Skip to content

Anthropic

Fable 5.2 matched Opus 5.5 on nine tasks at almost 4x cost

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: AI prices →
Image: thenewstack.io

Claude Fable 5.2 and Claude Opus 5.5 got the same scores on all nine coding and agent runs in a comparison The New Stack published on November 15, so on short jobs like these Opus matched Fable at about a quarter of the estimated API cost, $2.29 against $8.98. Both models ran at Anthropic’s API defaults, which put Fable at high effort and Opus at medium, and each test ran once. So the comparison can’t say how much of the gap comes from the effort setting, and the author says it can’t prove or disprove Anthropic’s benchmarks either.

Fable’s standard API rates are twice Opus’s: $8 per million uncached input tokens and $40 per million output tokens, against $4 and $20. The rest of the gap comes from volume. Across the nine runs Fable produced 168,099 output tokens to Opus’s 78,293, more than twice as many, and in The New Stack’s grading the extra reasoning raised no score. Fable also took 1,579 seconds in total against 701.

Where the tokens went

Every first version ended in a perfect tie, so The New Stack’s author made each job harder and reran it twice. The scores never split. The invoicing repo fix, the job closest to what agents do for developers daily, showed the widest cost and time gap of the three. Each model worked as an agent in a small invoicing library with interacting bugs in rounding, discounts, tax and shipping, plus one wrong test it should report but not edit. The harder versions added buy-two-get-one promotions, weight-based shipping, refunds, currency conversion and invoice numbering, each with its own bugs. Both models fixed the bugs the grader covered and reported the wrong test without editing it, every time. Opus finished each version in under a minute, between 28 and 51 seconds, on 519 to 1,178 reasoning tokens. Fable spent 5,909 to 14,022 reasoning tokens, took 106 to 223 seconds and cost eight times as much across the three versions.

On the cron scheduler, the models turned a written spec into code and then took a change request, graded by a hidden suite of 50 tests. In the hardest version they had to handle time zones and daylight saving time by hand, without Python’s time zone library. In The New Stack’s runs both scored 50 of 50 on every version, and Fable wrote two to three times as much reasoning each time. Its total came to 661 seconds and an estimated $3.29, against 358 seconds and $0.99 for Opus.

The incident investigation was the job the author expected to favor Fable. Each model used a terminal to trace a payments outage through logs, metrics and deployment records seeded with decoy deploys, a server clock running seven minutes slow and, in the last version, a server logging in Chicago time. Fable’s advantage showed up only as fewer turns, 9 to 11 against Opus’s 12 to 14, and it still cost $2.54 to Opus’s $0.91 across the three versions, the test’s figures show.

Effort settings and Anthropic’s own numbers

Effort is the setting that decides how deeply Claude thinks before answering, and with that its latency and cost. It runs low, medium, high, xhigh and max, and adaptive thinking stays on for both models and can’t be turned off, according to Anthropic’s platform docs. Called through the API with no effort value set, Fable 5.2 runs at high and Opus 5.5 at medium, the pairing The New Stack measured. In Claude Code and the Claude apps, Fable 5.2 starts at medium, so anyone comparing the two there is running a different pairing from this test.

When Anthropic released Fable 5.2 on November 9, it reported 49.1% on FrontierCode 2.0 at xhigh effort against 45.5% for Opus 5.5 at max, and 69.9% against 62.3% on FrontierSWE v2. Anthropic also says Fable scored 45.3% at high effort for roughly half the cost per task of that max-effort Opus run. “Anthropic has cut the price of a token, not the price of a thought,” Sanchit Vir Gogia of Greyhound Research said, per Computerworld.

The same launch announcement tells developers to start most coding and knowledge work on Opus 5.5, which it calls its “strongest daily driver”, and to move up to Fable for the most ambitious, open-ended technical work. The Claude developer blog’s model guide is more precise: start on Opus 5.5 at medium, and test Fable 5.2 before raising Opus past high. Opus at medium already scored full marks on all nine of The New Stack’s runs, so on jobs like these that switching point never came. Anthropic’s half-cost comparison sets Fable against Opus at max.

Fable 5.2 also requires 30-day data retention and isn’t available under zero data retention, the arrangement where Anthropic stores no prompts or outputs, unless Anthropic expressly authorizes it. A team that runs Opus under that arrangement needs the authorization before it can move any work up to Fable.

The New Stack’s author, who calls the tests “hard but short”, settles on a routing rule: “I’d start with Opus 5.5 and move up to Fable 5.2 only when Opus stops producing ideal results.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.