The New Stack ran Claude Fable 5.1 and Fable 5 on four real tasks, and both scored 24 out of 24

The New Stack published a hands-on comparison of Claude Fable 5.1 and its predecessor on 5 September. Reporter Jessica Wachtel ran both models on four tasks built to mimic day-to-day work and tracked every token: both models scored a perfect 24 out of 24, and the new model used 70 percent more tokens doing it.

Anthropic launched Fable 5.1 on 1 September, calling it "our most advanced model for coding and knowledge work". The announcement leans on Terminal-Bench-Science, an agentic research benchmark on which Anthropic says the new model scores 52.6 percent against Fable 5's 24.7 percent. Both models cost the same: $10 per million input tokens and $50 per million output tokens.

Wachtel's four tasks: excluding five documented bad rows from experiment data and computing batch averages, fixing two planted bugs in a small Python project until all 8 tests passed, two math problems with verified answers, and a harder sensor-data audit added as a tiebreaker after three rounds of ties. The New Stack reports both models completed every task correctly, including the audit's trickiest traps, a clock running fast and a mid-run hardware swap.

The extra turn is where the money went

Across the full run, Fable 5 used 22,219 tokens for $0.398 in 84.9 seconds, and Fable 5.1 used 37,809 tokens for $0.533 in 82.9 seconds. The gap came from the sensor audit, where Fable 5.1 needed five turns against Fable 5's four. Each turn resends the whole conversation, so the extra turn pushed 23,602 input tokens through the API against 7,940, and the task cost $0.304 against $0.134, more than double.

Wachtel states two caveats. Her short tasks used no prompt caching, so the piece does not test Anthropic's savings claim tied to cache read prices, which dropped by 75 percent. And Anthropic says it benchmarked 5.1 with its production safeguards on, which sometimes lowered its own scores. Her conclusion: on work like hers, code fixes, data cleanup and analysis with documented gotchas, the upgrade produced results she could not tell apart from the old model, and on a long agentic task it can cost more per run. She adds that Anthropic's numbers place the improvement in hours-long research agents, which her afternoon of testing could not reproduce. The New Stack discloses that its owner, Insight Partners, is an investor in Anthropic.

A team weighing the upgrade gets a copyable method out of this piece. Run your own task set on both models, count the turns, and compare tokens per task, because one extra turn late in a long conversation can double a run's cost while accuracy stays flat. If your agents finish in a few turns, the vendor's benchmark delta may say little about your bill or your output quality.

Source

Claude Fable 5.1 vs. Fable 5: On real work, I couldn't tell them apart, The New Stack, by Jessica Wachtel, 5 September 2026. Anthropic's claims are from its 1 September launch materials as quoted in the article.