Claude Fable 5.1: The Benchmark Story Everyone Will Cite — and the Effort Story Everyone Will Miss

Anthropic shipped its latest flagship today, and the headline framing is confident: the company is introducing Claude Fable 5.1 and Claude Mythos 5.1, calling them the world's most advanced models for coding and knowledge work. That's the kind of claim every frontier lab makes at every launch. What makes this one worth examining is that the benchmark deltas are unusually lopsided — and the real strategic move is hiding in the pricing table, not the leaderboard.

Let's get the scores on the table, because they are genuinely striking in one category and merely solid in the others:

- Agentic scientific research (Terminal-Bench-Science 0.1): This is the outlier. Anthropic reports Fable 5.1 scoring 52.6%, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol in its evaluation setup. A doubling on a hard agentic benchmark in a single point release is rare. It's worth noting the error bars, which Anthropic itself flags: the standard error is ±3.5–4.5 points per model, and the public leaderboard reports Opus 5 at 30.0% and Fable 5 at 21.4%, which Anthropic's setup reproduces at 29.0% and 24.7% respectively, both within noise. Even at the pessimistic end of that band, the jump is real. - Terminal coding (Terminal-Bench 4.0): Fable 5.1 scores 55.8%, versus 42.0% for Fable 5 and 52.3% for Opus 5 — and Mythos 5.1 reaches 60.9% on the same benchmark when operating under its more permissive cyber safeguards. That Mythos gap is one of the more honest data points in the release, and Anthropic addresses it directly: the two are the same underlying model, the gap reflects tasks where the earlier, less precise cyber safeguards intervened, and with today's safeguard improvements the company expects the difference to shrink considerably. - Knowledge work and workflows: Anthropic reports a GDPval-AA v2 score of 1,853, versus 1,824 for Opus 5 and 1,723 for Fable 5; on AutomationBench, Fable 5.1 scores 31.4% against 17.1% for Fable 5 and 26.9% for Opus 5; and on CursorBench 3.2.0 it reaches 73.4%.

For context, the Fable line was already at the top of most boards. Fable 5 leads SWE-bench Verified at 95.0% and SWE-bench Pro at 80.0%, so the 5.1 story is less about reclaiming coding crowns and more about extending the lead where tasks are long and messy.

Here's what I think most day-one coverage will underweight. The single most consequential line in the announcement isn't a score — it's this: when set to Low or Medium effort, Fable 5.1 achieves results similar to or better than Fable 5's at a much lower cost, and it defaults to High effort in Claude Code and Medium in Claude Cowork and on Claude.ai.

Read that again as an economist rather than a benchmark-watcher. Anthropic is effectively saying: yesterday's frontier is now available at a discount tier of today's model. Combine that with the pricing change — an estimated 25% cost reduction for typical workloads wherever usage is billed by token, driven by cheaper cache reads, with cache reads at a quarter of the cost while input and output prices stay the same — and the release starts to look like a repricing of agentic labor, not a model refresh.

VentureBeat gestured at this framing, describing the release as an attempt to solve three intertwined enterprise problems: making agents capable enough to finish difficult work, economical enough to leave running for hours, and governable enough to trust with sensitive systems. I'd push it further: cache reads are the dominant cost of long-running agents specifically, because an agent re-reads its own accumulated context on every step. Cutting cache-read prices by 75% is a targeted subsidy for exactly the multi-hour autonomous sessions the benchmarks are showcasing. The pricing and the Terminal-Bench-Science number are the same bet expressed in two languages.

The market is already responding to that math. Cognition, maker of Devin, said it's moving its Opus 5 traffic in Devin to Fable 5.1 on launch day, noting the model matched or edged out Fable 5 at a lower cost per task, and that the new cache read pricing finally makes a Fable-class model economical for workloads it had kept on Opus, starting with code review.

Two qualitative claims deserve scrutiny in the coming weeks, because if true they matter more than any leaderboard:

1. Honesty under failure. AWS's launch post claims the model is more honest: if it gets stuck it says so, and it is less likely to disable a failing test to pass. Test-gaming is the quiet plague of agentic coding, and benchmark suites mostly don't measure it. Independent replication of this claim should be priority one for evaluators.

2. Legibility over long horizons. Anthropic asserts that while prior models became hard to follow the longer they worked, Fable 5.1 remains readable over long, multi-step tasks. If agents are going to run unattended for hours, the audit trail is the product.

There's also one flashy anecdote making the rounds: in testing by the investment firm Millennium, Fable 5.1 reportedly found the cause of a rare crash on their internal systems that no engineer or other model had been able to explain. Compelling — and, like all vendor-supplied war stories, unverifiable from the outside.

Fable's benchmark numbers come with an asterisk that's actually to Anthropic's credit for disclosing: Fable 5.1 was evaluated with production safeguards enabled; on tasks where safeguards intervened it scored zero on OSWorld 2.0, other intervened cybersecurity tasks were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5, which likely reduces reported performance. In production, the same mechanism applies — flagged queries in these domains are automatically routed to less capable models, though you won't be charged Fable prices for rerouted requests. And there's a governance cost: using Fable requires 30-day data retention for safety monitoring by default. For regulated buyers, that retention clause may matter more than any benchmark.

Finally, remember the price of frontier status: Fable 5.1's ordinary input and output are twice as expensive as Opus 5's, though its cached input is half the cost of Opus 5's cache reads. Even Anthropic's own docs suggest restraint — start with Claude Opus 5 for most workloads, and reach for Fable 5.1 for demanding reasoning and long-horizon agentic work, or when Opus 5 at higher effort still falls short.

The benchmark chart says Fable 5.1 doubled the state of the art in agentic science. The pricing table says something more interesting: Anthropic just made last quarter's frontier a mid-tier commodity and repriced the marginal hour of autonomous agent work downward across the board. Three months from now, the number people remember probably won't be 52.6%. It'll be whatever happened to their agent compute bill.