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.
Yesterday we published Claude Fable 5.2 on 30 tasks from our coding-agent benchmark: at the API's default settings it solved the same tasks as Sonnet 5.5 and cost 8× as much as Opus 5.5. We noted one gap: the API default is high effort, while Claude Code defaults Fable to Medium.
So we ran the same 30 tasks again with one change, effort: medium. It solved the same 15 tasks as the high-effort run, for 72% less. It still cost 2.2× Opus.
The results
Same 30 tasks, same agent harness, same limits, one run. In the first run, six high-effort sessions were stopped by our budget for the run before they finished; this time all 30 finished. So the Medium-vs-High comparison uses the 24 tasks both runs completed, and the comparison with the other setups uses all 30. Costs are list prices.
Medium vs High, 24 tasks:
| Setup | Solved | Cost per task | Steps per task | Output tokens per step |
|---|---|---|---|---|
| Fable 5.2, Medium | 15 | $0.436 | 8.6 | 601 |
| Fable 5.2, High (API default) | 15 | $1.571 | 18.0 | 1,263 |
- Exactly the same 15 tasks solved (exact McNemar p = 1.0).
- Cost: 0.28× (95% interval 0.22–0.35×). Medium was cheaper on every one of the 24 tasks.
Medium vs the other setups, all 30 tasks (the others are the November 9 first run on the same tasks):
| Setup | Solved (of 30) | Cost per task | Cost per solved task |
|---|---|---|---|
| Claude Fable 5.2, Medium | 17 | $0.466 | $0.822 |
| All Claude Opus 5.5 | 18 | $0.216 | $0.360 |
| All Claude Sonnet 5.5 | 17 | $0.082 | $0.144 |
| CodeRouter default routing | 15 | $0.081 | $0.163 |
| All Claude Haiku 5.5 | 13 | $0.010 | $0.023 |
- Against Opus 5.5: one task fewer (17 vs 18, p = 1.0) at 2.15× the cost (interval 1.88–2.53×).
- Against Sonnet 5.5: the same number solved (17 vs 17, p = 1.0) at 5.7× the cost.
On 30 tasks, a completion rate has a wide interval (Fable Medium: 56.7%, 39–73%). We can't tell Fable, Opus and Sonnet apart on completion here. On cost, the gaps are clear.
Where the 72% went
Medium cut the cost two ways, about equally: Fable took half as many steps, and each step cost 42% less because it wrote and thought less and re-read a shorter conversation (24 tasks, per request unless noted):
| Medium | High | Ratio | |
|---|---|---|---|
| Steps per task | 8.6 | 18.0 | 0.48× |
| Output tokens per step, thinking included | 601 | 1,263 | 0.48× |
| Thinking tokens per step (estimated) | 172 | 711 | 0.24× |
| Steps with a thinking block | 54% | 78% | |
| Cache-read tokens per step | 14,603 | 37,850 | 0.39× |
| Cost per step | $0.051 | $0.087 | 0.58× |
Fewer steps also means a shorter conversation to re-read on each one, which is why cache reads per step fell too. No Medium response hit the 8,192-token output limit; ten High responses did.
Against Opus 5.5, Fable at Medium now takes fewer steps (8.6 vs 11.2), but each step costs 2.9× as much. Fable's input, output and cache-write prices are twice Opus's (cache reads cost the same), and it writes 1.6× the output per step.
What this means
- If you use Fable 5.2 through the API, set the effort explicitly. The API default is high. On our tasks, high cost 3.6× as much as Medium for the same tasks solved.
- In Claude Code, the default is already Medium, so these Medium numbers are the closer guide to what a Claude Code session costs.
- On these tasks, Fable at Medium still didn't do more than Opus or Sonnet, and it cost more than both.
- For CodeRouter users: smart routing doesn't send requests to Fable 5.2. You can pick it by name (Direct mode).
How we ran it
Everything as in the first run: the same 30 tasks, mini-swe-agent 2.4.6 with its bash tool, 150 steps, $10 and 60 minutes per session, 60 seconds per command, max_tokens 8,192, containers with no network access (checked with a probe before the start), graded with swebench 5.0.2 and SWE-rebench's evaluation spec. The only change is output_config: {"effort": "medium"} on every request, sent directly to Anthropic's Messages API with a separate key. Every one of the 282 responses reported claude-fable-5-2. On the same machine, every task's upstream fix passed and an empty change failed. We wrote the plan down in a dated internal document before the run.
Limitations
- 30 tasks, one run per setup; a one- or two-task difference is noise at this size.
- The High run was in the morning and the Medium run in the evening of the same day; the Opus, Sonnet, Haiku and CodeRouter runs were on November 9. Not interleaved.
- The High run had six sessions stopped by its budget, so Medium vs High uses 24 tasks.
- The other setups ran through CodeRouter's pass-through path, which adds its own prompt-cache markers; Fable was called directly.
- One agent harness with a bash tool. Claude Code uses other tools and prompts.
- List prices on November 10, 2026.
The data
Download the summary JSON and a per-session CSV (30 rows: task, solved, cost, steps and tokens), next to the High-effort run and the 99-task results. No agent trajectories.