Cursor's Claude Fable 5.2 Tops CursorBench at 60% With 20% Cheaper Tokens
Cursor adds Anthropic's updated Fable 5.2 model, pushing CursorBench to a new 60% ceiling while cutting token prices by 20% across the board.
- Claude Fable 5.2 is live in Cursor, topping CursorBench at 60% on Max effort.
- Default effort scores 56.9%, up 7.7 points from Fable 5.1's 49.2%.
- Priced at $8/M input and $40/M output tokens, 20% cheaper than Fable 5.1.
- Fable 5.2 Max costs $32.79 per task and takes 218 steps on average.
- Effort levels behave differently from 5.1; start at Default and recalibrate.
- Opus 5.5 High remains cheaper per task at $3.97 for 56.0%.
Cursor has added Claude Fable 5.2 to its model lineup. The release leads CursorBench 4.0 at Max effort with a 60.0% score. At the benchmark’s default High setting, it improves from Fable 5.1’s 49.2% to 56.9%. Token rates fell 20%, although the average benchmark task cost slightly more because total cost also depends on token usage.
Fable belongs to Anthropic’s Mythos-class model family. Cursor’s model documentation positions it as Anthropic’s most capable generally available option for autonomous knowledge work and coding. Its focus is long-running work that requires an agent to preserve intent while searching a repository, editing files, running tools, and correcting failures.
Behind the 60% result
CursorBench evaluates coding agents on ambiguous, multi-file tasks drawn from real Cursor sessions. Cursor cites one task that begins with a stack trace and the word “fix,” leaving the agent to infer the request, locate the root cause, implement a change, and validate it. The benchmark reflects Cursor’s own agent environment, so rankings may differ across IDEs, repositories, and independent evaluations.
| Model | Score | Average cost per task | Average steps per task |
|---|---|---|---|
| Fable 5.2 Max | 60.0% | $32.79 | 218 |
| Opus 5.5 Max | 57.8% | $13.43 | 185 |
| Fable 5.2 High (default) | 56.9% | $9.70 | 69 |
| Opus 5.5 High | 56.0% | $3.97 | 68 |
| Sonnet 5.5 Max | 55.5% | $7.05 | 170 |
| Fable 5.2 Medium | 54.8% | $4.66 | 35 |
| Fable 5.1 High | 49.2% | $9.08 | 77 |
Fable 5.2 High gains 7.7 percentage points over Fable 5.1 High while using eight fewer steps on average. Its average task cost rises from $9.08 to $9.70 despite the lower token rates, indicating that token volume and the input-output mix offset the price reduction in these runs.
Max effort adds 3.1 points over High while using about 3.2 times as many steps and costing about 3.4 times as much. Against Opus 5.5 Max, it gains 2.2 points at roughly 2.4 times the cost. Those figures make Max most suitable for difficult tasks where another attempt, a failed migration, or extended human review would cost more than the additional model usage.
Cheaper tokens, uneven task bills
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1 in each category. Per-task spending still varies with context size, generated output, retries, and the length of the agent loop.
- Fable 5.2 Low scores 53.2% at an average cost of $3.94 per task.
- Fable 5.2 Medium reaches 54.8% for $4.66 and uses 35 steps, about half as many as High while giving up 2.1 points.
- Opus 5.5 High scores 56.0% for $3.97, trailing Fable 5.2 High by 0.9 point at about 41% of its average task cost.
Cursor says effort levels behave differently in Fable 5.2. Existing High or Max preferences from Fable 5.1 should be reset to Default and recalibrated against representative repository tasks before wider deployment.
Where longer runs pay off
Fable’s main use case is autonomous work that spans many searches, edits, tests, and corrections. Cursor attributes stronger intent retention to the model, while the benchmark shows how much computation Max can use: 218 agent steps per task on average. That profile fits several workloads:
- Multi-file refactors that require investigation before editing.
- Underspecified bug reports where the agent must infer intent from logs, tests, and surrounding code.
- Background migrations or maintenance jobs that can run with limited supervision.
- Repository-wide changes where validation and recovery matter as much as code generation.
Fable applies strict safety guardrails. Cursor says requests stopped by those controls are routed automatically to Claude Opus, which can change model behavior during a task. Teams comparing output, latency, or cost across runs should account for those fallbacks.
Some Fable 5.1 users also reported token totals exceeding the figures shown in the session interface. Those reports concern the previous release and do not establish the same behavior in 5.2. Per-token deployments can reconcile the usage dashboard with session records before allowing unattended Max runs at scale.
Choosing an effort level
Opus 5.5 High remains the economical choice when its 56.0% benchmark score is sufficient. Fable 5.2 High offers a small score increase and a larger improvement over Fable 5.1 for long, ambiguous work. Medium and Low provide tighter cost controls, while Max reserves substantially more compute for tasks where a 3.1-point gain over High can justify an average cost near $33 per attempt.
Comments
