Claude Fable 5.2 vs Claude Opus 5.5
Should you upgrade from Claude Opus 5.5 to Claude Fable 5.2 at 2x base token pricing?
01 // VERDICT
The short answer
| At a glance | Claude Fable 5.2 | Claude Opus 5.5 |
|---|---|---|
| List priceinput / output per 1M tokens | $8 / $40 | $4 / $20 (better) |
| Benchmark task costArtificial Analysis, per Intelligence-Index task | $26.54 | $10.04 (better) |
| Intelligence IndexAA v5.0.1, higher is better | 65 (better) | 59 |
Highlighted values mark the lower rate or stronger measured score in the same documented run.
EXECUTIVE SUMMARY
Claude Fable 5.2 doubles Claude Opus 5.5's uncached token rates ($8/$40 vs $4/$20 per MTok, Batch $4/$20 vs $2/$10, Fast $16/$80 vs $8/$40), yet keeps cached reads identical at $0.20/MTok (2.5% of Fable input vs 5% on Opus) across the same flat 1M window. In Artificial Analysis Index v5.0.1, Fable 5.2 (Max) lifts Intelligence Index from 59 to 65 (+6 pts), AA-Omniscience from 48 to 67 (+19 pts), HLE-Diamond from 67.9% to 80.2%, and Terminal-Bench-Science from 59.0% to 73.3%, while Anthropic's own table shows Remote Labor Index nearly doubling (41.3% vs 21.3%) and FrontierSWE v2 jumping to 69.9% vs 62.3%. However, Fable Max burns 2.64x more spend per AA task ($26.54 vs $10.04), Opus 5.5 Max still edges AA-Briefcase (1800 vs 1785 Elo) and ties Terminal-Bench 4.0 (59.6%), and Fable enforces breaking API changes: adaptive thinking cannot be turned off, tool_choice 'any'/'tool' returns 400, and 30-day data retention is default.
Choose Claude Fable 5.2
Upgrading knowledge reliability (AA-Omniscience 67 vs 48), hard science/research (HLE-Diamond 80.2% vs 67.9%, Terminal-Bench-Science 73.3% vs 59.0%), FrontierSWE v2 (69.9% vs 62.3%), or high-cache workflows where cached reads stay $0.20/MTok for zero cache-read surcharge.
Choose Claude Opus 5.5
Cost-sensitive enterprise coding, AA-Briefcase office deliverables (1800 vs 1785 Elo at 2.6x lower cost), workloads requiring forced tool_choice ('any'/'tool') or zero-day retention defaults without new Enterprise authorization.
TIE-BREAKER RULE
Either — decide by evidence you generate
Drop-in cost parity trick: run Claude Fable 5.2 at 'medium' effort (AA Index 56, $3.80/task) — it outscores Opus 5.5 'high' (Index 52, $3.46/task) for nearly the exact same per-task spend.
02 // KEY FACTS
Key facts
- Base, cache-write, Batch, and Fast token rates double ($8/$40 vs $4/$20; 5m write $10 vs $5), EXCEPT cached reads stay $0.20/MTok on BOTH models (2.5% on Fable vs 5% on Opus).
- Both offer flat 1,000,000-token pricing (no >272K cliff), 128,000 max output, Batch (-50%), Fast (2x), and reasoning levels low/medium/high/xhigh/max.
- AA v5.0.1 overall gain: Fable 5.2 Max reaches Index 65 ($26.54/task, 41.1 min) vs Opus 5.5 Max Index 59 ($10.04/task, 24.5 min); Fable Medium ($3.80, Index 56) matches Opus High ($3.46, Index 52) spend.
- Largest domain jump is hallucination & knowledge accuracy (AA-Omniscience 67 vs 48, +19 pts) and research/STEM (HLE-Diamond 80.2% vs 67.9%, CritPt v2 76.5% vs 68.0%, Terminal-Bench-Science 73.3% vs 59.0%).
- Anthropic Nov 2026 table adds FrontierSWE v2 (69.9% vs 62.3%), Remote Labor Index (41.3% high vs 21.3%), and OSWorld 2.1 offline partial (89.3% vs 87.2%); Opus holds AA-Briefcase (1800 vs 1785) and ties Terminal-Bench 4.0 (59.6%).
- Breaking API changes on Fable 5.2: adaptive thinking is always enabled (default 'high'), forced tool_choice ('any'/'tool') returns HTTP 400, and 30-day data retention is required unless specifically authorized.
03 // TIER PRICING
Official pricing, every tier
USD per 1M tokens from vendor pricing pages. 'Not published' means exactly that — we don't extrapolate.
| Service tier | Claude Fable 5.2 | Claude Opus 5.5 |
|---|---|---|
| Standarddefault on-demand serving | in $8 · cached $0.20 · out $40 cache write $10 | in $4 · cached $0.20 · out $20 cache write $5 |
| Batchasync jobs, 50% off | in $4 · cached $0.10 · out $20 cache write $5 · derived | in $2 · cached $0.10 · out $10 cache write $2.5 · derived |
| Fast2× on supported OpenAI and Anthropic models | in $16 · cached $0.40 · out $80 cache write $20 · derived | in $8 · cached $0.40 · out $40 cache write $10 · derived |
| Long-context band | Claude Fable 5.2 | Claude Opus 5.5 |
|---|---|---|
| Long-context ratesapplies per request above the model threshold | Flat pricing same rates at any input size (up to context window) | Flat pricing same rates at any input size (up to context window) |
USD per 1M tokens. Every input token is billed in exactly one bucket — uncached, cached read, or cache write; a write replaces the input rate for those tokens (official rule), so write rates are per tier as listed. Anthropic rows bill the 5-minute cache-write TTL ($5/MTok on Opus 5.5, $10/MTok on Fable 5.2); 1-hour writes ($8 and $16) are also published, and cache multipliers stack with Batch and Fast-mode pricing. · Anthropic pricing (verified 2026-11-19)
Tier ≠ quality change
04 // COST SIMULATOR
Run this comparison on your own numbers
Preloaded with a workload typical for this comparison. Change anything — results recompute locally.
Estimated monthly ledger
browser-localToken-based workload estimate only; it does not measure quality or cost per successful task.
| Model | Per request | Monthly |
|---|---|---|
| Claude Fable 5.2 | $0.292 | $5,840.00 |
| Claude Opus 5.5cheapest | $0.148 | $2,960.00 |
Monthly spread between the cheapest and priciest selected model: $2,880.00.
Breakdown: Claude Fable 5.2 · standard tier · standard context
- Fresh input (uncached): 6,000 tok × $8/M$0.048
- Cached input (reads): 20,000 tok × $0.2/M$0.004
- Cache writes: 4,000 tok × $10/M$0.04
- Output (incl. reasoning tokens): 5,000 tok × $40/M$0.20
- Per request$0.292
- 4,000 tokens billed at the cache-write rate instead of the input rate (mutually exclusive buckets — no double charge).
Rates: official pricing · verified 2026-11-18
- Per-token arithmetic only: it does not model quality differences or cross-vendor output-token-count differences for the same task.
- Reasoning tokens are billed as output tokens at the chosen model's output rate.
- Long-context band is decided per request by total input tokens vs the provider threshold, not by monthly volume.
- Cache accounting uses the official mutually-exclusive buckets: each input token is billed once, as uncached input, cached read, OR cache write — writes replace the input rate for those tokens, never add to it.
- Claude rows bill the 5-minute-TTL cache write; the published 1-hour TTL ($8/MTok standard) is not modeled (no TTL input).
- Rows flagged 'derived' are computed from a stated official rule (e.g. Anthropic batch -50%), not read verbatim from a price table.
- Model and tier combinations without a published rate appear as Unsupported.
05 // BENCHMARK LOG
What the evidence actually says
Independent and vendor-run evals, each with publisher, runner, and harness. Numbers from different runs are never merged.
| Model | Score | Effort | Harness | Evidence |
|---|---|---|---|---|
| AutomationBench | ||||
| Claude Opus 5.5claude-opus-5-5 | 40 % | unspecified | Zapier (Opus 5.5 measured during Zapier's early access), published by Anthropic | Third-party |
| Terminal-Bench-Science (0.1)0.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 58.7 % | max (adaptive thinking) | Anthropic announcement evals | Vendor-reported |
| Terminal-Bench-Science cost per task0.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 23.21 usd-per-task | unspecified | OpenAI announcement cost table (who measured the competitor's cost is not stated) | Vendor-reported |
| Terminal-Bench (4.0)4.0⚠ run comparability unknown | ||||
| Claude Opus 5.5claude-opus-5-5 | 66.4 % | xhigh (adaptive thinking) | Anthropic announcement evals | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 60 % | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| FrontierCode (v1.1 (Main))v1.1 (Main)⚠ run comparability unknown | ||||
| Claude Opus 5.5claude-opus-5-5 | 54.4 % | max (adaptive thinking) | Anthropic announcement evals | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 54.6 % | medium (default) | Anthropic announcement evals | Vendor-reported |
| CursorBench (4.0)4.0 | ||||
| Claude Opus 5.5claude-opus-5-5 | 57.8 % | max (adaptive thinking) | Anthropic announcement evals | Vendor-reported |
| GDPval-AAv2.1⚠ run comparability unknown | ||||
| Claude Opus 5.5claude-opus-5-5 | 1846 elo | unspecified | Anthropic announcement evals | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 1866 elo | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| Humanity's Last Exam (with tools) | ||||
| Claude Opus 5.5claude-opus-5-5 | 67.7 % | unspecified | Anthropic announcement evals | Vendor-reported |
| OSWorld (2.1 (partial))2.1 (partial) | ||||
| Claude Opus 5.5claude-opus-5-5 | 81.8 % | unspecified | Anthropic announcement evals | Vendor-reported |
| AA Intelligence Index (v4.3.2)v4.3.2 | ||||
| Claude Opus 5.5claude-opus-5-5 | 58 index-points | config 'Opus 5.5 (Max, Default Fallback)' per AA | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| SciCode | ||||
| Claude Opus 5.5claude-opus-5-5 | 67 % | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| AA-Briefcase (v1.1)v1.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 1807 elo | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| Humanity's Last Exam | ||||
| Claude Opus 5.5claude-opus-5-5 | 61 % | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| AutomationBench-AA | ||||
| Claude Opus 5.5claude-opus-5-5 | 70 % | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| AA cost per Intelligence-Index task (v4.3.2)v4.3.2 | ||||
| Claude Opus 5.5claude-opus-5-5 | 5.98 usd-per-task | — | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models) | Third-party |
| Vals Index (DeepMind table)Oct 2026 | ||||
| Claude Opus 5.5claude-opus-5-5 | 67 % | — | Google DeepMind Model Card evaluation table (competitor column) | Vendor-reported |
| GraphWalks (256K–1M long context)BFS F1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 66.8 % | — | Google DeepMind Model Card evaluation table (competitor column) | Vendor-reported |
| LABBench 2v2 | ||||
| Claude Opus 5.5claude-opus-5-5 | 73.1 % | — | Google DeepMind Model Card evaluation table (competitor column) | Vendor-reported |
| GDP.pdf | ||||
| Claude Opus 5.5claude-opus-5-5 | 26 % | config 'Opus 5.5 (Max, Default Fallback)' per AA | Artificial Analysis Intelligence Index v4.3.2 (independent, same harness) | Third-party |
| AA Time to First Answer Tokenv4.3.2 | ||||
| Claude Opus 5.5claude-opus-5-5 | 748.61 seconds | config 'Opus 5.5 (Max, Default Fallback)' per AA | Artificial Analysis Intelligence Index v4.3.2 latency test | Third-party |
| DeepSWEv1.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 74.2 % | max | Meta Muse Wonder announcement evaluation table (competitor column) | Vendor-reported |
| SWE-Marathon | ||||
| Claude Opus 5.5claude-opus-5-5 | 46.3 % | max | Meta Muse Wonder announcement evaluation table (competitor column) | Vendor-reported |
| MemoryBench | ||||
| Claude Opus 5.5claude-opus-5-5 | 81.5 % | max | Meta Muse Wonder announcement evaluation table (competitor column) | Vendor-reported |
| Harvey LAB-AA | ||||
| Claude Opus 5.5claude-opus-5-5 | 4.2 % | max | Meta Muse Wonder announcement evaluation table (competitor column) | Vendor-reported |
| AA Intelligence Index (v5.0.1)v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 65 index-points | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 59 index-points | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| AA cost per Intelligence-Index task (v5.0.1)v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 26.54 usd-per-task | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 10.04 usd-per-task | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| AutomationBench-AA (AA v5.0.1)AA v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 76.4 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 69.5 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Terminal-Bench (4.0 (AA v5.0.1))4.0 (AA v5.0.1) | ||||
| Claude Fable 5.2claude-fable-5-2 | 59.6 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 59.6 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| FrontierCode-AA2.0 (AA v5.0.1) | ||||
| Claude Fable 5.2claude-fable-5-2 | 48.8 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 46.4 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| AA-Briefcase (v1.1 (AA v5.0.1))v1.1 (AA v5.0.1) | ||||
| Claude Fable 5.2claude-fable-5-2 | 1785 elo | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 1800 elo | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| AA-Omnisciencev1.1 (AA v5.0.1) | ||||
| Claude Fable 5.2claude-fable-5-2 | 67 index-points | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 48 index-points | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| HLE-DiamondAA v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 80.2 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 67.9 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| CritPt v2 (Hard Subset)AA v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 76.5 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 68 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Terminal-Bench-Science (0.1 (AA v5.0.1))0.1 (AA v5.0.1) | ||||
| Claude Fable 5.2claude-fable-5-2 | 73.3 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 59 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| GDP.pdf (AA v5.0.1)AA v5.0.1 | ||||
| Claude Fable 5.2claude-fable-5-2 | 24.2 % | config 'Claude Fable 5.2 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| Claude Opus 5.5claude-opus-5-5 | 26.2 % | config 'Claude Opus 5.5 (max)' per AA | Artificial Analysis Intelligence Index v5.0.1 (independent, same harness across all models) | Third-party |
| FrontierCode (v2.0 (Anthropic Nov 2026))v2.0 (Anthropic Nov 2026) | ||||
| Claude Fable 5.2claude-fable-5-2 | 49.1 % | xhigh | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 45.5 % | max | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| FrontierSWEv2 (Anthropic Nov 2026) | ||||
| Claude Fable 5.2claude-fable-5-2 | 69.9 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 62.3 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| OSWorld (2.1 (offline set partial, Anthropic Nov 2026))2.1 (offline set partial, Anthropic Nov 2026) | ||||
| Claude Fable 5.2claude-fable-5-2 | 89.3 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 87.2 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Remote Labor Index (RLI)Anthropic Nov 2026 | ||||
| Claude Fable 5.2claude-fable-5-2 | 41.3 % | high | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 21.3 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Humanity's Last Exam (with tools) (Anthropic Nov 2026)Anthropic Nov 2026 | ||||
| Claude Fable 5.2claude-fable-5-2 | 74.2 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| Claude Opus 5.5claude-opus-5-5 | 67.7 % | unspecified | Anthropic Fable 5.2 & Mythos 5.2 announcement benchmark table (2026-11-09) | Vendor-reported |
| CursorBench (4.1)4.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 58.1 % | — | Cursor supplied the score to Google DeepMind for its November 18 model card. | Third-party |
| Terminal-Bench (4.1)4.1 | ||||
| Claude Opus 5.5claude-opus-5-5 | 65.8 % | — | DeepMind took Opus’s cheating-adjusted accuracy from the public Terminal-Bench leaderboard. | Measured |
Same benchmark name, multiple rows — run comparability unknown
Terminal-Bench (4.0): claude-opus-5-5 = 66.4 (Anthropic announcement evals); claude-opus-5-5 = 60 (Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models)). Whether these rows share a run is not documented in the sources; treat them as not comparable.
FrontierCode (v1.1 (Main)): claude-opus-5-5 = 54.4 (Anthropic announcement evals); claude-opus-5-5 = 54.6 (Anthropic announcement evals). Whether these rows share a run is not documented in the sources; treat them as not comparable.
GDPval-AA: claude-opus-5-5 = 1846 (Anthropic announcement evals); claude-opus-5-5 = 1866 (Artificial Analysis Intelligence Index v4.3.2 (independent, same harness for all models)). Whether these rows share a run is not documented in the sources; treat them as not comparable.
We never average or merge these. See methodology.Sources: Anthropic — Introducing Claude Opus 5.5 · OpenAI — Introducing GPT-6.1 Sol (DevDay 2026) · Artificial Analysis (independent) — Independent same-harness comparison: GPT-6.1 Sol vs Claude Opus 5.5 (Intelligence Index v4.3.2) · Google DeepMind — Gemini 4 Argon Model Card (October 2026 evaluation table) · Artificial Analysis (independent) — Independent same-harness comparison: Muse Wonder (Max) vs Claude Opus 5.5 (Intelligence Index v4.3.2) · Meta — Put your ideas to work with Muse Wonder and Muse Wonder Full (official release blog & benchmark table) · Artificial Analysis (independent) — Independent same-harness comparison: Intelligence Index v5.0.1 (GPT-6.2 Astra, GPT-6.2 Sol, Claude Fable 5.2, Claude Opus 5.5) · Anthropic — Introducing Claude Fable 5.2 and Claude Mythos 5.2 (announcement & benchmark table) · Google DeepMind — Gemini 4 Neon model card
06 // BY WORKLOAD
Which model for which workload
Confidence labels: Measured = same-harness evidence · Price-derived = list-price arithmetic · Inferred = reasoned from docs · Insufficient evidence = no comparable public data.
| Workload | Our pick | Confidence | Why |
|---|---|---|---|
| Quality-first | Claude Fable 5.2 | Measured | Fable 5.2 delivers +6 AA Index points (65 vs 59), +19 pts on AA-Omniscience (67 vs 48), +12.3pp on HLE-Diamond, and nearly doubles RLI (41.3% vs 21.3%). |
| Complex repository maintenance | Claude Fable 5.2 | Measured | FrontierSWE v2 improves +7.6pp (69.9% vs 62.3% Anthropic), FrontierCode 2.0 rises to 49.1% vs 45.5%, and Terminal-Bench-Science jumps +14.3pp in AA v5.0.1 (73.3% vs 59.0%). |
| Daily coding tasks | Claude Opus 5.5 | Measured | Opus 5.5 ties Fable Max on Terminal-Bench 4.0 (59.6% vs 59.6%) at $10.04 vs $26.54/task; keep Opus for routine engineering unless Fable's medium tier ($3.80/task) fits your budget. |
| Agent workflows | Claude Fable 5.2 | Measured | AutomationBench-AA jumps 69.5% -> 76.4%, OSWorld 2.1 hits 89.3%, and Omniscience rises 48 -> 67 — BUT audit tool-calling code first because tool_choice 'any'/'tool' returns 400 on Fable 5.2. |
| Budget-sensitive / high volume | Claude Opus 5.5 | Measured | Opus 5.5 is half Fable's token rates ($4/$20) and 2.64x cheaper per Max AA task ($10.04 vs $26.54). On Opus 5.5 workflows heavy on office deliverables (AA-Briefcase 1800 vs 1785), upgrading wastes 2.6x budget. |
| Long-context tasks | Claude Fable 5.2 | Price-derived | With a flat 1M window on both, cache hits are the sleeper metric: Fable cached reads stay $0.20/MTok (identical to Opus!). Workloads >85% cache-hit can upgrade to Fable with far less than a 2x input-bill increase. |
| Latency-sensitive production | Claude Opus 5.5 | Measured | Opus Max finishes AA tasks in 24.5 min (118 tok/s) vs Fable Max's 41.1 min (111 tok/s), and Opus High takes 9.3 min vs Fable High 14.9 min. Both offer 2x Fast mode. |
| Offline batch jobs | Claude Opus 5.5 | Price-derived | Opus Batch ($2/$10, $0.10 cached, $2.50 5m write) is half Fable Batch ($4/$20, $0.10 cached, $5.00 5m write) unless task complexity specifically requires Fable's research/knowledge gains. |
HOLD CRITERIA
When staying on the lower-cost or incumbent model is reasonable
- Your agent harness forces tool_choice 'any' or a specific 'tool' — Fable 5.2 returns HTTP 400 when forced tool choice is paired with mandatory adaptive thinking.
- Zero-day data retention is non-negotiable without special authorization — Fable 5.2 defaults to 30-day retention.
- Your traffic is document/briefcase deliverables (Opus 1800 vs Fable 1785 Elo) or standard shell automation (both 59.6% Terminal-Bench 4.0) where doubling token cost adds no gain.
EVALUATION PROTOCOL
5-step Opus 5.5 -> Fable 5.2 upgrade verification
Public benchmarks don't run your tasks. Before committing a fleet, measure both models on your own work.
- 1Scan API logs and codebases for tool_choice: {type: 'any' | 'tool'} and thinking: {type: 'disabled'} — both must be refactored before sending traffic to claude-fable-5-2.
- 2Confirm 30-day data retention compliance with InfoSec / legal (Opus 5.5 remains available until at least 2027-09-22).
- 3Benchmark Fable 5.2 at 'medium' effort ($3.80 AA task cost) and 'high' ($9.08) against your current Opus 5.5 effort before turning on 'max' ($26.54).
- 4Compare monthly bill impact in the Cost Calculator using your actual cache-read ratio, since $0.20 cached reads dramatically lower upgrade cost for high-reuse prompts.
- 5Route selectively: promote hallucination-critical research or FrontierSWE failures to Fable 5.2 while leaving routine Opus traffic untouched.
QUESTIONS
Frequently asked questions
Why are Claude Fable 5.2's cached reads still $0.20/MTok like Opus 5.5?
Anthropic priced Fable 5.2 cache hits at 0.025x (2.5%) of base $8.00 input = $0.20/MTok, matching Opus 5.5's 0.05x (5%) of $4.00 = $0.20/MTok. Cache writes (5m) still double from $5 to $10/MTok.
What code breaks if I swap claude-opus-5-5 for claude-fable-5-2 in place?
Per Anthropic's What's New docs for Fable 5.2: adaptive thinking is always enabled, forced tool_choice ('any' or named 'tool') returns HTTP 400 (only 'auto' or 'none' work), and 30-day data retention is mandatory unless Enterprise authorization is configured.
Is Claude Mythos 5.2 the same API model as Fable 5.2?
Anthropic states Fable 5.2 and Mythos 5.2 share underlying capabilities, while Mythos 5.2 is verification-gated for Cybersecurity and Life Sciences programs and Fable 5.2 is generally available under model ID claude-fable-5-2.
CONTINUE EXPLORING
Related
VERIFICATION LOG
Sources & freshness
- Anthropicaccessed 2026-11-18Claude Fable 5.2 Overview (model ID, 1M context, limits, retirement floor)https://platform.claude.com/docs/en/models/fable-5-2/overview · published 2026-11-09claude-fable-5-2: 1,000,000 context window, 128,000 max output, knowledge cutoff Jun 2026, minimum retirement 2027-11-09, flat $8/$40 MTok pricing, adaptive thinking low/medium/high/xhigh/max (default 'high').
- Anthropicaccessed 2026-10-09Claude models overview (model IDs, context window, max output, retirement)https://platform.claude.com/docs/en/docs/about-claude/models/overviewclaude-opus-5-5: 1M context, 128K max output, training cutoff Jun 2026, retirement not sooner than 2027-09-22, cache reads '5% of input price', batch 50% off.
- Anthropicaccessed 2026-11-18What's new in Claude Fable 5.2 (adaptive thinking, breaking changes, data retention)https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2 · published 2026-11-09Documents breaking API changes relative to Opus 5.5 / Fable 5.1: adaptive thinking is always enabled, forced tool_choice ('any' or 'tool') returns 400 error, and 30-day data retention applies by default unless enterprise authorization is granted.
- Anthropicaccessed 2026-10-09Introducing Claude Opus 5.5https://www.anthropic.com/claude-opus-5-5 · published 2026-09-22Self-reported evals (Anthropic harness, adaptive thinking, max effort unless noted). Includes competitor scores incl. GPT-6 Astra as measured by Anthropic.
- Anthropicaccessed 2026-11-19Claude pricing docs (cache TTL prices, fast/batch stacking rules)https://platform.claude.com/docs/en/about-claude/pricingOpus 5.5 1h cache write $8/MTok; caching multipliers apply on top of fast-mode pricing and stack with the Batch discount; fast mode not available with Batch. Re-verified 2026-11-18: Fable 5.2 Batch 50% off ($4/$20), Fast 2x ($16/$80), 2.5% cache-read stacking, 1h cache write $16.00.
- Anthropicaccessed 2026-11-19Claude pricing (API pricing table incl. cache write/read, fast mode, batch)https://claude.com/pricingOpus 5.5: $4/$20 per MTok, cache read $0.20, cache write (5m TTL) $5, batch 'save 50%', fast mode 2x on input/output. Re-verified 2026-11-18: Claude Fable 5.2 ($8/$40, $0.20 cached read = 2.5%, $10.00 5m write).
- OpenAIaccessed 2026-10-09Introducing GPT-6.1 Sol (DevDay 2026)https://openai.com/index/introducing-gpt-6-1-sol · published 2026-09-29Self-reported evals (OpenAI harness). Includes per-task cost quotes (Terminal-Bench-Science). Announces GPT-6.1 Sol Ultrafast 'coming soon' in Codex. Availability: ChatGPT Work + Codex, not yet in Chat.
- Artificial Analysis (independent)accessed 2026-10-10Independent same-harness comparison: GPT-6.1 Sol vs Claude Opus 5.5 (Intelligence Index v4.3.2)https://artificialanalysis.ai/models/comparisons/gpt-6-1-sol-vs-claude-opus-5-5Independent same-harness evaluation of both models. Cost-per-task = weighted average per Intelligence Index task; reflects measured token verbosity. Effort configs as labeled by AA ('Sol (Max)' vs 'Opus 5.5 (Max, Default Fallback)').
- Google DeepMindaccessed 2026-10-13Gemini 4 Argon Model Card (October 2026 evaluation table)https://deepmind.google/models/model-cards/gemini-4-argon/ · published 2026-10-13Official evaluation table comparing Gemini 4 Argon, GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across Vals Index (68.9% vs 67.0% Opus), DeepSWE v1.1 (77.9% vs 74.2% Opus), Terminal-Bench 4.0 (57.4% vs 66.4% Opus), GraphWalks 256k-1M (84.2% vs 66.8% Opus), LABBench 2 (88.8% vs 73.1% Opus), and RiemannBench (76.0% vs 69.6% Opus). Knowledge cutoff March 2026.
- Artificial Analysis (independent)accessed 2026-10-13Independent same-harness comparison: Muse Wonder (Max) vs Claude Opus 5.5 (Intelligence Index v4.3.2)https://artificialanalysis.ai/models/comparisons/muse-wonder-vs-claude-opus-5-5 · published 2026-10-12Independent same-harness Intelligence Index v4.3.2 evaluation of Muse Wonder (Max): Index 52 (vs 58 Opus 5.5, 53 Gemini 4 Argon, 52 GPT-6.1 Sol; subscription-only Muse Wonder Full scores 56), cost per Intelligence-Index task $4.52 (54k output tokens/task, 134M output tokens to run Index, $5,733 total) vs $5.98 Opus 5.5 (119k output tokens/task, 260M total), output speed 85 tok/s, time to first answer token 60.82s vs 748.61s Opus 5.5.
- Metaaccessed 2026-10-13Put your ideas to work with Muse Wonder and Muse Wonder Full (official release blog & benchmark table)https://dev.meta.ai/resources/blog/build-with-muse-wonder · published 2026-10-12Announces Muse Wonder (muse-wonder-1.0 on Meta Model API and Muse Code) and Muse Wonder Full (exclusively in Muse Code Power/Maximum Usage subscriptions, no pay-as-you-go API). Official evaluation table compares Muse Wonder Full (max), Muse Wonder (max), Claude Fable 5.1 (max), GPT-6 Astra (max), and Claude Opus 5.5 (max) across DeepSWE 1.1 (78.8% Wonder vs 74.2% Opus), SWE-Marathon (53.1% vs 46.3%), MemoryBench (85.6% vs 81.5%), Harvey LAB-AA (10.0% vs 4.2%), Terminal-Bench 4.0-AA (52.0% vs 59.6%), OSWorld V2.1 partial (72.6% vs 81.8%), and GDP.pdf (30.8% vs 30.6%).
- Artificial Analysis (independent)accessed 2026-11-18Independent same-harness comparison: Intelligence Index v5.0.1 (GPT-6.2 Astra, GPT-6.2 Sol, Claude Fable 5.2, Claude Opus 5.5)https://artificialanalysis.ai/models/comparisons/gpt-6-2-astra-vs-claude-fable-5-2 · published 2026-11-18Independent same-harness Artificial Analysis Intelligence Index v5.0.1 suite (also mirrored cleanly across /models/comparisons/gpt-6-2-astra-vs-gpt-6-2-sol, /claude-fable-5-2-vs-claude-opus-5-5, and /gpt-6-2-astra-vs-claude-opus-5-5). Incorporates 9 evaluations: AA-Briefcase v1.1, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience v1.1, GDP.pdf, HLE-Diamond, CritPt v2 (Hard Subset), and Terminal-Bench-Science 0.1.
- Anthropicaccessed 2026-11-18Introducing Claude Fable 5.2 and Claude Mythos 5.2 (announcement & benchmark table)https://www.anthropic.com/claude-fable-and-mythos-5-2 · published 2026-11-09Nov 9, 2026 launch post and official benchmark table comparing Fable 5.2, Opus 5.5, and GPT-6 Astra across FrontierCode 2.0 (49.1% vs 45.5% vs 42.8%), FrontierSWE v2 (69.9% vs 62.3% vs 65.5%), OSWorld 2.1 offline partial (89.3% vs 87.2% vs 79.5%), RLI (41.3% high effort vs 21.3% vs 20.8%), Humanity's Last Exam with tools (74.2% vs 67.7% vs 57.2%), Chartography (73.4% vs 64.4% vs 71.0%), and AA-Briefcase v1.1 (1785 vs 1800 vs 1569 Elo).
- Google DeepMindaccessed 2026-11-19Gemini 4 Neon model cardhttps://deepmind.google/models/model-cards/gemini-4-neon/ · published 2026-11-18DeepMind identifies Carbon as the basis for Neon and dates the knowledge cutoff to March 2026. The evaluation table includes CursorBench 4.1 and cheating-adjusted Terminal-Bench 4.1.
Pricing data last verified 2026-11-19. Vendors ship frequently — re-verify before making spend commitments. See the methodology page for update cadence and stale-data flags.