Anthropic added Claude Fable 5.2 on November 9, positioning it as a higher‑scoring alternative to Claude Opus 5.5. The headline claim is that Fable outperforms Opus on internal benchmarks, yet its standard API rates are twice as high ($8 / M input tokens, $40 / M output tokens versus $4 / M and $20 / M). Independent tests on three engineering‑focused scenarios reveal that Opus matches Fable’s functional scores while using far fewer tokens, costing less, and completing faster.
What Changed in Claude Models
Anthropic’s public numbers show Fable achieving 49.1% on FrontierCode 2.0 (high effort) versus Opus’s 45.5% at max effort, and 69.9% versus 62.3% on FrontierSWE v2. The pricing table confirms Fable’s input and output rates are exactly double Opus’s. Crucially, the API defaults differ: Fable runs at high effort while Opus defaults to medium effort, meaning the out‑of‑the‑box experience already favors Opus in token efficiency.
Operational Impact on Coding and Agent Workloads
Three custom tests were executed once per model using default settings.
- Cron scheduler: Both models scored 50/50 across three spec versions. Fable generated 2–3× more reasoning tokens, leading to total costs of $0.87, $1.12, and $1.30 versus Opus’s $0.22, $0.38, and $0.39. Execution time was roughly double for Fable (≈660 s total) compared with Opus (≈358 s).
- Invoicing repo fix: Again a perfect 50/50 score, but Fable’s token usage produced costs of $0.58, $1.39, and $1.18, while Opus stayed under $0.20 per run. Opus completed each version in under a minute; Fable took up to four minutes.
- Incident investigation: The test emphasized careful log analysis. While the source cuts off before reporting final numbers, the methodology mirrors the earlier patterns—extra reasoning tokens from Fable did not translate into higher scores.
Across all scenarios, the additional reasoning generated by Fable did not improve functional outcomes, yet it inflated token consumption, cost, and latency.
Architectural and Implementation Considerations
For teams integrating LLMs into CI pipelines, automated agents, or SRE tooling, the following factors emerge from the data:
- Token budgeting: Model choice directly impacts budget forecasts. Doubling the per‑token price without a commensurate quality gain can quickly exceed allocated spend.
- Default effort levels: High‑effort defaults increase token counts. Adjusting the
effortparameter to match workload needs can reduce cost without sacrificing accuracy. - Latency budgeting: Opus’s faster turn times (≈80–150 s per test) make it more suitable for real‑time or near‑real‑time automation, whereas Fable’s longer runtimes may be acceptable only for batch‑oriented jobs.
- Observability: Monitoring token usage per request becomes essential to detect when a model’s reasoning overhead outweighs its benefits.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
- Run a small‑scale benchmark on your own workloads before adopting a higher‑priced model.
- Explicitly set the
effortlevel to the lowest acceptable tier; the default may be unnecessarily expensive. - Incorporate token‑count logging into your observability stack to catch cost overruns early.
- Prioritize models that deliver the required functional score with the smallest token footprint, especially for high‑frequency automation.
- Re‑evaluate model selection periodically as pricing or benchmark updates are announced.


