Summary
- Anthropic has released Claude Fable 5.1, claiming top scores on the Terminal-Bench-Science 0.1 benchmark at 52.6%, well ahead of comparable models.
- The model offers five reasoning effort levels — low, medium, high, xhigh, and max — with no option to disable reasoning entirely.
- Cost and latency vary dramatically across tiers: a single max-effort prompt cost USD $3.30 and took nearly 14 minutes; a subsequent animation pass cost an additional $1.37.
- At low and medium settings, the model appeared to skip active reasoning for at least one tested prompt, producing similar output and token counts to higher tiers.
- Organisations deploying AI coding or research tools need to understand which reasoning tier their teams are using, as costs can escalate sharply without explicit controls.
What Anthropic is claiming
Anthropic has released Claude Fable 5.1, positioning it as a step forward for coding, knowledge work, and long-running problem-solving. The company’s own announcement highlights performance on Terminal-Bench-Science 0.1, a benchmark first published on 27 August, where Fable 5.1 scored 52.6% — compared with 24.7% for the previous Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. Other benchmark improvements are described as modest by comparison. The scientific research framing is notable: it signals that Anthropic sees professional and enterprise use cases, not just consumer applications, as the primary growth market.
Five reasoning tiers, no off switch
From a deployment standpoint, the most operationally significant feature is the five-tier reasoning system: low, medium, high, xhigh, and max. Crucially, there is no option to disable reasoning entirely. That design decision means every API call will consume some reasoning capacity, and the resource profile of a given request depends entirely on which tier is selected — a choice that may sit with individual developers or end users rather than with security or procurement teams.
The cost curve is steep
Independent researcher Simon Willison documented the cost and latency spread across all five tiers using a consistent SVG-generation prompt. At low effort, the call took 23.8 seconds and cost approximately USD $0.10. At max effort, the same category of task consumed 65,927 output tokens, ran for nearly 14 minutes, and cost $3.30. A follow-on animation prompt at the high reasoning tier added $1.37. These are single-prompt figures. At organisational scale — across development teams, research functions, or automated pipelines — unconstrained access to upper reasoning tiers represents a material and unpredictable cost exposure.
Behaviour at lower tiers was inconsistent
Willison also observed that at low and medium reasoning settings, the model produced similar token counts and showed no visible reasoning trace for the tested prompt — suggesting the model may elect to skip active reasoning on certain inputs regardless of the tier selected. This inconsistency matters for organisations that assume a lower tier setting equates to predictable, bounded behaviour. The actual compute consumed may not correspond neatly to the tier label.
Output quality does scale with effort
On the purely technical side, Willison’s testing does demonstrate that output quality improved meaningfully at xhigh and max settings. The model’s reasoning traces at higher tiers showed detailed iterative problem-solving — adjusting geometry, reconsidering design choices, and self-correcting errors. That level of deliberative processing may be genuinely useful for complex coding or research tasks, which is precisely the use case Anthropic is marketing. The tradeoff between quality and cost is real, not merely theoretical.
No corroborating sources
It should be noted that the performance figures and cost data reported here come from a single independent researcher’s testing. Anthropic’s benchmark claims have not been independently verified by any corroborating source included in this briefing. CISOs should treat the numbers as illustrative rather than definitive until broader third-party evaluation is available.
Why it matters
As AI models become embedded in development pipelines, research workflows, and internal tooling, the reasoning tier a user selects is no longer just a quality setting — it is a cost and governance variable. A model with no reasoning-off switch and a 33-fold cost difference between tiers creates real exposure if organisations have not set explicit policies around tier selection. Procurement controls, API usage monitoring, and clear guidance on appropriate tier use for different task types should be treated as baseline hygiene for any enterprise deployment of Fable 5.1 or similar tiered-reasoning models.
What to do now
- Audit which reasoning tier is configured as default in any internal tools or API integrations that use Claude Fable 5.1, and confirm that tier selection is not left to individual end users without guardrails.
- Establish spending limits or usage caps at the API level to prevent unconstrained use of xhigh or max reasoning tiers in non-supervised contexts.
- Review whether automated pipelines that call Claude models have hardcoded tier selections, and assess whether those selections are appropriate for the sensitivity and complexity of the tasks involved.
- When evaluating Anthropic’s benchmark claims, seek third-party validation before making procurement or architecture decisions based on Terminal-Bench-Science 0.1 scores, given the benchmark’s recent publication date.
- Document internal guidance on which reasoning tier is appropriate for which class of task, so that developers and analysts have a clear framework rather than defaulting to the highest available setting.
