Models

Claude Fable 5.2 Is Cheaper per Token and Pricier per Task, and That Is the Whole Trade

The token price fell by a fifth and the average professional task got more expensive. Anthropic shipped a model that writes its own test cases, and the invoice now reflects how hard it checks itself.

Emily CarterEmily Carter
Heat: 2,100
A sand-coloured effort ladder rising beside the words Fable 5.2 and a note that tokens got cheaper

Dek: The token price fell by a fifth and the average professional task got more expensive. Anthropic shipped a model that writes its own test cases, and the invoice now reflects how hard it checks itself.

Claude Fable 5.2 Cut Token Prices and Raised the Task Bill

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. Input tokens cost $8 per million and output tokens $40, down from $10 and $50 for the 5.1 generation. That's a 20 percent cut at both ends.

Then SemiAnalysis ran the GDPx professional-task benchmark and found the average task cost $88.15, up from $67.16 for Fable 5.1. The newer model took first place at 16.4 percent while spending more per answer.

Cheaper per token. Not per answer.

The reason is in how it works. Fable 5.2 writes hundreds or thousands of test cases, then spins up subagents to check what it produced. Verification costs tokens, and a model that trusts its first draft spends none of them.

So is it cheaper or not? Per word, yes. Per finished job, only if the older model would have got that job wrong.

Claude Mythos 5.2 Is the Same Model Behind a Gate

The two names cover one underlying model. Anthropic says Fable carries strong cybersecurity and biology safeguards so it can be offered to anyone, while Claude Mythos 5.2 goes to organisations approved through its Life Sciences and Cyber Verification Programs.

That's an access control rather than a different product, so the security figures for Mythos describe the family with the guardrails off: 10.2 percent more vulnerabilities found than Opus 5.5, and 14.4 percent more than Mythos 5.1, on an internal CVE Rediscovery test. One security company told Anthropic that one in six confirmed bugs it reported had been missed by every other frontier model tested.

Mythos isn't on the menu unless you're vetted for security or life sciences work.

What Anthropic Says Claude Fable 5.2 Built

The capability claims will get screenshotted, so be precise about what was and wasn't done. In prerelease testing, 1,000 Fable 5.2 agents formalised the proof of Fermat's Last Theorem in 30 hours. The theorem was proved in 1995. Formalising means translating that proof into a form a machine can check, and Anthropic notes that Opus 5.5 completed 26 percent of the same task in 18 hours.

Separately, an engineer asked for a prototype operating system. From an empty folder, the model coordinated more than 1,000 subagents over eight days and produced something that boots on a laptop with its own kernel, disk and network drivers, a login screen and an installer.

There's more: three working electronic devices, and two open problems in black-hole physics. All of it comes from the company, with a system card behind it.

Anthropic's Own Benchmark Table, Read Carefully

Everything below comes from Anthropic, so treat it as a company claim. It also produces output up to 30 percent faster than the 5.1 generation, which matters for agent loops.

Test

Claude Fable 5.2

Comparison

FrontierCode 2.0, Xhigh effort

49.1%

Opus 5.5 at 45.5%

SWE v2

69.9%

Anthropic's own scoring setup

OSWorld 2.1

89.3%

Anthropic's table

Terminal-Bench-Science 0.1

74.6%

Anthropic's table

Company numbers, then. Almost every figure here is Anthropic's, measured against Anthropic's previous model on Anthropic's setup.

Effort Levels, Fast Mode and the 1M Context Bill

The pricing setup decides whether this model is manageable or ruinous. There are five effort levels: low, medium, high, xhigh and max. Medium is the default in the Claude apps and in Claude Code.

Sensible default.

Fast mode generates tokens up to 2.5 times quicker at $16 per million input and $80 per million output, so double the standard rate. Context is 1 million tokens with text and image input, cache writes cost $10 with a five-minute lifetime or $16 with an hour, and cache reads cost $0.20.

Medium effort on a 1M context is a normal day's work. Max effort with fast mode on is a research budget.

Artificial Analysis Puts Claude Fable 5.2 First at 65

Independent numbers look strong. On the Artificial Analysis Intelligence Index v5, Claude Fable 5.2 ranks first at 65 at max effort. Gemini 4 Carbon sits at 60, Opus 5.5 at 59, GPT-6 Astra at 57 and Fable 5.1 at 49.

It records new highs on three tests: HLE-Diamond at 80 percent against a previous 69, AA-Omniscience at 67 against 54, and Terminal-Bench-Science at 73 against 65.

A six-point lead is a real lead, and it was measured at max effort, the setting behind that $88.15 task.

What Simon Willison Noticed About Prompting Claude Fable 5.2

Simon Willison quoted Anthropic's prompting guidance: telling the model it owns and maintains the codebase raised consolidation of duplicated code from under 10 percent of runs to about 75 percent.

Read that as a claim about context. It behaves differently depending on whether it thinks it's a contractor passing through or the person who gets the call when the build breaks. One phrase in a system prompt moved a behaviour from rare to usual.

Who Should Switch, and Who Anthropic Says Shouldn't

Anthropic's own recommendation is the least promotional thing in the post. Start most coding and knowledge work with Opus 5.5, move up to Fable or Mythos for harder problems, drop to Sonnet 5.5 or Haiku 5.5 for easy ones. Fable can also act as an advisor to a cheaper model or as the orchestrator of a team of them.

The subscription terms are tighter than the API terms. Ultra and Team Premium users get resets scaled by tenure, usable until 8 March. Pro and Team Standard users get a one-week trial capped at half the weekly limit, and Fable can consume at most 50 percent of any weekly limit.

Data retention is on by default so Anthropic's safety classifiers can run, and zero data retention is available only to eligible enterprise customers.

Day one access covers Cursor, GitHub Copilot and VS Code, Perplexity, Factory, Hermes Agent through OpenRouter, plus Arena. The release landed as Anthropic's $ANT IPO roadshow began, which is worth one clause and no more.

If you pay per token and your jobs are small, the cut is a straight saving. If you pay per outcome, test the thing you actually run before switching, because the model's habit of checking its own work is exactly what you're now paying for.

Share This Story

Sources

Related AI News

A terminal window with a rising reward curve and a 27B badge beside it
Models

RadixArk Fine-Tuned a 27B Model on 1,200 Tasks and It Beat the Original Everywhere

RadixArk took an existing open model, spent 18 days and about 1,200 hand-built tasks on it, and changed nothing about its shape. The gains are real, and the absolute scores are still low.

Heat: 1,330
A laptop outline with three stacked layer sheets passing through a loop arrow inside it
Models

Inkling 2 Mini Fits on a Laptop by Running the Same Layers Three Times

Thinking Machines Lab built a model that thinks longer by running the same block of layers three times. The trick buys reasoning without extra memory, and the bill comes due somewhere else.

Heat: 1,710
A long teal audio waveform with a single orange accent bar on a dark field
Models

A 0.5B Speech Model Just Matched a 30B One on Real Phone Calls

Reson8 shipped a half-billion-parameter transcription model that gets within a tenth of a point of a 30B rival on noisy phone calls. The interesting part is what that does to a per-minute bill.

Heat: 560
Three neon rings orbiting a small glowing core on a deep blue field
Models

Gemini 4 Neon Lands in the App, and the Model Picker Loses Two Names

Google's fourth Gemini model in a month slots into the middle of the lineup, priced to be used every day. The app now asks subscribers to choose by effort, and two familiar names are gone.

Heat: 1,500