Anthropic released Claude Fable 5.2 on 9 November 2026. On paper it is a point release: cheaper, faster, a little stronger. For us it is the first model that makes “general intelligence” read like a description instead of a forecast. This is what it does, what the evidence says and where the feeling runs ahead of the facts.
What Fable 5.2 is
Fable is Anthropic’s tier above Opus: the generally available version of its most capable model. Fable 5.2 shares its underlying model with Mythos 5.2, which is reserved for vetted security and life-science teams. Fable adds safeguards in those two fields and is open to everyone on a paid plan.
It is built for long, demanding technical work: multi-day agent runs, large code migrations, research problems and engineering design. It reads up to a million tokens in one go and can direct a team of smaller models. It costs $8 per million input tokens and $40 per million output tokens, 20% less than its predecessor.
What the benchmarks say
These are Anthropic’s own measurements, so read them as the maker’s case. Each row compares the new model with the one it replaces.
Benchmark | Fable 5.1 to 5.2 |
|---|---|
Remote Labor Index (real freelance-style projects) | 17.9% to 41.3% |
FrontierSWE v2 (agentic coding) | 56.3% to 69.9% |
Terminal-Bench-Science 0.1 (scientific research) | 52.6% to 74.6% |
OSWorld 2.1, partial credit (computer use) | 83.5% to 89.3% |
RoboDojo Sim (robotics in simulation) | 27.6% to 64.9% |
Humanity’s Last Exam, with tools (multidisciplinary reasoning) | 65.6% to 74.2% |
Three results stand out. The Remote Labor Index, which measures how well a model completes real freelance-style projects, more than doubled, and is about twice the score of the next model Anthropic lists. Simulated robotics also more than doubled. On agentic coding it leads GPT-6 Astra, at 65.5%, and Claude Opus 5.5, at 62.3%.
The first independent check arrived the same day. Vals AI ranked it first of 36 models on its overall index, with 71.3%. The lead over the next two models is smaller than the margin of error, and on legal work it placed 20th of 37.
Why it feels like AGI
No single score explains the feeling. Breadth and stamina together do. One model writes and migrates software, runs research, operates a desktop, plans robot tasks in simulation and takes freelance-style projects from brief to delivery.
And it keeps going. Customers quoted by Anthropic describe a security audit that ran for ten to twelve hours, and another that coordinated 170 reviewing agents. Anthropic says the model checks itself on long runs, building controls and flagging results that do not reproduce.
For decades, “general” meant a system that moves between kinds of work the way a capable person does. Completing four in ten freelance-style projects is no longer a rounding error.
Where the feeling runs ahead of the facts
Anthropic does not call it AGI. Its own advice is to start with Claude Opus 5.5 for most everyday work.
Four in ten is not ten in ten. Most of those projects still fail.
It is not first everywhere. Gemini 4 Carbon leads on business workflow automation, 53.9% to 51.4%.
Some numbers flatter. The computer-use score counts partial credit, and the robotics test was run by Anthropic itself.
Some doors stay closed. Offensive security work and most research-level biology are held back by design.
What it could change
For a company, the shift is from asking for answers to delegating outcomes: a migration, an audit, a market study, a working prototype.
Work that was too long or too tedious to staff becomes feasible. One strong model directing a team of cheaper ones changes the cost of ambitious projects. And for a studio that builds robots and connected devices, the jump in simulated robotics is the one we will watch most closely.
What does not change is the part that was always ours: choosing the problem, judging the result and answering for it.
We will test these claims in the Lab and report what holds.
Figures come from Anthropic’s Claude Fable page and from Vals AI, as published on 9 and 10 November 2026. This article was drafted with Claude Fable 5.2 and reviewed by Pixelgrammar.