Model comparison · Updated Nov 9, 2026
Claude Fable 5.2 vs Opus 5.5: is the stronger model worth twice the token price?
Anthropic shipped Claude Fable 5.2 on November 9, 2026. It costs $8 per million input tokens and $40 per million output tokens. Opus 5.5 costs $4 and $20. Same 1M context, same 128K output cap. So the question is narrow: on your workload, does Fable 5.2 finish enough extra tasks to cover the doubled rate? Every number on this page comes from Anthropic's docs and launch post.
The short answer
- Start on Opus 5.5. That is Anthropic's own advice for most workloads.
- Move a workload to Fable 5.2 when your evals on Opus 5.5 at high or xhigh effort still fall short.
- Per token, Fable costs 2x. Per successful task, it can cost less if it saves retries.
- Fable on the Batch API ($4 / $20) costs the same as Opus 5.5 at standard rates.
Fable 5.2 vs Opus 5.5 side by side
From Anthropic's models overview, pricing page and Fable 5.2 model page. The Bedrock ID for Fable 5.2 was blank in the docs on launch day and now reads anthropic.claude-fable-5-2. Check your Bedrock region before you plan around it.
| Claude Opus 5.5 | Claude Fable 5.2 | |
|---|---|---|
| Claude API model ID | claude-opus-5-5 | claude-fable-5-2 |
| Amazon Bedrock ID | anthropic.claude-opus-5-5 | anthropic.claude-fable-5-2 (added to the docs after launch) |
| Anthropic's description | Long-running agentic coding and knowledge work | Demanding reasoning and long-horizon agentic work |
| Input price | $4 / MTok | $8 / MTok |
| Output price | $20 / MTok | $40 / MTok |
| Cache read (hit) | $0.20 / MTok (5% of input) | $0.20 / MTok (2.5% of input) |
| Cache write, 5 min / 1 hour | $5 / $8 per MTok | $10 / $16 per MTok |
| Batch API (in / out) | $2 / $10 per MTok | $4 / $20 per MTok |
| Fast mode (in / out) | $8 / $40 per MTok | $16 / $80 per MTok |
| Context window | 1M tokens | 1M tokens |
| Max output (sync Messages API) | 128K tokens | 128K tokens |
| Thinking | Adaptive (always on) | Adaptive (always on) |
| Default effort on the API | medium | high |
| Comparative latency | Moderate | Moderate |
| Reliable knowledge cutoff | Jun 2026 | Jun 2026 |
| Data retention | Can run under zero data retention | Requires 30-day retention unless Anthropic authorizes ZDR |
| Retirement | Not sooner than Sep 22, 2027 | Not sooner than Nov 9, 2027 |
Two rows decide most of the bill. Cache reads cost the same $0.20 per million tokens on both models, because Fable's cache discount is deeper (2.5% of input against 5%). And the default effort differs: a request that leaves effort unset runs at high on Fable and medium on Opus, so Fable thinks longer and bills more output with no change in your code.
Benchmarks Anthropic published
These come from the launch announcement. They are Anthropic's runs, and effort settings differ by cell, so read the footnotes there before quoting one number.
| Claude Opus 5.5 | Claude Fable 5.2 | |
|---|---|---|
| FrontierSWE v2 (agentic coding) | 62.3% | 69.9% |
| FrontierCode 2.0 (Fable at xhigh) | 45.5% | 49.1% |
| AutomationBench (business workflows) | 42.5% | 51.4% |
| Humanity's Last Exam (with tools) | 67.7% | 74.2% |
| OSWorld 2.1 (offline, partial) | 87.2% | 89.3% |
| Remote Labor Index | 21.3% | 41.3% |
The row I'd look at first isn't in the table. Anthropic says that on FrontierCode 2.0, Fable 5.2 at high effort scores about the same as Opus 5.5's best score (45.3% against 45.5%) "for roughly half the cost per task." So Fable at a lower effort can match Opus at its highest, and spend less doing it. That is one benchmark, measured by the vendor. Test it on your own tasks before you budget around it.
When each one wins
Anthropic's models overview says: start with Opus 5.5 for most workloads, and use Fable 5.2 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." The launch post calls Opus 5.5 "our strongest daily driver" and says Fable is for "the most ambitious, open-ended technical work."
Opus 5.5 wins when
- It already passes your evals at high or xhigh effort.
- Volume is high and each task is short: chat, code review comments, ticket triage.
- Your organization runs under zero data retention.
- You need the lowest cost per call and a retry is cheap.
Fable 5.2 wins when
- Agents run for hours and Opus 5.5 stalls or loses the thread.
- Opus 5.5 at its best effort still fails tasks you care about.
- A failed run costs more than the token gap, like a human redo or a bad deploy.
- You are on Fable 5.1 now. Fable 5.2 is 20% cheaper.
Cost per task, worked out
Arithmetic from Anthropic's list prices only. The token counts are example sizes for a mid-size coding agent task, not measurements. Swap in your own counts from a real eval run.
| Scenario | Opus 5.5 | Fable 5.2 | How |
|---|---|---|---|
| A. Agent task, no cache: 200K input, 20K output | $1.20 | $2.40 | Plain 2x. Opus: 0.2 × $4 + 0.02 × $20. Fable: 0.2 × $8 + 0.02 × $40. |
| B. Same task, 180K of the input read from a warm cache | $0.52 | $1.00 | Cache reads cost $0.20 / MTok on both, so the gap narrows to about 1.9x. Cache writes not counted. |
| C. Task A on the Batch API | $0.60 | $1.20 | Fable on Batch costs the same as Opus 5.5 at standard rates. |
| D. Task A in fast mode | $2.40 | $4.80 | Fast mode doubles the standard rate on both and can't be combined with Batch. |
The number that matters: cost per successful task
Divide the cost per attempt by the pass rate. Take scenario A. If Opus 5.5 passes 30% of a hard task set and Fable 5.2 passes 70%, Opus costs $1.20 / 0.30 = $4.00 per success and Fable costs $2.40 / 0.70 = $3.43. Those pass rates are made up to show the math. The rule they show is real: at equal token counts, Fable 5.2 is cheaper per success once its pass rate is more than double Opus 5.5's. On easy tasks where both pass, Opus 5.5 is half the price.
Token counts won't stay equal either. Fable at high may think more than Opus at medium, or finish in fewer turns. Only your logs can tell you which. For full price lists across every model, see our LLM API pricing comparison and the AI model price index.
Decision checklist
- Have you run your eval set on Opus 5.5 at high and xhigh effort? If not, do that before testing Fable 5.2.
- Does Opus 5.5 still fail tasks that matter, at its best effort? If it passes, stay. Fable can't beat a pass.
- Is the workload long-horizon: multi-hour agent runs, large refactors, open-ended research? That is where Anthropic points Fable.
- Does a failed run cost more than the token gap? A redo by a senior engineer or a broken deploy usually does.
- Does your organization need zero data retention? Fable 5.2 needs Anthropic's approval for ZDR. Opus 5.5 doesn't.
- Have you set effort explicitly? Fable defaults to high on the API, Opus to medium. Compare at matched settings.
- Can the job wait for Batch? At 50% off, Fable on Batch costs what Opus costs live.
- Are you on Fable 5.1 today? Fable 5.2 is cheaper ($8 / $40 against $10 / $50). That switch needs no debate.
Ready to move a workload? Our Opus 5.5 to Fable 5.2 migration guide covers the model swap, effort settings, the 400s that catch teams out, and the rollback plan.
Want someone to run the evals for you?
The comparison above needs a test harness, two models and a sweep of effort levels on your real tasks. AY Automate places AI engineers inside your team to do that work, from $3K a month with a 90-day replacement guarantee.
FAQ
Is Claude Fable 5.2 better than Opus 5.5?
On the benchmark rows Anthropic published for both models, Fable 5.2 scores higher: FrontierSWE v2 69.9% vs 62.3%, AutomationBench 51.4% vs 42.5%, Remote Labor Index 41.3% vs 21.3%. Anthropic still recommends starting most coding and knowledge work on Opus 5.5 and moving to Fable 5.2 when evals on Opus 5.5 at higher effort fall short.
How much more does Fable 5.2 cost than Opus 5.5?
Twice as much per token. Fable 5.2 is $8 per million input tokens and $40 per million output tokens. Opus 5.5 is $4 and $20. Cache reads are $0.20 per million on both, so heavily cached workloads see a gap a little under 2x.
Can Fable 5.2 cost less per task than Opus 5.5?
Yes, in two ways. Anthropic reports that on FrontierCode 2.0, Fable 5.2 at high effort scores about the same as Opus 5.5's best score (45.3% vs 45.5%) for roughly half the cost per task. And if Fable passes a task in one attempt where Opus needs two or more, the cost per successful task can come out lower.
Do Fable 5.2 and Opus 5.5 have the same context window?
Yes. Both have a 1M token context window and 128K max output tokens on the synchronous Messages API. Both bill the full 1M window at standard rates.
Is Claude Fable 5.2 faster than Opus 5.5?
Anthropic's models overview rates both as moderate latency. The launch post says outputs are up to 30% faster than Fable 5.1, which is a comparison with the older Fable, not with Opus 5.5. Fast mode on Fable 5.2 speeds up generation by up to 2.5x at $16 / $80 per MTok.
What are the model IDs?
claude-fable-5-2 and claude-opus-5-5 on the Claude API, Claude Platform on AWS, Microsoft Foundry and Claude Models on Google Cloud. On Amazon Bedrock the docs now list anthropic.claude-fable-5-2 and anthropic.claude-opus-5-5.
Has Artificial Analysis or another third party compared them?
Third-party leaderboards are still adding Fable 5.2 in the days after launch. This page uses Anthropic's published numbers only. Run your own evals before you trust anyone's table, ours included.
Keep reading
- Claude Fable 5.2 pricing, with fast mode, batch, cache and monthly cost examples
- Migrate Claude Opus 5.5 to Fable 5.2
- Claude Fable 5.2 for engineering teams
- Hire Claude Fable 5.2 engineers
- Compare every Claude model on scores and specs
- Claude Fable 5.1 vs Opus 5, the last round of this decision