Disclosure: we build CodeRouter, a router for coding agents. For this test we called Claude Fable 5.2 directly through Anthropic's API, not through CodeRouter.
Update, Nov 12, 2026: since November 11, CodeRouter bills model list price with no markup, plus a 5.5% payment fee when you buy credit (a $20 top-up costs $21.10). Where this post says list price +8%, that was CodeRouter's price when it was published. Current terms: pricing.
Anthropic released Claude Fable 5.2 on November 9. Its announcement says that in early testing on FrontierCode 2.0, Fable 5.2 at High effort "scored about level with Opus 5.5 at Max effort, for roughly half the cost per task."
The same day we ran our coding-agent benchmark on Opus 5.5, Sonnet 5.5, Haiku 5.5, OpenRouter Auto Code and our own router (results). Then we added Fable 5.2 on a random 30 of those 99 tasks, one run, at the API's default settings. It solved the same 15 tasks as Sonnet 5.5, at 20× Sonnet's cost and 8× Opus's.
This is a quick read on a small sample, not a ranking. Here is what we measured.
The results
Fable 5.2 ran on 30 tasks drawn at random (with a fixed seed) from the benchmark's 99. Six of its sessions were stopped by our budget for the run before they finished. As we decided before the run, those are left out, so the comparison is on the other 24 tasks. Every other row is the November 9 first run on the same 24 tasks. A task counts as solved when the repository's tests pass after the agent's change. Costs are list prices.
| Setup | Solved (of 24) | Cost per task | Cost per solved task |
|---|---|---|---|
| Claude Fable 5.2 | 15 | $1.571 | $2.513 |
| All Claude Opus 5.5 | 16 | $0.196 | $0.293 |
| OpenRouter Auto Code | 16 | $0.228 | $0.342 |
| All Claude Sonnet 5.5 | 15 | $0.077 | $0.123 |
| CodeRouter default routing | 14 | $0.070 | $0.120 |
| All Claude Haiku 5.5 | 11 | $0.008 | $0.017 |
- Against Opus 5.5: one task fewer (15 vs 16: −4.2 points, 95% interval −20.8 to +8.3, exact McNemar p = 1.0). Fable solved one task Opus didn't; Opus solved two Fable didn't. Cost per task 8.0× (interval 6.3–10.2×), cost per solved task 8.6×.
- Against Sonnet 5.5: exactly the same 15 tasks, at 20.4× the cost.
- Counting the six budget-stopped sessions as failures: Fable 15 of 30, Opus 18 of 30 (p = 0.375). In the November 9 run, Opus had solved 2 of those 6 tasks, and no setup had solved 3 of them.
With 24 tasks, a completion rate has a wide interval (Fable: 62.5%, 43–79%). On completion, we can't tell Fable, Opus and Sonnet apart here. The cost difference is large and clear.
Where 8× comes from
Two factors multiply:
- Price, 1.8×. Fable 5.2 lists at $8 per million input tokens and $40 per million output tokens, twice Opus 5.5 ($4 and $20). Cache reads are $0.20 per million for both, so on an agent's mix of tokens the gap is less than 2×: Fable's own sessions cost 1.8× what they would have at Opus's rates.
- Behavior, 4.4×. Fable's own sessions, priced at Opus's rates, cost 4.4× what Opus's did. Fable thinks more on every step and takes more steps, so the context it re-reads keeps growing:
| Per request or session (24 tasks) | Fable 5.2 | All Opus 5.5 | Ratio |
|---|---|---|---|
| Cost per request | $0.087 | $0.018 | 5.0× |
| Output tokens per request, thinking included | 1,263 | 374 | 3.4× |
| Requests per session | 18.0 | 11.2 | 1.6× |
| Output tokens per session | 22,681 | 4,182 | 5.4× |
| Cache-read tokens per request | 37,850 | 12,077 | 3.1× |
78% of Fable's requests included a thinking block, against 45% of Opus's. 10 of Fable's requests hit the 8,192-token output limit; none of Opus's did.
What this means if you use Fable 5.2 for coding agents
- Check the effort setting. Our run used the API default, which for Fable 5.2 is high effort. Anthropic says Medium is the default in Claude apps and Claude Code, so a Claude Code session would likely think less and cost less than this. We haven't measured that yet.
- Budget per task, not per token. At $2.50 per task, the limit the other setups ran with, 7 of Fable's 30 sessions would have been cut off. We raised it to $10 for this run so the limit wouldn't decide the result.
- For CodeRouter users: smart routing doesn't send requests to Fable 5.2. You can still pick it by name (Direct mode), billed at list price +8%.
Why this differs from Anthropic's result: a different benchmark, different settings (Anthropic compared Fable at High with Opus at Max; we ran each model at its API default) and a different agent. We're not saying Anthropic's number is wrong on FrontierCode. On our tasks, with our harness, we didn't see Fable do more for its cost.
How we ran it
- Tasks: 30 drawn with Python's
random.Random(20261110).samplefrom the 99 tasks of our November 9 benchmark (SWE-rebench 2026_03 split, real pull requests from Python repositories). We wrote the sample, the limits and how we'd compare into a dated internal document before the run. - Agent: mini-swe-agent 2.4.6 with its bash tool and its SWE-bench prompts unchanged,
max_tokens8,192, no temperature, thinking or effort parameter. Limits per session: 150 steps, $10, 60 minutes, 60 seconds per command. - Model:
claude-fable-5-2through Anthropic's Messages API with a separate key. Every one of the 611 responses reportedclaude-fable-5-2. - Offline: containers had no network access; we checked that with a probe before each start. 0 of the 30 patches matched the upstream fix.
- Grading: swebench 5.0.2 with SWE-rebench's evaluation spec. On the same machine, every task's upstream fix passed and an empty change failed.
- Cost: every request's input, cache-read, cache-write and output tokens × Anthropic's list price on November 10, 2026.
- One interruption: the run stopped once on an API billing error on our side. The 12 sessions that had already finished were kept; the other 18 tasks were rerun from scratch about 45 minutes later, with the same settings. We recorded that change before the rerun.
Limitations
- 30 tasks, one run, 24 in the main comparison. A one- or two-task difference is noise at this size.
- The other setups ran on November 9 and Fable on November 10, not interleaved, so day-to-day changes in models or serving aren't controlled.
- The other setups were called through CodeRouter's pass-through path, which adds its own prompt-cache markers; Fable was called directly, with only the agent's own cache markers. So prompt caching wasn't set up identically.
- Fable had a $10 per-session limit and the others $2.50. No Opus session in the full benchmark reached $2.50, so the higher limit only mattered for Fable.
- The six budget-stopped sessions were among the longest and most expensive, so leaving them out may flatter Fable's cost. Counting them, Fable cost $1.76 per task.
- One agent harness with a bash tool, at the API's default effort. Claude Code, Codex or Cursor sessions use other tools, prompts and effort settings.
- List prices on November 10, 2026. Models and prices change.
The data
Download the summary JSON and a per-session CSV for Fable 5.2 (30 rows: task, solved, stopped by budget, cost, steps and tokens), next to the 99-task results for the other setups. No agent trajectories.