Twice the Price, Half the Bill

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The price list still calls Fable the expensive model: $8 per million input tokens and $40 per million output, exactly twice Opus 5.5 at $4 and $20.

The number that matters to anyone paying for agents isn't on the price list. On FrontierCode 2.0, a benchmark that asks whether open-source maintainers would merge a model's code changes, Fable 5.2 at high effort scored 45.3%. Opus 5.5 at max effort scored 45.5%. Anthropic says Fable got there "for roughly half the cost per task."

That line puts one of my own rules on trial. My harness runs Opus 5.5 at max effort in every project and calls Fable only when I ask for it, because I judged Fable too expensive. I judged it per token. This launch says to judge it per finished task.

What shipped

In my September post on Fable 5.1, Fable cost exactly twice Opus 5. Both prices fell, so the ratio held.

Anthropic's own advice is not to start with Fable. "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver," then move up to Fable or Mythos, or down to Sonnet or Haiku, as the work requires. Fable can also serve as an Advisor, consulted by Haiku, Sonnet or Opus on harder problems, or as an Orchestrator that delegates simpler tasks to smaller models. I covered the economics of that pattern in Advisor and Executor, back when it was Fable 5 advising Sonnet 5.

The benchmarks, losses included

Anthropic calls Fable 5.2 "a step above Opus 5.5 in almost all the capabilities we evaluated." Its own table shows the losses:

Gemini 4 Carbon wins business workflows. Opus 5.5 wins the Elo half of the knowledge-work benchmark. Against Fable 5.1, FrontierCode 2.0 rose from 32.4% to 49.1%, and the Remote Labor Index more than doubled, from 17.9% to 41.3%. Read the fine print before quoting any row. Scores are "each model's highest-available effort unless otherwise noted," and the notes matter: Fable's 49.1% on FrontierCode is marked xhigh, and Gemini 4 Carbon's 41.1% there is marked medium. Fable's scores are with safeguards on. And RoboDojo Sim, where Fable 5.2's 64.9% more than doubles GPT-6 Astra and Fable 5.1, "was run internally."

Price per token is the wrong unit

The more useful charts in the launch post plot accuracy against cost per task. In four of them, Fable 5.2 roughly matches Opus 5.5's best score, or beats it, for a fraction of what Opus spent:

When Opus 5.5 shipped, I called it Fable's level at 40% of the price. Fable 5.2 returns the favor.

The lever is effort. Anthropic published a same-day guide, Which Claude model should you build with?, and it says so plainly: "Past high, one more effort level can cost as much as a bigger model, so compare both on your own work. If you're past high, test the next model up at a lower effort instead." Higher effort "helps when the model needs to think longer, call more tools, or check its own work more," but it "won't give a model knowledge or skill it doesn't have."

The same guide shows the curve cutting the other way. On Anthropic's runs of Terminal-Bench 3.0, "the previous Fable, 5.1, at max scored 58.0%, while Opus 5.5 at high scored 58.9% with half the tokens." That time the bigger model at max was the expensive corner.

My read: the lesson isn't "Fable is cheaper." It's that max effort is where the waste collects, on either model, and the fix may be a different model at a lower setting rather than more effort on the same one. I wrote about when max loses before; the launch charts put a price per task on it.

On a subscription the math is murkier. On plans that include Fable, it draws on the same weekly allowance, and when 5.1 shipped the plan docs said only that it uses the allowance faster, not by how much.

A thousand agents, one database

The most instructive part of the launch is an engineering write-up, Safely scaling multi-agent teams with Claude Fable 5.2.

In September, Anthropic published the first complete computer-checked proof of Fermat's Last Theorem. A researcher and dozens of agents on an internal model "roughly comparable to Claude Fable 5.1" took 11 days and roughly 6 billion output tokens, and wrote 13 million lines of Lean. They coordinated through Prove2Me, a purpose-built platform from Columbia University, after earlier attempts failed because the agents "lost track of the project's state."

The rerun dropped the platform and the researcher. A team of 1,000 agents on an early version of Fable 5.2 got:

• a shell and a file editor, each in its own sandbox with no network access;
• its own Git branch and the right to open merge requests into main;
• one shared SQL database;
• 60 statements to prove, a one-sentence sketch of the classical route, the definitions those statements use, and 47 lemmas already proved in the September run;
• one structural rule: one lead and 999 helpers.

The team finished in 28.7 hours, with a third of the output tokens and a proof a seventh the length of the original. The launch post gives it as 30 hours and adds the comparison: Opus 5.5, with the same 1,000-agent setup, completed 26% of the formalization in 18 hours "and barely made any more progress after 30."

What the lead did with that freedom is the part worth copying:

• It built its own coordination layer. In the first ten minutes it created five shared tables, among them an announcement channel, an assignment roster, an open notes board and a claimable subtask list. It gave each statement a crew of 16 or 17 helpers with a captain.
• It triaged, then reversed itself. Helpers judged 47 of the 60 statements infeasible within the budget, so 27 minutes in, with three already proved, the lead moved nearly everyone onto the remaining ten, and crews grew to between 46 and 139 helpers. Around 45 minutes in, it wrote, "the team's throughput is far higher than the early 'infeasible-in-budget' verdicts assumed." With those ten done, it sent the team back to the hard ones: "there is no reason to idle."
• It let go. By the third hour it told free helpers, "do not wait for me." By hour 5.6, 73 helpers had written 756 tasks, and only 16 of those helpers were captains.
• Its team grew a tree for the hardest piece. Langlands–Tunnell, which its own captain had given less than a 1% chance of finishing in time, got about 750 helpers ("Honest odds are low.") and a task tree of 2,418 tasks, up to 10 levels deep. It was proved at hour 28.4, and the theorem at 28.7.

The safety design matters as much as the result. Each agent ran on its own lightweight virtual machine, because a single shared machine "hit its limits at about 100 agents." A classifier checked every shell command and file edit before it ran. Anthropic's standing rule is that a blocked tool call stops the whole team until a person reviews it; in this run, a block on a helper would have removed only that helper, and only a block on the lead would have ended the run. An offline monitor reviewed the team's actions as well. "Across more than a million tool calls, none in this run met the bar to be blocked."

Two caveats come from Anthropic's own pages. The launch post says the agents "used no mathematical references, only a list of intermediate theorem statements to score progress"; the engineering post also lists the definitions and the 47 proved lemmas they started with. And its lesson isn't that agents no longer need people: "Agents tend to give up on hard problems and misjudge their difficulty." The harness sent the lead 19 automatic reminders of the time left, "each encouraging the lead to keep going."

For most teams the lesson is the architecture, not the headcount. Isolate every worker. Keep shared state in a store with transactions. Merge through a gate, never into a shared tree by hand. Give the monitor the power to stop everything. And plan for the moment the agents decide the hard part is impossible, because they tend to. My own setup runs one task per session, with worktrees wherever work could collide, for the same reasons at a much smaller scale.

Same model, two doors

Fable's classifiers block "high-risk cyber activities, including penetration testing, exploit generation, and binary-based vulnerability" work, plus most research biology. In early testing, the 55% drop in blocks split into 60% fewer in cybersecurity and 40% fewer in biology.

Mythos is the other door, for organizations vetted through the Cyber Verification Program or the Life Sciences program, and inside Claude Security for Enterprise customers. On CVE Rediscovery, an internal test of whether a model finds a published vulnerability in the code as it was before the fix, Mythos 5.2 scores 89.6%, against 81.3% for Opus 5.5 without safeguards and 78.3% for Mythos 5.1. Mozilla reports that it "roughly doubled the vulnerability discovery metrics of Mythos 5.1" in Firefox scanning campaigns.

Read footnote 2 before repeating the bigger claims. The calibration claim, that earlier models overstated the seriousness of about 1 in 20 findings and Mythos 5.2 of none, rests on "raw findings from one tester's audit," although testers "consistently found" that Mythos 5.2 identified more vulnerabilities than earlier models. Then: "On our own evaluations, vulnerability identification is roughly on par with Opus 5.5."

Two limits are deliberate. During training Anthropic "took steps to limit advances" in the research capabilities that could most directly speed up AI development, so in accelerator kernel optimization and language-model architecture work, Fable 5.2 and Mythos 5.2 "perform at about the level of Opus 5.5," with transparent fallbacks to Opus 5 on a few topics. And Fable 5.2 has Anthropic's strongest protections against distillation, techniques it says cut the data that distillers harvest by 90% to 99%.

One change reaches every API developer: preserved thinking "now applies to all accounts." It prevents callers from editing Claude's prior context to extract its reasoning. Anthropic points developers to those docs for updating integrations; if your code rewrites earlier turns, start there.

The safety numbers worth reading

On an adversarial test of more than 1,800 scenarios, most of them unsolvable without escaping a deliberately weakened sandbox, Mythos 5.2 used credentials or files planted outside its sandbox in 3.1% of scenarios. Opus 5.5: 3.3%. Sonnet 5.5: 5.1%. Mythos 5 and 5.1: 39% and 21%.

In the automated behavioral audit of about 1,900 scenarios, "hallucinating inputs, falsely claiming to have completed a task, omitting important information, and failing to disclose bad or lazy work were all lower" in Mythos 5.2. Early testers describe the same thing in practice. Cognition's Adhyyan Sekhsaria: "Claude Fable 5.2 says it doesn't know things instead of hallucinating." Uber's Andrew Mager ran a 10–12 hour security audit with about 170 reviewer agents and found that the model keeps working "instead of stopping at a plausible checkpoint."

Two results cut the other way, and Anthropic prints both. Mythos 5.2 scored slightly worse than Opus 5.5 on taking harmful actions after wrongly concluding it was in a simulation. And Fable 5.2 and Mythos 5.2 "each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings." External evaluators, including the UK AI Security Institute, Frontier Design and METR, tested the model before release; a footnote adds that none of them reviewed the launch post or the system card.

What else it did

• Physics. Fable 5.2 completed a black-hole calculation laid out in 2003 but too laborious to finish, showing that black holes form abruptly in very small universes too, and found the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity in a simplified universe.
• Hardware. From short prompts, it designed a USB speakerphone, a voice-assistant hub and a handheld virtual pet: circuit boards, enclosures and firmware. Each board was ready for manufacturing in under a day, and all three devices worked "without needing any electrical modifications." On PCB Gym, an internal benchmark run by the board maker Diode, it scored 69% against 25% for Fable 5.
• An operating system. One Anthropic engineer gave Fable 5.2 an empty folder. It "coordinated over 1,000 subagents, working around the clock for eight days," and produced a system that boots on a laptop, with drivers for the disk, keyboard and network, permanent file storage, windows, a login screen and an installer.
• Biology. Mythos 5.2 found a way to predict whole gene-therapy capsids from two open-source protein-folding models, and accurately modeled 28 of the 34 it attempted.
• Code at Anthropic. "Over 90% of the changes merged into Anthropic's core codebase are now written by Claude (with human guidance)." A team of agents on an earlier version of Fable 5.2 moved a critical production service from Python to Rust; it now starts about 26 times faster.

My read

The headlines are a thousand agents and Fermat's Last Theorem. The change that will reach most bills is smaller: effort and model size are now one dial, and the price per token no longer tells you where to set it.

Anthropic still says to start low: Opus 5.5 at medium for everyday work, Fable 5.2 when Opus at high falls short or the run is long and unattended. What the 5.2 numbers add is a reason not to rescue a hard task by turning Opus up to max.

My "too expensive" rule was a per-token judgment. The per-task numbers are Anthropic's, from its own runs, with effort levels that differ between columns. That's enough to re-test the rule, and not enough to replace it. The five steps at the top are the test.