Sarah Hooker said export controls have been difficult to enforce because Chinese firms can use data centers in nearby countries, while China is also developing domestic chips and optimizing models for efficiency. Manu Sharma added that Chinese labs can use remote compute through APIs, so chips need not sit on the mainland.
  • The host reported allegations that Nvidia executive Alice Wong helped divert chips to Chinese companies, including an alleged $7.5 billion shipment, and said Nvidia's board found no reason to believe secrets or national security were compromised. The board opened an independent investigation, and Reuters had not independently verified the reported figures.

Chips (1)

  • Alex Finn argued Nvidia should not remove Jensen Huang absent evidence he deliberately harmed US security, calling Huang central to the company. Sarah Hooker said the export-control story highlights compute as a strategic resource, while Manu Sharma questioned whether restrictions can stop Chinese AI development.

Regulation (1)

  • The host described HR 3447, the Chip Security Act, as requiring location-verification mechanisms for covered chips and reporting when hardware is diverted. Alex Finn argued export controls will be circumvented when demand is strong, while the host predicted the bill would advance after the Nvidia allegations.

AI Infrastructure (2)

  • Sarah Hooker said US opposition to data centers reflects broader anxiety about AI and can obstruct essential infrastructure through local processes. Manu Sharma said developers and governments need clear, locally relevant evidence about jobs and tax revenue; both rejected unsupported claims about water and pollution as poor grounds for debate.
  • A proposed $5 billion data-center campus at the former Landover Mall site stalled after local protests and petitions; the mall had been demolished, leaving vacant land. The host alleged foreign-backed groups may be fueling opposition, citing Sam Lyman's claims, while Sarah Hooker said public receptivity to AI matters more than tracing money flows.

Adoption (1)

  • Alex Finn said his parents became enthusiastic about AI after ChatGPT helped them repair a broken door, and argued that practical AI education would build lasting support for data centers better than giveaways. The discussion also cited a Geely model priced at $9,000 that had displaced Tesla's Model Y as China's top-selling car.

Startups (2)

  • Sarah Hooker said Adaption Labs automates fine-tuning and continuously updates models using a company's context, aiming to make smaller models useful without the manual tuning process she said many firms found difficult. Manu Sharma described Labelbox's Recursion as an enterprise agent platform for recurring technical work such as underwriting, research, and model training.
  • Alex Finn described Henry Intelligent Machines as an agent that learns a user's skills and helps build businesses around problems that user can solve. Finn said the product was preparing a waitlist and envisioned agents creating value while their users pursue other work and interests.

Can Jensen Huang remain CEO? Inside NVIDIA’s next steps| MOONSHOTS #310 • Nov 11

  • Anthropic’s Fable 5.2 and Mythos 5.2 were described as 30% faster than their 5.1 versions, with Fable available broadly and Mythos limited to vetted partners. The models reportedly solved two open black-hole physics problems, designed working circuit-board devices, modeled 28 of 34 gene-therapy capsids, and helped about 1,000 agents formalize Fermat’s Last Theorem in 30 hours.
  • Alec Hosare-Gross said Fable’s strongest demonstrations were robotics, product design, and code-generated media, rather than physics benchmarks that he found less impressive. The panel cited an ECI score of 174 and a GDP X result of 16.4%, while Alec predicted GDP X would saturate within six months and said remote work could be automatable within one or two years.
Also discussed on this episode: (23)

Chips (3)

  • Alec Hosare-Gross said reports allege that NVIDIA GPUs worth $8 billion were sold to China through Alice Wong and Mega Speed via a Singaporean syndicate, with activity traced to 2023. He argued the investigation could expose weaknesses in Pax Silica, the effort to keep Western chip and AI infrastructure within the West’s sphere of influence.
  • The speakers said NVIDIA relies on TSMC rather than manufacturing chips itself, so changes in capacity allocation could shift where much of the sector’s value lands. They described the alleged diversion as a human vulnerability in chip-export controls and warned that a prolonged board investigation could produce weekly disclosures.
  • Elon Musk said a Terrafab could produce a terawatt of AI chips, about 30 times projected global annual output, though the panel stressed this was a target rather than existing capacity. Musk also cited more than $100 billion in December revenue run rate and said $500 billion the following year, or even $1 trillion, was possible; he said the project could cost $1 trillion.

Regulation (2)

  • The panel debated whether scrutiny of Chinese-born researchers should turn on ethnicity, birthplace, or nationality. One speaker argued export controls generally distinguish US persons from foreign nationals; the group expected stronger security clearances for frontier-model work and possible limits on Chinese models in banking, insurance, cars, drones, and government systems.
  • Delaware was drafting an artificial-intelligence-company entity that could let an AI agent manage a firm, sign contracts, hold property, and sue, while shielding human members except in cases such as fraud or underfunding. The panel debated whether AI corporate status should remain state law or become a federal framework; the proposal was still a draft.

China (1)

  • A speaker said MIT had restricted Chinese robots to a separate room after an internal investigation found they transmitted video and audio to China. The panel also cited reports of Chinese open-source models generating backdoors and argued that internet-connected models can leak information through covert channels.

AI Infrastructure (3)

  • Peter Diamandis estimated Anthropic produces $4.39 of economic output per kilowatt-hour, compared with $6.30 for New York City workers; his projection puts Claude at $6.20 in 2027 and roughly 40% above humans by 2030. The discussion disputed whether human education costs and productive years change the comparison.
  • The panel described SpaceX as shifting from rockets toward cloud computing, chip manufacturing, and large-scale infrastructure, with Anthropic as its largest customer. Salim Ismail argued vertical integration can bypass supply-chain bottlenecks; Alec Hosare-Gross characterized the SpaceX-Anthropic arrangement as a partnership that could position SpaceX as Anthropic’s infrastructure layer.
  • The panel debated whether future compute will concentrate in data centers or move to robot and vehicle edge chips. Alec Hosare-Gross expected data centers to dominate inference, while Imad Mostaque argued robot, car, and home brains could absorb much of global compute; they also discussed possible 20-megawatt limits on data centers.

Startups (2)

  • Peter Diamandis recounted Zipline’s announced $1.5 billion financing at a $20 billion valuation, an Uber agreement for one million deliveries annually, and a nationwide Walmart arrangement promising delivery in eight minutes. He also described Varda’s orbital protein-crystallization business and said three vehicles were already in orbit.
  • Anthropic launched an IPO roadshow at a proposed $1.3 trillion to $1.4 trillion valuation, below earlier figures of $1.5 trillion and $2 trillion; the offering was expected the week before Thanksgiving. Dave Blunden said the company could build its cash reserves through later offerings rather than raise $100 billion at once.

Agents (1)

  • Alec Hosare-Gross said Anthropic’s internal model four had an estimated 240-hour task horizon, while another estimate put model time horizons at doubling every two months. He argued internal models may be reserved for training stronger models, and Salim Ismail said longer horizons could let companies delegate outcomes rather than discrete tasks.

Health (3)

  • Dave Blunden’s agent framed curing 5,000 diseases as a platform problem, not 5,000 separate drug inventions: approve a gene editor, delivery method, manufacturing process, and safety protocol once, then adapt the sequence. The thought experiment estimated that lowering each disease-specific addition from $2 billion to $2 million would cut the total from $10 trillion to $10 billion.
  • The same thought experiment said delivery currently works for liver, blood, and eye tissue, but not reliably for kidneys or most of the brain. It cited roughly half of rare-disease patients lacking a molecular diagnosis, a five-year average wait for others, and about 1,200 certified medical geneticists serving 25 million to 30 million people.
  • Dr. Dawn Musalem said Fountain Life found previously undetected cancers in 3.3% of members who considered themselves healthy. The clinic uses full-body MRI and early-detection screening, which she said conventional preventive care does not typically provide and insurance does not yet cover.

History (1)

  • Peter Diamandis said an AI-assisted comparison ranked current civil unrest ninth among 250 years of US upheaval and AI and robotics seventh among technological disruptions. Salim Ismail argued the comparison misses that AI automates human adaptation itself and that institutions struggle to absorb multiple fast-moving technological shifts.

Robotics (3)

  • Elon Musk estimated that 20 million to 30 million humanoid robots could match the productive output of US workers, assuming each robot produces about five times as much as a person. The panel debated robot rentals, maintenance, and affordability; Musk put initial costs near a car’s price, with one speaker citing $30,000 falling toward $20,000.
  • The panel said AI-designed mechanical devices could be paired with automated manufacturing, producing many more specialized products rather than a few standardized designs. Speakers expected robot subscriptions and rentals to become common, but disagreed over whether general-purpose humanoids or simpler local systems would be economical for supermarket work.
  • Theseus’s Theo robot was presented as a supermarket restocking and online-order picker, with the company claiming a $60 billion annual market. Speakers disagreed on whether specialized shelf robots can beat general-purpose humanoids: some favored domain-specific systems, while others said Figure and Optimus would outcompete them.

Big Tech (1)

  • Elon Musk said SpaceX could use acquired spectrum more efficiently than legacy telecoms and might buy a carrier. The panel linked this strategy to Starlink’s launch capacity and need for demand, while another speaker said Starlink could weaken national phone companies and encourage global infrastructure monopolies.

Space (2)

  • NASA announced the next Commercial Lunar Payload Services contracts as part of a mission-by-mission moon-base plan, with eight companies competing under a 10-year program capped at $10 billion. Blue Origin’s Promise nuclear rover is scheduled for the lunar South Pole, where permanently shadowed craters may contain water ice.
  • The panel argued that space could offer a fresh setting for new technology and social structures, while smaller countries could attract AI or biotech activity through accessible governments and permissive regulation. Salim Ismail also described inner space and Eastern traditions as another frontier for understanding human experience.

Markets (1)

  • Peter Diamandis cited economist Christopher Clark’s cross-country analysis linking more billionaires per capita with less poverty, while speakers stressed that correlation does not prove causation. Salim Ismail said Sweden challenges the idea that entrepreneurship and a safety net conflict; Dave Blunden argued institutional quality matters more than partisan labels.

Fable 5.2 Is the Best AI Ever. So Why Am I Not Switching Yet? • Nov 10

  • Nathaniel Whittemore says Fable 5.2 leads coding and agentic scientific research, while Mythos 5.2 sets a new cybersecurity bar. The reported margins over the previous leaders were 4 points in coding, 6 in research, and about 10 percentage points in cybersecurity.
  • Fable 5.2 completed more agentic knowledge-work tasks than several rivals, but its AA Briefcase ELO remained below Opus 5.5. On the Remote Labor Index, it nearly doubled Opus 5.5, a result Garrison Lovely called alarming while noting the tasks are more defined than typical office work.
  • Fable 5.2 took the top three positions on the Artificial Analysis Intelligence Index across effort settings. Artificial Analysis credited gains in science, knowledge, and coding, with less progress in agentic knowledge work and terminal coding.
  • Anthropic cut Fable 5.2 token prices by 20%, but high-effort use can still cost more per task than Opus 5.5. Anthropic recommends medium by default, high for everyday coding, extra high for measurable long-running work, and max only for testing the model’s ceiling.
  • Anthropic says Claude writes 90% of its code changes under human supervision and uses Fable 5.2 for difficult bugs and projects other models cannot justify. Its showcase operating system involved more than 1,000 subagents working for eight days, while the model also handled a Python-to-Rust production-service migration.
  • Anthropic reported scientific work spanning theoretical physics, protein design, and formal proof. Fable 5.2 formalized Fermat’s last theorem with agents, while its capsid modeling could help biologists design gene-therapy delivery structures faster and more cheaply.
  • Anthropic’s robotics demo showed Fable 5.2 controlling two robotic arms on tasks it had not explicitly been trained to perform. The company said it outperformed most models built specifically to control robotic arms, extending the general-purpose model advantage into physical control.
  • Early users reported strong vulnerability discovery. Leander Builds said Fable 5.2 found severe issues created by Opus 5.5, while Pavel Hurin’s planted-bug test found 58 bugs, seven more than the previous best, with fewer turns and lower cost than Sonnet 5.5 Max.
  • Fable 5.2 produced more detailed Blender scenes than Opus 5.5 and generated a polished short video advertisement live. Nathaniel Whittemore says the demos show that code generation can create visual work with richer detail, lighting, and physics, though those displays do not establish broad everyday usefulness.
  • Ali K. Miller used Fable 5.2 for projects ranging from a government-data map to an ADHD-themed YouTube channel and a 3D task-management game. Miller trusted it more than other tested models, but noted obvious color and unit mistakes and overlapping slide text that its secondary agents missed.
  • Every’s team said Fable 5.2 could build a production-grade video editor by combining editing, effects, and music tools, with less supervision than Opus 5.5 required. The result revived the idea of professionals building custom software, though the time and maintenance burden may still make established SaaS products preferable.
  • Ethan Mollick called Fable 5.2 “almost AGI” because it handled complex work but failed to judge its own quality. He described accurate yet joyless course material, repetitive writing, and overbuilt rules; critics and fact-checkers caught errors, but did not recognize when the work was dull or poorly judged.
  • Nathaniel Whittemore found Fable 5.2’s sense of human interest unreliable: its standard-interface showcase favored difficult-to-verify intellectual tricks, while its coding-harness film was more compelling. Its first proposed research episode also measured launch-to-access delays that Whittemore judged irrelevant to current model users.
  • Fable 5.2’s clearest new use may be ambitious work rather than routine efficiency. Nathaniel Whittemore sees playable games as a promising test: users reported functional game mechanics, including a Harry Potter-themed battle royale, and Anthropic’s Sholto Douglas said the model cleared his bar for fun, beautiful games.
  • Advice from Anthropic and early testers favors assigning Fable 5.2 whole, open-ended projects, setting approval boundaries, and asking it to critique an idea before building. Anthropic also recommends telling the model it owns the codebase; Nathaniel Whittemore says the model’s strengths and best workflows may take weeks of experimentation to understand.

More in AI & TECH