AnalysisNov 11, 202610 min read4 sources

Fable 5.2's Vals #1 is within the error bar; the effort setting moves scores more than the rank does

Claude Fable 5.2 leads the Vals Index by 0.46 points, against an uncertainty of ±0.96. Vals' own reasoning-effort study shows effort settings swing scores by up to 13.7 points. For buyers, the useful number is cost per point at the effort level you actually run.

In one minute

Claude Fable 5.2 leads the Vals Index by 0.46 points, against an uncertainty of ±0.96. Vals' own reasoning-effort study shows effort settings swing scores by up to 13.7 points.

For buyers, the useful number is cost per point at the effort level you actually run.

Summary

Vals AI ranks Anthropic's Claude Fable 5.2 first of 36 models on its Vals Index at 71.31%, and Epoch AI puts it first of 239 on its Capabilities Index with a score of 174. The Vals lead over Claude Sonnet 5.5 is smaller than the stated margin of error. A Vals study published a week earlier shows that changing the reasoning-effort setting moves scores far more than that gap.

What happened

Vals AI tested Claude Fable 5.2 at High effort. It scored 71.31% ±0.96 on Vals Index v2.2, ahead of Claude Sonnet 5.5 at 70.85% and Gemini 4 Carbon at 70.45%. Vals lists a cost of $17.94 per Index test and a latency of 43 minutes 42 seconds. Token prices are $8 per million input tokens and $40 per million output tokens. The model came first on Finance Agent v2.1 (69.47%), Vibe Code Bench v1.1 (93.77%) and EMB v1.1 (77.38%), and tied for first with Claude Opus 5.5 on Terminal-Bench 4.0 (65.15%). It was weak on HLAB-Vals-1.1 (5.00%, tied #29 of 37) and Tax Agent Bench (32.09%, #27 of 84). On the Vals Science Index it placed second, behind GPT-6 Astra.

Epoch AI separately gives Fable 5.2 an ECI of 174. That is ahead of Claude Opus 5.5 and GPT-6 Astra, both at 169, and Gemini 4 Argon at 167. Fable 5.2 ties other models on several rows, including GPQA Diamond, where it shares 96% with Astra and Argon.

On the Vibe Code Bench leaderboard, Fable 5.2 cost $22.13 per test and took 40m37s. Fable 5.1 scored 90.26% at $33.37 and took 57m40s. Gemini 4 Argon scored 91.91% for $8.11. DeepSeek V4.1 Flash 1017 scored 87.73% for $0.58.

On November 3 Vals published a study of reasoning effort across seven frontier models. At their lowest setting, four of the seven kept at least 80% of their Index score for half the cost or less.

What the archive adds

New Runtime's earlier coverage gives two useful reference points. In October, Gemini 4 Argon was reported as #1 on the Vals Index, finishing agentic tasks at about half the cost of Opus 5.5. Less than a month later Fable 5.2 holds the top spot, so the leadership has changed hands within weeks. The Vals Gemini 4 Argon profile shows a similar uneven spread, with ranks between #4 and #6 across many components.

Anthropic's launch post for Fable 5.2 and Mythos 5.2 makes its own effort-versus-cost case. It says Fable 5.2 at Medium effort beats Opus 5.5 at Max effort on a chart-reading test for less than a tenth of the cost per task. Vals' Finance Agent v2 page warns that v2 and v2.1 results should not be compared as if they were the same benchmark. Version drift is a second reason to be careful when comparing aggregate ranks over time. Third-party leaderboards keep multiplying, among them Cognition's FrontierCode, which Artificial Analysis now runs independently, and Hex's DataBench. Each new one is another place a "#1" can be claimed.

Trend signal

Frontier leaderboards now separate the top models by less than their own noise. Vals' top three models are within 0.86 points of each other, and the leader's stated uncertainty alone is ±0.96. On Epoch the top four are within 7 ECI points. At the same time, the evaluators are publishing cost, latency, token counts and effort settings next to accuracy. Our read is that this is a quiet admission that the accuracy rank no longer separates the models, while the cost per point still does. Fable 5.2's clearest improvement over Fable 5.1 shows up in that second column. On Vibe Code Bench it is 3.51 points better, about a third cheaper per test, and 17 minutes faster.

Why now

1) Anthropic seems to have targeted cost as well as capability. Fable 5.2 is priced below Fable 5.1 ($8/$40 against $10/$50, according to OpenCode's model list). It also uses fewer tokens and less time on Vibe Code Bench (34.71M tokens against 57.39M). 2) Coding and agentic post-training appear to be the focus. Fable's first-place finishes cluster in Vibe Code, Terminal-Bench, Finance Agent and ProgramBench, which matches Anthropic's emphasis on coding in its launch post. 3) The frontier has compressed. Several labs now ship models within a few points of each other, so aggregate indices change hands within weeks, as Gemini 4 Argon's October Vals lead showed. 4) Evaluators now report effort settings. Vals' November 3 study made effort a visible variable, which exposes how much a "#1" depends on configuration.

Connections to other events

Here is the arithmetic. The Vals lead is 0.46 points, about half of Fable's own ±0.96, so the rank is not statistically settled. Vals ran Fable at High effort, not Max. In the same lab's study, dropping from the leaderboard setting to the lowest setting cost Opus 5.5 13.7 points and GPT-6 Astra 7.5 points, and Gemini 4 Argon lost only 0.5. Those swings are 15 to 30 times the size of the leader's margin. A lab could reshuffle the top of the Index just by choosing a different default effort.

The Vibe Code table makes the same point in dollars. Over Gemini 4 Argon, Fable 5.2 gains 1.86 points and costs $14.02 more per test, about $7.50 per extra point. DeepSeek V4.1 Flash 1017 gets within 6 points for roughly 1/38th of the cost. Vals also found that in 22 of 56 model-benchmark pairs, a cheaper effort setting scored just as well. Together, our read is that "which model is #1" matters less than "which model at which effort clears my quality bar most cheaply". Fable 5.2 is a strong answer to the second question in coding and finance-agent work, not across the board.

Why it matters

Builders should treat Fable 5.2 as a top candidate for agentic coding, finance-agent tasks and EMB-style tasks. At $8/$40 per million tokens it is also cheaper than Fable 5.1 was. Buyers routing tax or legal-research workloads have no reason to switch on the strength of the aggregate rank. Fable sits at #27 and #19 on those benchmarks.

The strongest counter-argument is that two independent evaluators, Vals and Epoch, both put Fable first, and agreement between different methods is real signal. That is fair, and it supports "Fable 5.2 is among the top two or three". It does not support "Fable 5.2 is clearly best". The Epoch lead also depends on rows such as FrontierMath Tier 4 at 100% and SimpleQA Verified at 94%, which are close to saturation. Neither index tells you how Fable 5.2 performs at Medium or Low effort, and that is the setting most production spend will run at.

What happens next

  • A model other than Claude Fable 5.2 will hold the #1 spot on the Vals Index within three months. (3 months, 55%). Watch for: A new frontier release from OpenAI, Google or Anthropic is evaluated by Vals and scores within 1 point of 71.31%. Resolves on 2027-02-10, yes if: The Vals AI Vals Index leaderboard page lists a model other than Claude Fable 5.2 at rank 1 on 2027-02-10.
  • Vals will publish Claude Fable 5.2 results at an effort setting other than High within three months. (3 months, 45%). Watch for: Fable 5.2 appears in the reasoning-effort chart on the Vals Index page. Resolves on 2027-02-10, yes if: The Vals AI reasoning-effort chart or the Claude Fable 5.2 model page on vals.ai shows Claude Fable 5.2 scores at an effort setting other than High by 2027-02-10.
  • Another model will match or beat Fable 5.2's ECI of 174 on Epoch AI within three months. (3 months, 45%). Watch for: Epoch AI adds a new OpenAI or Google DeepMind frontier model to its ECI rankings. Resolves on 2027-02-10, yes if: The Epoch AI Capabilities Index (ECI) leaderboard on the Epoch AI website lists a model other than Claude Fable 5.2 with an ECI score of 174 or higher on 2027-02-10.

Limits and uncertainty

Vals ran Fable 5.2 at a single effort level, so its own cost-versus-score curve is unmeasured. Anthropic's Medium-effort figure comes from the vendor and has not been independently checked. The Vals effort study used about half of each benchmark's tasks and seven models that did not include Fable 5.2. Applying its findings to Fable is an inference. Cost per test also depends on the harness: every Vibe Code run used OpenHands, and results in Claude Code or other agents may differ. Finally, the ±0.96 is the uncertainty on Fable's own score. A proper test of the gap would need both models' errors, which could make the lead look slightly more or less robust.

Competing readings

  • The #1 rank is statistically thin and depends on configuration. Fable 5.2's real advance is cost-efficiency in coding and agentic work, not clear overall superiority (likely, about 60%). What would tell it apart: The 0.46-point Vals lead is smaller than the ±0.96 uncertainty (vals.ai); Effort settings move scores by far more than the gap, and cheaper settings often match (vals.ai)(vals.ai); Ranks are very uneven across components (vals.ai).
  • Fable 5.2 is a genuine broad step up, confirmed because two independent evaluators both rank it first (unlikely, about 30%). What would tell it apart: Epoch ECI 174 against the previous best of 169, from a different methodology (epoch.ai); Vals also ranks it #1 (vals.ai).
  • Gains come from narrow tuning toward coding and finance benchmarks, with real regressions or weaknesses elsewhere (very unlikely, about 10%). What would tell it apart: Weak tax and HLAB results next to first-place coding and finance results (vals.ai); Vendor claims rest on selected effort comparisons (anthropic.com).

Base rate

Of 21 earlier research stories about companies or models in the New Runtime archive over the last 12 months that are at least 3 months old, 19 (90%) were followed within 3 months by a release story on the same company or product line. Of 21 comparable research stories, 19 were followed by a release on the same line within 3 months. That suggests the competitive field around this leaderboard will turn over quickly, and today's 0.46-point lead is unlikely to define the next quarter.

What would change this view

  • Vals publishes Fable 5.2 at Medium and Low effort, and it still leads the Index by more than 1.5 points at matched cost (model-listing signal): The lead is robust to configuration and reflects real capability, not effort choice.
  • Anthropic cuts Fable 5.2 prices, or rivals raise theirs, so that its Vibe Code cost per test falls below Gemini 4 Argon's (pricing signal): The cost-per-point argument against defaulting to Fable would no longer hold.
  • Independent evaluators such as Artificial Analysis' FrontierCode or Hex's DataBench show Fable 5.2 leading by wide margins (signal): Broad superiority would be better supported than a noise-level lead.

Skill to build now

Measure cost per point at several effort levels instead of reading aggregate ranks. First exercise: take 30 representative tasks from your own backlog and run them through Fable 5.2 at Medium and at High effort. Record accuracy, dollars and wall-clock time for each run. Plot the two points next to Gemini 4 Argon's, and work out the marginal dollars per extra correct task.

Reader takeaway

Ignore the #1 badge. Pull 30–50 tasks from your own workload and run Fable 5.2 at Medium and High effort, alongside Gemini 4 Argon and one cheap model such as DeepSeek V4.1 Flash. Pick the cheapest model-and-effort pair that clears your accuracy bar. Revisit the choice when the next frontier release lands, which on recent form will be within weeks.

Open question

With lab defaults and evaluator effort choices now moving scores by more than the gaps between top models, should leaderboards rank models at a fixed cost budget rather than at each model's highest setting?

Evidence / sources

  1. 1vals.ai/models/anthropic_claude-fable-5-2
  2. 2epoch.ai/models/claude-fable-5-2
  3. 3vals.ai/benchmarks/vibe-code
  4. 4vals.ai/blogs/reasoning-effort
All Analysis

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02archiveAnalysisOpen the latest authorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract