Claude Fable 5.2 vs Opus 5.5: What It REALLY Costs
Anthropic says Claude Fable 5.2 on high effort matches Opus 5.5 at its best for about half the cost. So I gave both the same three tasks and counted every dollar. Fable on high finished all three for $0.73. Opus on max cost $2.60 and finished one. The surprise: Opus on its default setting was cheapest of all, $0.16. What you'll learn: → What launched, the price and the effort setting → What my three tests cost on each setting → Where "half the cost" comes from, and what outside testers found → What people built with it on X → The catch: weekly limits and resets → Which model and which effort to pick My tests (9 Nov 2026, Anthropic API, one run each, list prices) Fable 5.2 on HIGH, Opus 5.5 on MAX, and Opus 5.5 on DEFAULT (no effort sent, so medium). • Coding: one Python file that prints the next run times of a cron schedule and rejects bad input, scored by my 41 hidden tests. Fable high 41/41, 110 s, $0.44 (limit 40,000 tokens); Opus max no code, 568 s, $1.28 (64,000); Opus default 41/41, 32 s, $0.08 (24,000). • Maths: "How many subsets of {1, 2, 3, ..., 30} (including the empty set) have a sum that is divisible by 7?" Answer 153,391,696, checked with a short program. All right. Fable high 12 s, $0.05 (12,000); Opus max 15 s, $0.04 (24,000); Opus default 10 s, $0.02 (24,000). • Picture: "Draw a pelican riding a bicycle as a single SVG." Fable high drawn, 69 s, $0.24 (16,000); Opus max no drawing, 648 s, $1.28 (64,000); Opus default drawn, simpler, 30 s, $0.06 (12,000). Opus max returned nothing twice: it was still thinking at my 64,000-token limit (my limit, not a broken model; Opus had more room than Fable). Output tokens, thinking included: Fable high 18,226, Opus max 129,863. Honest trade-offs: single runs on small tasks, one sample, not a score. Anthropic's claim is about long coding tasks. With a bigger limit, Opus max might have finished, and cost more. Fable hits limits too: at its default it used my whole 32,000-token limit on a game and a solar system page, and gave me nothing ($1.28 each). X builds use each poster's own estimates. Also in the launch (not covered in the video) • Fermat's Last Theorem: 1,000 agents on an early version of Fable 5.2 reproduced a computer-checked proof in 28.7 hours. It didn't start from nothing: it got 60 statements, a one-sentence sketch, the definitions and 47 lemmas proved in earlier work. • Electronics: Fable 5.2 designed three devices with Diode; Anthropic says the prototypes worked without fixes to their circuits. No electrical validation, thermal testing or certification yet, and Anthropic has a commercial relationship with Diode. • Safety (system card, Mythos 5.2, the same model): it may have known it was being tested in up to 43% of automated audit transcripts. In sandbox tests built to tempt it, it reached or tried to reach outside in 31% (Opus 5.5: 20%), nearly always just looking around; 1.2% once told the rules. Timestamps 0:00 Fable 5.2 vs Opus: the pelican test 0:31 What you'll learn 0:56 What launched: price and effort 2:15 My tests: High vs Max 5:28 Where "half the cost" comes from 7:09 What people built on X 8:16 The catch: limits and resets 9:31 Verdict: which model, which effort 10:34 Recap Numbers used (checked 9 Nov 2026) • Per million tokens in/out: Fable 5.2 $8/$40, Opus 5.5 $4/$20. https://platform.claude.com/docs/en/about-claude/pricing • FrontierCode 2.0, run by Cognition: Fable high $6.72 a task, Opus best $12.82, Fable max $44.70. https://www.anthropic.com/claude-fable-and-mythos-5-2 • Artificial Analysis: Fable high $9.08, Opus max $10.04, Fable max $26.54 a task. https://artificialanalysis.ai/models/claude-fable-5-2 • Vals AI: Vals Index $17.94 vs $24.77 a test; Terminal-Bench 4.0 $14.56 vs $12.04 (Fable high vs Opus max). https://www.vals.ai/benchmarks/vals_index • Plans: https://support.claude.com/en/articles/17506494 Credits Pelican test: Simon Willison. Shown from X: https://x.com/kevin_t_ngo/status/2119868968307695818 https://x.com/BertrandDiouly/status/2119886197648765001 https://x.com/marcthecreatorr/status/2119890973560054072 https://x.com/DanielZambrini/status/2119893834276507954 https://x.com/notjazii/status/2119886035421442097 https://x.com/mattshumer_/status/2119879860344766782 https://x.com/Yuchenj_UW/status/2119867942603264279 https://x.com/ValsAI/status/2119869713782378658 Not independent: @mattshumer_ had early access and invests in Spawn, which hosts his game; @kevin_t_ngo very likely had early access. YouTube comment: @jonjohnson2844. Tests: Cognition, Artificial Analysis, Vals AI. Anthropic system card: https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card Which effort setting do you use? Tell me in the comments. Follow Mervin Praison for more AI news, explained simply. #ClaudeFable #Claude #Anthropic #ClaudeOpus #AINews
