Why Claude Fable 5.1’s 75% Cheaper Cache Reads Still Lead to Higher Costs

Release Overview

Anthropic announced Claude Fable 5.1 and Mythos 5.1, continuing a rapid three‑month iteration cycle after the original Fable 5 launch.

Performance: Benchmarks Double

Terminal‑Bench‑Science 0.1 : 52.6% improvement over Fable 5’s 24.7%.

Terminal‑Bench 4.0 (programming) : 55.8% (Mythos 5.1 60.9%) surpassing Opus 5’s 52.3%.

AutomationBench (workflow automation) : 31.4% vs 17.1%.

CursorBench 3.2.0 : 73.4% at max effort.

Independent evaluator Artificial Analysis ran the model in pre‑release and reported that Fable 5.1 (max) achieved an Intelligence Index score of 66, the highest measured so far, beating Claude Opus 5 (63), Claude Fable 5 (62), GPT‑5.6 Sol (61) and Grok 4.6 (61). About 4% of output tokens came from a server‑side fallback to older models, so the score is not purely from Fable 5.1.

HLE and Other Scores

On the HLE benchmark Fable 5.1 scored 59.1% versus 55.5% for Fable 5. It also set record scores on Terminal‑Bench v2.1 (91.4%) and SciCode (62.0%). The τ³‑Banking test was 9 points higher than Fable 5.

Felix Rieseberg noted that Fable 5.1’s writing style feels more natural, with less overuse of bold, lists and headings.

Pricing Paradox: Cheaper Cache Reads, Higher Overall Cost

Fable 5.1’s per‑million‑token pricing remains $10 for input and $50 for output, but cache‑read cost dropped from $1 to $0.25 (75% reduction). Anthropic’s internal data suggest a ~25% total cost reduction for typical workloads and up to 45% for highly agentized tasks.

Artificial Analysis calculated that a single Intelligence Index task with Fable 5.1 (max) costs $3.76, 20% more than Fable 5’s $3.14 and 1.6× the $2.34 cost of Opus 5 (max). The higher cost stems from output token usage about 1.7× that of Fable 5, which consumes the savings from cheaper cache reads.

Effort Levels and Token Efficiency

Fable 5.1 offers five effort tiers, with output tokens ranging from 13.1 M (low) to 143.7 M (max), an 11‑fold span, while Intelligence Index scores rise from 58 to 66. On the token‑vs‑score Pareto front, Fable 5.1 occupies the upper‑right corner, unmatched by other models at comparable token counts.

At the xhigh tier the model scores 65 for $2.72 per task, $1.04 cheaper than the max tier, showing that the highest‑cost tier is not always the most cost‑effective.

AI/ML API Comparison: Fable 5.1 vs GPT‑5.6 Sol

Fable 5.1 generated five three.js scenes (futuristic mansion, crystal chandelier, water‑park, Lamborghini garage, helipad) using 69,108 output tokens, costing $5.69.

GPT‑5.6 Sol produced the same prompts with 29,086 output tokens, costing $0.88.

Fable 5.1 is roughly six times more expensive but delivers richer, more detailed scenes; a notable bug was a water‑park without water.

Agent‑Class Tasks: Near Parity with Opus 5

In agentic knowledge‑work benchmarks, Fable 5.1 (max) achieved the highest scores: GDPval‑AA v2 1853 Elo (+130 over Fable 5) and AA‑Briefcase 1694 Elo (+122). Compared with Opus 5, the advantage is modest—GDPval‑AA v2’s lead falls within the confidence interval and AA‑Briefcase is essentially tied (1694 vs 1685). The tests used Artificial Analysis’s open‑source Stirrup framework.

On AA‑Omniscience, Fable 5.1 attempted 93.4% of questions (Opus 5 87.8%) and achieved a 67.2% accuracy—the highest recorded—surpassing Fable 5’s 65.4%. However, higher attempt rates led to more hallucinations, resulting in a net parity with Fable 5.

Enterprise Frontier Safeguards (EFS)

The biggest controversy at Fable 5 launch was Anthropic’s 30‑day data‑retention requirement for Mythos‑level models, problematic for regulated sectors. Anthropic responded with Enterprise Frontier Safeguards, which stores data in the customer’s own cloud infrastructure and lets customers perform manual reviews. Over 100 customers, including major banks, co‑developed the solution.

EFS will roll out in phases this fall; qualified customers can already use Fable 5.1 with zero data retention at no extra Anthropic fee, though they bear their own cloud storage and I/O costs.

Research Demonstrations

Fable 5.1 used 30‑year‑old Magellan radar data to train a neural network that generates high‑resolution (300 m) DEMs of one‑third of Venus’s surface, improving spatial resolution from 10‑20 km to 2‑3 km and raising height accuracy by up to 25%.

Mythos 5.1 showed early drug‑discovery capability: after integrating open‑source protein‑design tools, it produced high‑affinity binders that were ten times stronger than the best submissions in the Adaptyv Bio competition, with a success rate near 50% (industry baseline 10‑15%).

In computational biology, Mythos 5.1 wrote custom GPU kernels and cached intermediate results, accelerating inference of seven open‑source deep‑learning models by 1.4‑2.5× and cutting whole‑genome‑analysis GPU costs by 30‑60%.

Other Notable Changes

Safety guard false‑positive rate : network‑security false triggers down ~60%; basic biology/medicine false blocks down ~85%. Model now permits software‑vulnerability discovery (but not exploit generation).

Anti‑distillation mechanism : new API accounts can no longer edit Claude’s prior context while preserving chain‑of‑thought records in multi‑turn dialogs.

EU AI Act compliance : invisible watermarks added to output text, API detection currently limited to EU‑required organizations.

Context window : 1 M tokens, supporting image and text inputs, aligned with Anthropic’s recent models.

API identifier: claude-fable-5-1. Mythos 5.1 remains limited to US‑based organizations.

Real‑World Cases

Anthropic shared feedback from 22 early partners. Millennium used Fable 5.1 to locate a rare internal system crash root cause that had eluded engineers for years, successfully disassembling an external library and matching a core dump to a bug.

Jane Street’s quantitative research lead reported better readability on long tasks, while Cognition switched traffic to Fable 5.1 on launch day because the cheaper cache reads finally made code‑review use cases economically viable.

Takeaways

Fable 5.1’s benchmark results are impressive, yet three tensions remain:

Price paradox : cheaper cache reads are offset by a 1.7× increase in output tokens, making the max effort tier costly ($3.76 per task) compared with Opus 5 ($2.34). Lower effort tiers improve cost‑effectiveness.

Agent‑task parity : performance is roughly on par with Opus 5, but Opus 5 remains cheaper; the advantage lies in writing style and adherence to scoring criteria.

EFS rollout speed : phased deployment means regulated industries still face a gap before zero‑retention usage is broadly available.