VynarisEarly betaGet your API key

Blog · 2026-09-02 · Vynaris Team

Claude Fable 5.1 Low Effort Beats Fable 5 Max at 74.8% Lower Task Cost

Claude Fable 5.1 Low effort scores 26.3% at $11.10 per task. Fable 5 Max costs $44.10 at 24.7%. The cheaper model wins on both axes.

Claude Fable 5.1 at Low effort scores 26.3% on Terminal-Bench-Science 0.1 at $11.10 per task. Claude Fable 5 at Max effort scores 24.7% at $44.10. The newer model's cheapest setting beats the old model's most expensive setting on both axes: 74.83% lower cost and 1.6 points higher score. Prices verified 2026-09-02.

TL;DR

What we computed and why

Anthropic's Fable 5.1 launch page publishes a Terminal-Bench-Science 0.1 cost-quality chart with five effort levels per model. The chart embeds a CSV with mean cost per task and score for all ten points. We extracted that CSV, computed cost per score point for each, and compared every Fable 5.1 effort level against its Fable 5 counterpart at the same setting and against Fable 5's maximum.

The question is not whether Fable 5.1 is better. Anthropic's own benchmark table says it is. The question is which effort level a buyer should select, and whether the cheapest setting that clears a quality threshold is also the cheapest overall. The cost-quality frontier answers that.

Assumptions table

Assumption              Value                           Source
----------------------  ------------------------------  ---------------------------------------------------------------------------------------------------
Benchmark               Terminal-Bench-Science 0.1      Anthropic launch page
Harness                 Claude Code, 3 trials per task  Anthropic chart footnote
Standard error          ±3.5 to 4.5 points per model    Anthropic chart footnote
Fable 5.1 input price   $10/MTok                        [Claude API pricing](https://platform.claude.com/docs/en/about-claude/pricing), verified 2026-09-02
Fable 5.1 output price  $50/MTok                        Same, verified 2026-09-02
Fable 5.1 cache read    $0.25/MTok                      Same, verified 2026-09-02
Fable 5 input price     $10/MTok                        Same, verified 2026-09-02
Fable 5 output price    $50/MTok                        Same, verified 2026-09-02
Fable 5 cache read      $1.00/MTok                      Same, verified 2026-09-02
Cost unit               Mean USD per task               Anthropic chart x-axis
Score unit              Percent of tasks solved         Anthropic chart y-axis

The cost figures are Anthropic's reported mean cost per task on their harness, not our derivation from token counts. Anthropic does not publish the token traces behind each point. We treat the ten cost/score pairs as first-party receipts and compute ratios from them.

The full frontier

Effort  Fable 5.1 cost  Fable 5.1 score  Fable 5.1 cost/pt  Fable 5 cost  Fable 5 score  Fable 5 cost/pt
------  --------------  ---------------  -----------------  ------------  -------------  ---------------
Low     $11.10          26.3%            $0.422             $17.10        12.3%          $1.390
Medium  $14.90          35.7%            $0.417             $25.00        21.4%          $1.168
High    $20.30          40.0%            $0.508             $34.30        25.0%          $1.372
xHigh   $31.80          49.5%            $0.642             $36.00        23.4%          $1.538
Max     $37.90          52.6%            $0.721             $44.10        24.7%          $1.785

Source: Anthropic Terminal-Bench-Science 0.1 chart, embedded CSV extracted 2026-09-02. Standard error ±3.5 to 4.5 points per model.

Fable 5.1 is monotonically increasing: each effort level costs more and scores higher. Fable 5 is not. Its score drops 1.6 points from High (25.0%) to xHigh (23.4%) while cost rises $1.70, then gains only 1.3 points from xHigh to Max while cost rises $8.10. A buyer running Fable 5 at xHigh pays more for less than at High.

Same-effort comparison

Effort  Fable 5.1 vs Fable 5 cost  Fable 5.1 vs Fable 5 score  Cost-per-point saving
------  -------------------------  --------------------------  ---------------------
Low     -35.09%                    +14.0 pts                   -69.64%
Medium  -40.40%                    +14.3 pts                   -64.27%
High    -40.82%                    +15.0 pts                   -63.01%
xHigh   -11.67%                    +26.1 pts                   -58.24%
Max     -14.06%                    +27.9 pts                   -59.64%

At Low, Medium, and High effort, Fable 5.1 cuts cost by 35% to 41% while adding 14 to 15 score points. The cost-per-score-point saving is 63% to 70% at those levels. At xHigh and Max the cost gap narrows to 12% to 14% because Fable 5's cost flattens, but the score gap widens to 26 to 28 points.

The headline: cheapest Fable 5.1 beats most expensive Fable 5

Fable 5.1 Low at $11.10 and 26.3% beats Fable 5 Max at $44.10 and 24.7% on both axes. The cost saving is $33.00 per task, or 74.83%. The score gain is 1.6 points.

Fable 5.1 Medium at $14.90 and 35.7% is 66.21% cheaper than Fable 5 Max and 11.0 points higher. If your eval threshold sits between 25% and 35%, Medium is the setting that clears it at the lowest cost.

The best cost per score point across all ten measurements is Fable 5.1 Medium at $0.417 per point. The worst is Fable 5 Max at $1.785 per point. That is a 4.28x spread on the cost-per-point axis.

Put your own task counts and token mix into the LLM cost calculator before extrapolating these per-task figures to a monthly bill.

Standard error: the score gap is within noise, the cost gap is not

Anthropic states a standard error of ±3.5 to 4.5 points per model. Using the wider bound:

Model and effort  Point estimate  Range
----------------  --------------  --------------
Fable 5.1 Low     26.3%           21.8% to 30.8%
Fable 5 Max       24.7%           20.2% to 29.2%

The ranges overlap from 21.8% to 29.2%. The 1.6-point score advantage of Fable 5.1 Low over Fable 5 Max does not clear the noise threshold. A buyer should not claim Fable 5.1 Low is definitively more accurate than Fable 5 Max on this benchmark alone.

The cost advantage is different. The $33.00 per-task gap is a measured mean, not a statistical estimate with a published confidence interval. Anthropic reports it as a deterministic billing figure on their harness. No standard error applies to the dollar column.

What this means for routing

Three decisions follow from the frontier.

First, if you are running Claude Fable 5 at Max effort on Terminal-Bench-class tasks, migrate to Fable 5.1 Low. The cost falls 74.83% and the point estimate rises 1.6 points. Even under the standard-error overlap, you are paying $33.00 less per task for a score that is statistically indistinguishable.

Second, if your workload needs a higher accuracy threshold than 26.3%, step up to Fable 5.1 Medium. At $14.90 and 35.7%, it costs $3.80 more than Low but gains 9.4 points. The cost-per-score-point at Medium ($0.417) is marginally better than at Low ($0.422), making Medium the price-performance optimum across the full frontier.

Third, do not assume the effort level that was optimal on Fable 5 carries over. Fable 5's non-monotonicity at xHigh means a buyer who selected xHigh for a marginal score gain was paying $1.70 more for 1.6 points less. On Fable 5.1, every step up the effort ladder buys more score. Re-run your eval at each effort level rather than inheriting the old setting.

For model routing rules that key on effort level, the right-sizing logic needs to change. A rule that sent hard tasks to Fable 5 Max should now send them to Fable 5.1 Medium or High, not Max, unless the eval threshold exceeds 49.5%.

Honest tradeoff: the cost win is real, the quality win is not proven

The 74.83% cost saving is a first-party billing figure. It holds regardless of standard error. But the 1.6-point score advantage of Fable 5.1 Low over Fable 5 Max sits inside the ±4.5-point noise band. If your decision criterion is "strictly higher accuracy with statistical confidence," this chart does not prove that. You need your own task success rate eval on your workload, not Anthropic's aggregate.

The stronger claim is the same-effort comparison. At every effort level, Fable 5.1 scores 14.0 to 27.9 points higher than Fable 5 at lower cost. Those gaps exceed the standard error. A buyer migrating from Fable 5 High to Fable 5.1 High gets 15.0 more points for 40.82% less money, and that quality gain is outside the noise.

The cache-read price cut that drove the news analysis we published earlier today explains part of the cost reduction. Fable 5.1's cache reads cost $0.25/MTok versus $1.00 on Fable 5. On a coding agent harness that re-reads large context prefixes, cache reads dominate the bill. Our prompt-caching break-even analysis covers the write side: a cheaper read does not rescue a prefix that changes before reuse. But the effort-level frontier shows that Fable 5.1 also uses tokens more efficiently per task at every effort setting, not just cheaper per token. The cost-per-score-point improvement ranges from 58.24% to 69.64% across effort levels, which exceeds what a 75% cache-read cut alone would produce on a fixed-token workload.

Caveats

Sources