Benchmark guide
Claude Fable 5.2 Benchmarks and What They Measure
A benchmark table can help you choose which model to trial, provided you know what each row actually measures. We reviewed Anthropic's Fable 5.2 launch table and its methodology notes. These are reported evaluation scores, followed by an illustrative way to turn one result into a useful internal test.
The quick answer
Use Chartography to motivate a chart-reading trial, FrontierCode to motivate a software trial, and AutomationBench to motivate a workflow trial. Compare rows with their effort and tool conditions attached. For a creative team deciding whether it can trust an image audit, start with Chartography and test a narrow visual task.
| Evaluation | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Chartography without tools | 73.4% | 44.8% | 64.4% |
| FrontierCode 2.0 | 49.1% at Xhigh | 32.4% at Xhigh | 45.5% |
| AutomationBench | 51.4% | 31.4% | 42.5% |
Read across a single row
Anthropic generally reports the highest available effort unless it calls out a different setting. Compare the models under the conditions of one task and read the footnote before assuming effort or tool use was identical. Rows explicitly labeled Mythos describe Mythos under the stated conditions.
Chartography for visual evidence
The published task is chart recognition without tools. A relevant studio pilot might check whether each model can identify the unit, category and date on a campaign report. Score exact evidence and correct abstentions when a value is unreadable. A strong chart score leaves composition and brand judgment to a separate review.
FrontierCode for coding work
This evaluation concerns changes in software repositories. If the design team hands prototypes to engineers, test one representative fix with versioned inputs and automated checks. Keep the same review process for each candidate and record whether the patch meets its actual acceptance criteria.
AutomationBench for handoffs
Turn one repetitive handoff into a local pilot with a known answer. For example, map each approved asset to one channel and size. Record whether the model assigns the right item and whether it flags missing dimensions before proposing an export.
Keep a pilot result you can audit
Give each case an ID, expected facts and a pass rule. Capture effort, actual tools, cost and reviewer corrections for each model. Make the decision on accepted deliverables. Put the references and accepted findings on the same canvas when the work moves to production.
Keep a result row that another reviewer can reproduce
A useful pilot row contains the source-packet ID, question, expected fact, model, settings, finding, correction and final decision. For a chart check, the expected fact might be the axis unit; do not replace it with an impression such as “good analysis.”
Score an unreadable label as unresolved when the model correctly asks for a clearer source. Keep unsupported guesses as failures even if the guessed number happens to match. That distinction tests whether the review can be trusted when the team does not already know the answer.
Frequently asked questions
Can I treat the table as an independent benchmark?
These rows come from Anthropic's launch report. Use the cited methodology and a local pilot to interpret them.
Should I use a single aggregate score?
Choose the evaluation that matches the job. A separate creative acceptance test is more useful than an aggregate for deciding whether a design proof is ready.
Keep pilot evidence on one board
Bring in the original charts and screenshots, then place the accepted findings next to the evidence the team can verify.