ModelsArchitectures & capability
Ethan Mollick calls Anthropic's Claude Fable 5.2 "almost" AGI after a business-launch test
Wharton professor Ethan Mollick says Anthropic's Claude Fable 5.2 is close to AGI but still uneven. Its weak spots are now in carrying work out, not in errors. Artificial Analysis ranks it first for intelligence but finds it slow and costly per task.
From the One Useful Thing article of November 9, 2026 and cited sources: Mollick gave Fable 5.2 one prompt: launch a business that makes money. [1] [3]
SOURCES[1] One Useful Thing: Fable 5.2 and Jagged AGI oneusefulthing.org[3] One Useful Thing: On Jagged AGI: o3, Gemini 2.5, and… oneusefulthing.org
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- One Useful Thing article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 9, 2026 article published
- Covered
- Nov 12 edition
- Evidence
- 10 sources across 9 sites
- 1One Useful Thing: Fable 5.2 and Jagged AGIoneusefulthing.org
- 2Artificial Analysis: Claude Fable 5.2artificialanalysis.ai
- 3One Useful Thing: On Jagged AGI: o3, Gemini 2.5, and…oneusefulthing.org
- 4Simon Willison: Ethan Mollick on Jagged AGIsimonwillison.net
- 5Anthropic: Claude Fable 5.2 and Mythos 5.2 announcementanthropic.com
- 6The Decoder: Fable 5.2 and Mythos 5.2 cement Anthropic's…the-decoder.com
RisingHow it’s measured
Hacker News · 24 likes · 2 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Mollick gave Fable 5.2 one prompt: launch a business that makes money. The model screened 22 ideas, read public agendas from 201 city and county governments, and proposed a plausible service that tracks municipal procurement. It rated its own chances of success at 20–40%. Mollick stopped the test before the model could run the business. He worried it would take money without delivering quality, and the infrastructure it assumed did not exist yet. He calls the model "Jagged AGI": a term he first used in 2025 for models that are superhuman in some areas and unreliable in others. In his new view, the gaps have moved away from hallucinations and toward execution, such as infrastructure, sales and the many tasks involved in running a business. [1] [3]
Artificial Analysis gives Fable 5.2 a score of 65 on its Intelligence Index, which ranks it first of 74 models. The same testing found it slow and expensive. A task takes about 47.8 minutes, against a median of 12.7 minutes, and costs about $26.54. The model also produced 275M output tokens across the test suite, against a median of 72.8M. Anthropic's own coding and robotics benchmark scores have not been reproduced by anyone else. Only the overall ranking has been. [2] [5] [6]
Anthropic cut the price from $10/$50 to $8/$40 per million input/output tokens. Even so, Artificial Analysis puts the average cost per task at 2.6 times that of Opus 5.5. Developers also face breaking changes. Adaptive thinking is always on. Forced tool use returns an error, and temperature and top-p/top-k settings should be left out. The model also requires 30-day data retention unless Anthropic authorizes otherwise. [5] [6] [8]
one expert's hands-on tests, plus leaderboard results showing top scores with long, costly runs. Implication: the model may suit long research and planning work. Running operations still needs people, and teams should budget per task, not per token.
Executive brief
Ethan Mollick, a Wharton professor who had early access to Anthropic's Claude Fable 5.2, gave it one prompt: "launch a business to make me money." The model screened 22 ideas, read public agendas from 201 city and county governments, and produced a plausible municipal procurement-intelligence service. It also put its own odds of success at 20–40% (One Useful Thing). Mollick calls the model "almost" AGI, or "Jagged AGI." He argues the remaining gaps are no longer simple errors or hallucinations. Separately, Artificial Analysis ranks Fable 5.2 first on its Intelligence Index, and also finds it slow, verbose and expensive per task (Artificial Analysis).
What changed and event timeline
Mollick coins "Jagged AGI."
Writing about OpenAI's o3 and Gemini 2.5, Mollick described models that are superhuman in many areas yet unreliable at some mundane tasks (;).
Independent leaderboard result
Artificial Analysis scores Fable 5.2 at 65 on Intelligence Index v5, ranked #1 of 74 models. The cost is about $26.54 per task and 47.8 minutes per task ().
Anthropic ships Fable 5.2 and Mythos 5.2
Fable is generally available. Mythos has fewer safeguards and is limited to verified cyber and life-sciences organizations (;).
Price cut ahead of IPO
Pricing drops from $10/$50 to $8/$40 per million input/output tokens. Yahoo Finance links the launch to Anthropic's expected November IPO, now targeting a $1–1.5 trillion valuation ().
Mollick publishes "Fable 5.2 and Jagged AGI."
He tests the model on generality: a cinematic one-shot demo built in code, a hardware hack, and a business launch. His conclusion is "almost" AGI, still jagged ().
Capabilities and access
- Model ID:
claude-fable-5-2. Context window is 1M tokens and maximum output is 128K. Inputs are text and images, output is text. - Platforms: Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Retirement is no sooner than Nov 9, 2027 (Claude docs).
- Fast mode: $16/$80 per million tokens (Anthropic).
Read the full section
- Model ID:
claude-fable-5-2. Context window is 1M tokens and maximum output is 128K. Inputs are text and images, output is text. Knowledge cutoff is June 2026 (Claude docs). - Platforms: Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Retirement is no sooner than Nov 9, 2027 (Claude docs).
- Fast mode: $16/$80 per million tokens (Anthropic).
- Mythos 5.2: same specifications and pricing, but only through the Cyber or Life Sciences Verification Programs (Claude docs).
Technical analysis for researchers and developers
- Thinking: adaptive thinking is always on and cannot be disabled. Depth is set with the
effortparameter, which defaults tohigh(Claude docs). - Breaking API changes: forced tool use (
tool_choiceset toanyortool) returns a 400 error. - Data retention: the model requires 30-day retention and is unavailable under zero data retention unless Anthropic authorizes it (Claude docs).
Read the full section
- Thinking: adaptive thinking is always on and cannot be disabled. Depth is set with the
effortparameter, which defaults tohigh(Claude docs). - Breaking API changes: forced tool use (
tool_choiceset toanyortool) returns a 400 error.temperature,top_pandtop_kshould be omitted. Agent frameworks that rely on either need changes (Claude docs). - Data retention: the model requires 30-day retention and is unavailable under zero data retention unless Anthropic authorizes it (Claude docs).
- Caching: cache reads cost 2.5% of the base input price (Claude docs). That matters because Artificial Analysis measured 275M output tokens on its suite, against a 72.8M median (Artificial Analysis).
- Methodology: the Intelligence Index combines 9 evaluations covering knowledge work, coding, science and reasoning (Artificial Analysis).
Claims and evidence
- Anthropic-reported benchmarks: FrontierCode 2.0 at 49.1% ("xhigh" effort), FrontierSWE v2 at 69.9% versus 62.3% for Opus 5.5, and RoboDojo Sim at 64.9% versus 29.8% for GPT-6 Astra (Anthropic).
- Fermat's Last Theorem in Lean: Anthropic says about 1,000 agents formalized it in 30 hours with minimal human help, while Opus 5.5 managed 26% of the work in 18 hours (Anthropic).
- Mollick's demos: these are single-user, pre-launch anecdotes, not controlled evaluations (One Useful Thing).
Read the full section
- Anthropic-reported benchmarks: FrontierCode 2.0 at 49.1% ("xhigh" effort), FrontierSWE v2 at 69.9% versus 62.3% for Opus 5.5, and RoboDojo Sim at 64.9% versus 29.8% for GPT-6 Astra (Anthropic). Only the overall ranking has been reproduced independently: 65 points, 5 ahead of Gemini 4 Carbon (The Decoder).
- Fermat's Last Theorem in Lean: Anthropic says about 1,000 agents formalized it in 30 hours with minimal human help, while Opus 5.5 managed 26% of the work in 18 hours (Anthropic). Independent confirmation is not available.
- Mollick's demos: these are single-user, pre-launch anecdotes, not controlled evaluations (One Useful Thing).
Context and prior work
Mollick's 2025 "Jagged AGI" essay described o3 and Gemini 2.5 Pro as superhuman in places but unreliable at some mundane tasks (Gigazine). Gary Marcus criticized similar AGI claims about o3 at the time (Marcus on AI). The new essay keeps the jaggedness idea but changes what it means.
Read the full section
Mollick's 2025 "Jagged AGI" essay described o3 and Gemini 2.5 Pro as superhuman in places but unreliable at some mundane tasks (Gigazine). Gary Marcus criticized similar AGI claims about o3 at the time (Marcus on AI). The new essay keeps the jaggedness idea but changes what it means. The gaps are no longer obvious errors or hallucinations; they now sit in execution: infrastructure, sales and the many moving pieces of running a business (One Useful Thing).
Limitations, safety and contested findings
- Business test: Mollick stopped the experiment before Fable could run the business.
- Speed is contested: Anthropic claims outputs up to 30% faster than Fable 5.1 (Anthropic).
- Deliberate capability limits: Anthropic held AI R&D capabilities, such as kernel optimization and architecture work, to Opus 5.5 level.
Read the full section
- Business test: Mollick stopped the experiment before Fable could run the business. He worried it would take money without delivering quality, and the infrastructure it assumed did not yet exist (One Useful Thing).
- Speed is contested: Anthropic claims outputs up to 30% faster than Fable 5.1 (Anthropic). Artificial Analysis rates the model "notably slow" per task, at 47.8 minutes against a 12.7-minute median (Artificial Analysis).
- Deliberate capability limits: Anthropic held AI R&D capabilities, such as kernel optimization and architecture work, to Opus 5.5 level. Fable routes most cybersecurity requests to Opus 4.8 responses (Anthropic).
- Evaluation caveat: Anthropic says the model voices suspicion that it is being tested, which makes it harder to tell how it behaves in real use (Anthropic).
Business and practitioner implications
- Cost: per-token prices fell, but Artificial Analysis puts the average cost per task at 2.6× Opus 5.5 (The Decoder). Budget per task, not per token.
- Where it fits: Mollick's results point to long-horizon research and planning as the strength.
- Market: Yahoo Finance describes a pricing war in which enterprises mix closed and open-weight models to avoid lock-in (Yahoo Finance).
Read the full section
- Cost: per-token prices fell, but Artificial Analysis puts the average cost per task at 2.6× Opus 5.5 (The Decoder). Budget per task, not per token.
- Where it fits: Mollick's results point to long-horizon research and planning as the strength. Operations still need human oversight; he estimated a few hours of human work a week (One Useful Thing).
- Market: Yahoo Finance describes a pricing war in which enterprises mix closed and open-weight models to avoid lock-in (Yahoo Finance).