Claude Fable 5.2 beats Claude Opus 5.5 on 11 of the 12 benchmarks in Anthropic's own table, and it lists at exactly twice the price. Anthropic announced Fable 5.2 and Mythos 5.2 on November 9, 2026, a point release that reverses the September story where Opus 5.5 beat Fable 5.1 on every published benchmark. This post answers the question builders ask next: do I switch, and for which jobs?
The short version is that "Fable 5.2 vs Opus 5.5" is not a winner-takes-all contest. Fable leads by wide margins on long-horizon, search and robotics work. Opus 5.5 still ties it on coding and it wins one knowledge-work score. Every number below is lab-reported unless we say otherwise.
TL;DR: the questions people ask first
| Question | Direct answer |
|---|---|
| Which is stronger? | Fable 5.2 on 11 of 12 rows in Anthropic's table; Opus 5.5 wins AA-Briefcase Elo and edges FrontierCode at High effort |
| List price? | Fable 5.2 $8 / $40 per million tokens; Opus 5.5 $4 / $20 |
| Cheaper per task? | Depends on effort. Anthropic reports Fable matching Opus at about half the cost per task on FrontierCode and about 40% on WANDR |
| Third-party score? | Artificial Analysis lists Fable 5.2 at 65 on its Intelligence Index; Opus 5.5 led v5 at 58 on October 27 |
| Context and output? | Fable 5.2 has 1M context and 128K max output |
| Code changes needed? | Prefix checks on all accounts; computer use needs computer_toolset_20260801 |
| My default? | Keep Opus 5.5 as the workhorse. Route hard, long, ambiguous work to Fable 5.2 |
What did Anthropic publish: Fable 5.2 vs Opus 5.5 head to head?
Anthropic's launch table compares Fable 5.2 against Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. Here are the Fable 5.2 and Opus 5.5 columns, plus Fable 5.1 so you can see how much moved in one point release.
| Benchmark | Fable 5.2 | Opus 5.5 | Fable 5.1 |
|---|---|---|---|
| FrontierCode 2.0 (agentic coding, Xhigh for Fable) | 49.1% | 45.5% | 32.4% |
| FrontierSWE v2 (agentic coding) | 69.9% | 62.3% | 56.3% |
| AutomationBench (business workflows) | 51.4% | 42.5% | 31.4% |
| Terminal-Bench-Science 0.1 | 74.6% | 58.7% | 52.6% |
| OSWorld 2.1 (computer use) | 89.3% | 87.2% | 83.5% |
| WANDR (agentic search, relaxed) | 82.1% | 72.3% | 68.7% |
| RoboDojo Sim (robotics) | 64.9% | 39.4% | 27.6% |
| AA-Briefcase v1.1 (rubric) | 66.1% | 61.2% | 61.5% |
| AA-Briefcase v1.1 (Elo) | 1785 | 1800 | 1669 |
| Remote Labor Index (High effort) | 41.3% | 21.3% | 17.9% |
| Humanity's Last Exam (with tools) | 74.2% | 67.7% | 65.6% |
| Chartography (no tools) | 73.4% | 64.4% | 44.8% |
The biggest gaps are not in coding. On Remote Labor Index, Fable 5.2 roughly doubles Opus 5.5 (41.3% vs 21.3%). On RoboDojo it leads by about 25 points. On Terminal-Bench-Science it leads by about 16. Coding gaps are smaller: 3.6 points on FrontierCode 2.0 and 7.6 on FrontierSWE v2.
A fine pen beside a heavy hammer, showing Fable 5.2 vs Opus 5.5 as a precision tool and a heavy tool
The same table says something quieter. Fable 5.1 trailed Opus 5.5 on every row where both were published, so the 5.2 jump is the real news: on FrontierCode 2.0 it climbed from 32.4% to 49.1%, and on Remote Labor Index from 17.9% to 41.3%. That is why the old "what is the point of Fable" argument from our September comparison no longer holds.
Does Fable 5.2 cost more per finished task than Opus 5.5?
On list price, yes, exactly 2x. A job with 1M input tokens and 200K output tokens costs $8 on Opus 5.5 and $16 on Fable 5.2 at standard rates ($4 plus $4, against $8 plus $8). Per-token price is only half the bill, though. The other half is how many tokens each model spends.
Anthropic's launch post gives three matched-quality data points:
- FrontierCode at High effort: Fable 5.2 scores 45.3% against Opus 5.5's 45.5%, at roughly half the cost per task.
- WANDR at Medium effort: Fable 5.2 matches Opus 5.5's top score at about 40% of the cost per task.
- Humanity's Last Exam at Low effort: Fable 5.2 matches Opus 5.5's top score at under one-tenth of the cost.
Read that carefully. Anthropic is not saying Fable is cheap. It is saying Fable at a low effort setting can match Opus at its best setting for less money, because Opus needs more reasoning tokens to land in the same place. That fits what developers said in September, when one commenter described Fable as the model that "just knows stuff" while Opus "needs to reason more to end in the same place."
The counter-evidence comes from outside Anthropic. Playcode's founder, in the test we covered in AI labs raise prices without raising prices, found Fable 5.2 more expensive on one drawing task even after the 20% price cut: at xhigh and max effort it reportedly spent 128,000 output tokens thinking and returned no image, about $5.12 per attempt at list rates. Their figure for Fable 5.2 at max on another run was 250,082 output tokens, against 99,072 for Fable 5.1.
Both things can be true. Fable 5.2 is cheaper on the tasks Anthropic picked, at the effort levels Anthropic picked. It can be pricier on open-ended generation at max effort. The only number that settles it is cost per successful artifact on your own traces.
Where does Opus 5.5 still win?
Opus 5.5 keeps one row and one near-tie in Anthropic's own table:
- AA-Briefcase Elo: 1800 for Opus 5.5 against 1785 for Fable 5.2. The rubric version of the same benchmark goes the other way (66.1% vs 61.2%), which tells you the two scoring methods disagree by more than the gap itself.
- FrontierCode 2.0 at High effort: 45.5% against 45.3%, a statistical tie.
Opus 5.5 also holds a price advantage that does not show up in a benchmark. At $4/$20 it is the cheaper default for any workload where the extra Fable margin is small. Our Opus 5.5 launch coverage has the original reaction, and the Opus 5.5 vs Sonnet 5.5 comparison shows where the cheaper rung of the ladder fits.
There is a further data point on the independent side. The SemiAnalysis GDPx benchmark found Opus 5.5 leading at 12.0% with agents failing most tasks, a reminder that the best model on a hard benchmark still fails often. See our GDPx write-up.
What does the independent scoreboard say?
Artificial Analysis lists Fable 5.2 at 65 on its Intelligence Index as of November 9, per our Fable 5.2 launch coverage. For context, the rebuilt v5 index had Opus 5.5 (max) on top at 58 on October 27, per our v5 write-up. Two caveats apply before you quote that seven-point gap.
First, we have not re-run either leaderboard, and index composition and effort settings change between versions. Second, the v5 index weights about 60% of its score on private or held-out data, so a lab's own table and the index can legitimately rank models differently.
A scoreboard being redrawn over eraser marks, a reminder that benchmark margins shift
Anthropic itself hedged this way in September. Its Opus 5.5 page warned that "benchmark margins have become a less reliable guide to real-world differences." The same caution applies in reverse now.
Which jobs should go to Fable 5.2?
Based on where the published gaps are largest, these workloads have the strongest case for Fable 5.2:
- Long-horizon agent runs. Remote Labor Index and WANDR are multi-step, tool-heavy evaluations. Fable leads by 20 and 10 points respectively.
- Agentic search and research. WANDR at 82.1% against 72.3%, and Terminal-Bench-Science at 74.6% against 58.7%.
- Robotics and simulation. RoboDojo at 64.9% against 39.4%.
- Very large parallel jobs. Anthropic says about 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in roughly 30 hours, while Opus 5.5 reached about 26% in 18 hours. Our earlier Fermat Lean coverage explains what that does and does not prove.
- Electronics and CAD work. Anthropic and Diode report Fable 5.2 scoring 69% on Diode PCB Gym; see the PCB design post.
Keep Opus 5.5 for everything else:
- Well-scoped code changes where FrontierCode shows a tie.
- High-volume routine agent loops where $4/$20 matters.
- Pipelines you have already tuned, tested and cached against Opus 5.5.
How should I route between them?
A rail switch sending a track to one of three destinations, a picture of routing Fable 5.2 and Opus 5.5 by task
The pattern that emerged in September still works, with the roles sharpened. Use Fable 5.2 to plan, search and decide. Use Opus 5.5 to execute the well-defined steps. In September, one Hacker News commenter summed it up as "plan with fable, implement with opus," and Anthropic's own advisor tool is built for this shape.
Here is a routing sketch you can adapt. It is a heuristic, not an Anthropic recommendation.
def pick_claude_model(task):
# Hard, open-ended or long-horizon work: pay for Fable 5.2.
if task.horizon_steps > 50 or task.needs_web_research or task.kind in {"robotics", "science"}:
return "claude-fable-5-2", {"effort": "medium"}
# Well-scoped coding and routine agent loops: Opus 5.5 is cheaper and ties on FrontierCode.
return "claude-opus-5-5", {"effort": "high"}
Two practical notes. Start Fable at Medium effort, not Max, because the matched-cost results Anthropic cites were at Low and Medium. And re-measure tokens per successful task for a week before you commit, because the Playcode result shows max-effort Fable can overspend.
What breaks when you move from Opus 5.5 to Fable 5.2?
Anthropic's release notes list hard failures, not warnings. Our Fable 5.2 migration guide covers each in detail. The short version:
| Change | What happens if you ignore it |
|---|---|
| Prefix check enforced on every account | A 400 error if you edit the system prompt, tools or earlier turns and replay a later thinking block |
| Computer use toolset | computer_20251124 returns a 400; declare computer_toolset_20260801 |
| Advisor tool pairing (beta) | A Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as advisor |
| Model switches mid-thread | Moving onto Fable 5.2 keeps Opus 5.5 reasoning blocks; moving away drops Fable's thinking blocks |
| Data retention | 30-day retention applies; zero data retention needs separate authorization |
Fast mode (speed: "fast") doubles the rate to $16/$80 for up to 2.5x output speed. Treat it as a latency purchase for interactive use, and benchmark it separately.
What are people saying about Fable 5.2 vs Opus 5.5?
Developer reaction on Hacker News focused on pricing and versioning more than math. Commenters noted immediate Claude Code availability, debated whether Fable outranking Opus sits next to alleged Opus regressions, and flagged writing-style complaints compared with GPT or Gemini. One thread cited Remote Labor Index near 41.3%. These are commenter reports, not independent benchmarks.
The writing-style point deserves weight. In September, the most common reason developers preferred Opus 5.5 over Fable 5.1 was how it communicates, not raw capability. Nothing in Anthropic's benchmark table measures that. If your product shows model output to users, test readability on your own prompts.
Honest limitations
- Every score above is lab-reported except the Artificial Analysis figure, and even that mixes evaluations that change between versions.
- Opus 5.5's price is not restated in the November 9 post. The $4/$20 figure comes from the earlier Opus 5.5 launch and Anthropic's models table.
- Effort settings differ across rows. Fable 5.2 is listed at Xhigh on FrontierCode and at Max on RoboDojo; Opus 5.5's effort level is not always stated, so rows are not strictly like for like.
- Cost-at-equal-quality claims come from Anthropic. The Playcode data point shows the opposite on one task, so verify on your own workload.
- Safety caveat. Anthropic says both models voiced suspicion of being evaluated more often than recent models, and it limited AI R&D capability to roughly Opus 5.5 level. See our system card and METR coverage.
- We did not run these models ourselves for this post. We compiled published numbers, so treat the routing sketch as a starting hypothesis.
What this means for what you build or pay
If your Claude bill is dominated by routine coding loops, stay on Opus 5.5. The coding tie at half the per-token price is the strongest fact in the table. If you run long agents, deep research or robotics-style work where Remote Labor Index and WANDR look like your task, test Fable 5.2 at Medium effort this week and compare cost per successful artifact.
Do not switch on benchmark deltas alone. Run the same 20 to 50 real tasks through both models, record tokens and pass rate, and decide on dollars per solved task. For a wider view of how Fable 5.2 sits next to Google, OpenAI, SpaceXAI and open-weight models, read our frontier model comparison.
Update — November 10, 2026: Remote Labor Index coverage explains the Scale/CAIS 41.3% automation metric cited on the launch table.
Related on explainx.ai
- Frontier models compared: Fable 5.2, GPT-6 Astra, Gemini 4 Carbon, Grok 4.7, Kimi K3.1 and DeepSeek
- Claude Fable 5.2 and Mythos 5.2: API pricing, migration and what changed
- Fable 5.1 vs Claude Opus 5.5: which one do you actually need
- Claude Opus 5.5 launch: benchmarks, pricing and reaction
- Claude Opus 5.5 vs Sonnet 5.5
- Artificial Analysis Intelligence Index v5
- Output-token inflation: why cheaper per token is not cheaper per task
- How to read AI benchmarks without getting fooled
Primary sources: Anthropic's Fable 5.2 and Mythos 5.2 announcement, Claude Platform release notes, Claude models overview.
Benchmarks, prices and API rules are accurate as of November 10, 2026. Check Anthropic's pricing page before you change production defaults.
Update — November 10, 2026: Dedicated coverage of the Remote Labor Index explains what Fable 5.2's 41.3% automation rate measures on Scale's leaderboard.