Claude Fable 5.2 Scores 65 on Artificial Analysis Intelligence Index – What the Numbers Mean

Bottom Line

Claude Fable 5.2 tops the Artificial Analysis Intelligence Index with a score of 65, but its cost per task more than doubles compared to Fable 5.1, and output tokens per task jump from ~118 k to ~274 k, making efficiency a major concern for most users.

Intelligence Index Score

  • Score: 65 on the Artificial Analysis Intelligence Index (AA‑II) v5.0.1, the highest reported for any model so far.
  • Why it matters: The AA‑II aggregates nine heavyweight benchmarks (AA‑Briefcase, AutomationBench‑AA, Terminal‑Bench 4.0, FrontierCode 2.0, AA‑Omniscience, GDP.pdf, HLE‑Diamond, CritPt v2, Terminal‑Bench‑Science 0.1). A higher score signals better overall reasoning, knowledge reliability, and agentic capability across a broad spectrum of tasks.

"Fable 5.2 at high is the real thing here – scoring 62, second highest ever, while being 3x less expensive." – nedsma (HN comment)

Cost per Task – The Upside‑Down Trade‑off

  • Cost increase: Fable 5.2’s weighted‑average cost per Intelligence Index task is roughly 2× that of Fable 5.1, even after the announced price cut.
  • Token explosion: Output tokens per task rise from 118 k (≈46 k answer + 71 k reasoning) in 5.1 to 274 k (≈81 k answer + 192 k reasoning) in 5.2.
  • Implication: Users pay more while consuming far more tokens, which directly inflates API bills.

"Forget about the scores… Fable 5.1 was already the most expensive by a good margin, and now Fable 5.2 doubles it!" – rgbrenner

Speed vs. Intelligence – Pareto Position

  • Pareto frontier: On the Artificial Analysis chart that plots Intelligence Index against Weighted‑average cost per task, Fable 5.2 (especially low and medium settings) sits on the frontier, meaning it offers the best trade‑off among models that achieve similar intelligence levels.
  • Time per task: Despite higher token counts, the model’s output speed (tokens / second) remains competitive, keeping total minutes per task comparable to prior releases.
  • User perspective: For workloads where raw intelligence outweighs cost—e.g., high‑stakes planning or complex scientific reasoning—the model may be justified.

"Fable 5.2 (especially low and medium) also sit on the pareto frontier between time per task and intelligence… the new models seem to be meaningfully better here." – oliwary

Capability Highlights

Capability Metric (higher = better) Fable 5.2 Observation
AA‑Briefcase Elo Composite of rubric pass, analytical quality, presentation Leads the leaderboard; combines head‑to‑head Elo scoring with 95 % confidence intervals clamped at 0.
AA‑Omniscience Net Score -100 → 100 (knowledge reliability, hallucination penalty) Near‑top tier; rewards correct answers, penalizes hallucinations, no penalty for refusals.
GDP.pdf reasoning All‑pass on professional document reasoning Consistently passes the hardest legal‑document benchmark.
Agentic workflows (terminal, software engineering) Elo on task completion speed & correctness Shows strong performance on agentic scientific and software tasks, though anecdotal reports note occasional drift.

Community Sentiment – Pros and Cons

Positive takeaways

  • Intelligence lead: Several commenters note the model’s superior scores, especially on high‑effort settings.
  • Planner role: Users who need a high‑quality planner or reviewer find the improvements worthwhile despite higher cost.

"To me Fable's role was always as a planner or plan reviewer… these improvements seem worthwhile in exchange for the cost." – djacksondev

Criticisms

  • Cost explosion: The most repeated complaint is the dramatic rise in cost per task and token usage.
  • Real‑world relevance: Some users report modest practical gains, with hallucinations and “shallow logic” still present.
  • Benchmark trustworthiness: A few commenters question the transparency of Artificial Analysis’s methodology and its paid relationship with AI labs.

"Artificial Analysis is paid by AI labs and they don't publish the full method. To me it's meaningless…" – rodorgas

When Might Fable 5.2 Be Worth It?

  1. High‑stakes planning or legal analysis where the AA‑Omniscience net score’s low hallucination rate is critical.
  2. Agentic research pipelines that benefit from strong terminal‑bench performance, even if token costs are high.
  3. Benchmark‑driven competitions where the headline score can influence funding or partnership decisions.

Conversely, for routine customer‑facing chat, code generation, or content creation, cheaper alternatives (e.g., Opus 5.5, DeepSeek‑Flash) often deliver comparable quality at a fraction of the cost.

Bottom Line Revisited

Claude Fable 5.2 pushes the frontier of measured intelligence, but the steep rise in cost and token consumption makes it a niche tool rather than a universal replacement. Teams should weigh the marginal intelligence gain against the doubled expense, especially when cheaper models already meet most production needs.

Sources

Related