Fable 5.2 vs Opus 5.5: Same Accuracy, 2–5× Cost in Coding Tasks

Overview

On November 9, Anthropic released Claude Fable 5.2, integrated into Claude Code 2.1.321. The author conducted hands-on tests comparing Fable 5.2 and Opus 5.5 at medium and high effort levels across three typical coding tasks in Claude Code's non-interactive mode ( claude -p).

Models and Pricing

Fable is Anthropic's top public tier, with added security guards for cyber and bio; Mythos 5.2 shares the same base model but requires institutional access. Fable 5.2 pricing dropped 20% vs Fable 5.1 but remains double Opus 5.5's price. A new fast mode offers up to 2.5× output speed at double the per-token cost. Official benchmarks show Fable 5.2 outperforming Opus 5.5 on most metrics, and long-horizon demonstrations include 1,000 agents formalizing Fermat's Last Theorem in 30 hours (vs Opus 5.5 completing 26% in 18 hours) and an OS prototype built over eight days.

Test Methodology

Three Python micro-projects were used:

Bug fix : A meeting-room booking library with three failing tests; prompt asked to fix them and check for similar issues. Two additional hidden bugs existed.

Spec-driven parser : Implement a parser for a 9-rule config format given only three examples; hidden suite contained 30 edge cases (escaping, line continuation, comments, section conflicts).

Scheduling : 130 meetings, 5 rooms, 20 people with priorities and constraints; maximize score within 60-second runtime. No optimal solution known; a naive greedy baseline reaches ~70-80% of max.

Each configuration ran in a clean project copy via claude -p without user CLAUDE.md, memory, hooks, skills, or MCP. Costs are Claude Code's official price estimates for relative comparison. The author also ran the first two tasks in their normal setup and observed the same correctness/cost relationship.

Results: Bug Fix and Parser — Perfect Scores, Divergent Costs

All four configurations (Fable 5.2 medium/high, Opus 5.5 medium/high) passed all tests on both tasks, including the hidden bugs. On the bug fix, Fable 5.2 medium finished in ~30 seconds; Opus 5.5 medium took ~1 minute. Estimated cost for Fable was ~1.5× Opus at the same effort level. On the parser, Opus 5.5 completed in 1–2 minutes; Fable 5.2 needed 3–3.5 minutes, produced 2–3× more output tokens, and cost 4–5.5× more. Combined, Fable 5.2 showed no correctness advantage over Opus 5.5.

Results: Scheduling — Same Wall, More Effort

All configurations adopted a greedy initial solution followed by simulated annealing within the 60-second limit. At medium effort, scores differed by 1–8 points (out of 852) across runs; Fable 5.2 took 43 minutes vs Opus 5.5's 27 minutes. Both models reported diminishing returns: Fable 5.2 medium noted extending runtime to 300 seconds yielded ~1 extra point; Opus 5.5 medium estimated a few points from optimum but unverifiable, with 1–3 point variance across random seeds.

At high effort, both switched to multi-process search across multiple random seeds. Opus 5.5 high ran 71 minutes, generated new test data, compared two approaches, and concluded differences were noise. Fable 5.2 high explored tabu search, constraint weighting, time-slice swapping, discarded ineffective ideas, and compared two multi-process collaboration schemes. The author stopped it after 2 hours 24 minutes while it was still running a final evaluation.

Across four rigorous re-runs (alternating runs on a quiet machine), Fable 5.2 high scored 4–7 points higher than Opus 5.5 high (average 1–2 points per dataset). This margin falls within the 2–5 point run-to-run variance observed for a single program, and each configuration was only run once per model, so the difference is not statistically significant. The author characterizes it as both models hitting the same wall, with Fable 5.2 willing to spend more time searching for cracks.

Token Output Speed: Fable 5.2 Catches Opus 5.5

Measured by having each model write the same short story three times at low effort, Fable 5.2 averaged 63 tokens/second, Fable 5.1 averaged 39 tokens/second, and Opus 5.5 averaged 66 tokens/second (end-to-end latency including time to first token). The observed 60%+ speedup over Fable 5.1 exceeds the official "up to 30%" claim; the author attributes the discrepancy to uncontrolled variables. The key takeaway: Fable 5.2's raw generation speed now matches Opus 5.5; its longer wall-clock times on coding tasks stem from more extensive reasoning and verification (e.g., parser task output tokens were 2–3× Opus).

When to Use Fable 5.2

For tasks completable in minutes to an hour — the scope of these three tests — Opus 5.5 delivers equal correctness at far lower cost and time; the author defaults to Opus 5.5. Fable 5.2's demonstrated strengths lie in very long-horizon, massively parallel workloads (tens of hours, thousands of agents) which were not tested here. The author suggests switching to Fable 5.2 only when a problem has stalled even at high Opus effort. Pro subscribers can trial Fable 5.2 within their weekly allowance during the launch promotion (until Nov 16 PT). This aligns with the author's prior five-task evaluation showing that across Opus, Sonnet, and Fable 5.1, low effort often sufficed and higher effort mainly added verification and cost.

Reference: previous article Claude Code Model and Effort Pairing Tested with Five Tasks .