Did Claude Opus 5 Overtake Fable 5? Benchmarks, API Pricing, and a Repository Test Plan
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
Last checked: 2026-07-27.
The short answer: several wins, but not a universal overtake
Claude Opus 5 became available on July 24, 2026 under the API ID claude-opus-5, priced at $5/M input and $25/M output. Its appeal is not a single overall leaderboard. It is that Anthropic has brought several results close to, or above, Fable 5 while charging half the base API rate.
Calling this a universal overtake would still be wrong. Opus 5 leads in Anthropic's stated Frontier-Bench v0.1 and OSWorld 2.0 runs. On CursorBench 3.2, maximum effort comes within 0.5% of the Fable 5 peak, but Cursor's wording is still just under Fable 5. The useful conclusion is that Opus 5 has shifted the price-performance frontier, with task-specific wins rather than a clean sweep.
| Test | Reported Opus 5 result | What the test mainly covers |
|---|---|---|
| Frontier-Bench v0.1 | Ahead of other models in Anthropic's stated run | Real software tasks using mini-SWE-agent on GKE, averaged over five attempts |
| CursorBench 3.2 | Within 0.5% of the Fable 5 peak at maximum effort | Cursor's coding tasks; Opus 5 remains slightly below Fable 5 here |
| OSWorld 2.0 | Above Fable 5's best result at just over one-third of the cost | Computer-use tasks, not coding performance in general |
Source image: Anthropic's Claude Opus 5 launch page, accessed 2026-07-27. Pricing and benchmark wording in this article are checked against the launch page, model pages, and system card.
Price is the clearest change
Fable 5 is listed at $10/M input and $50/M output, exactly twice the Opus 5 base rates. For a 24-task repository packet consuming 18M input tokens and 1.2M output tokens, the simple list-price calculation is about $120 on Opus 5 and $240 on Fable 5.
Equal token volume is not equal work. Models may consume different amounts of context, retry differently, or diverge on refusals and tool calls. A useful comparison asks how many identical tasks passed their tests, how much maintainer rework remained, and what the final bill was.
Fast mode is a separate trade-off. Anthropic says it runs at about 2.5x default speed for 2x the base price. That can make sense for interactive work where waiting is expensive. Batch jobs may be better served by the default mode. In either case, measure first-token latency, total completion time, retries, and cost in the team's own environment.
Move from token price to cost per accepted task
Add one more column to the model sheet: cost per accepted task. Take the model bill, add the team's internal cost for maintainer rework, and divide by tasks that actually pass acceptance. A generated patch does not enter the denominator until tests pass, unrelated changes are acceptable, and a maintainer considers it mergeable.
A cheaper model can become the expensive option if people repeatedly restore context, rerun tools, or repair broad edits. Fast mode can produce the opposite result: even at twice the list price, it may be economical when it removes costly waiting from an interactive task. Preserve four values for every run: model charges, machine runtime, maintainer minutes, and accepted tasks.
Public benchmarks are therefore candidate filters, not a weighted scorecard. Frontier-Bench, CursorBench, and OSWorld use different work and environments, so averaging them would hide the distinction. In an internal test, count “produced an answer” separately from “delivered a mergeable result,” and do not let strength on one task type cancel repeated failure on another.
Payment-service repository: a fair 24-task comparison
A practical internal test starts with one frozen payment-service repository commit and 24 defect or cross-file change tasks. Both candidates receive the same code, task text, and test environment.
Record the exact model ID, Fast mode state, input and output tokens, patch, and reconstructable cost. Apply each patch, run the existing suite and new regression tests, then have a maintainer inspect unrelated-file changes. Exclude a result if tests were not run, cost cannot be reconstructed, or an automatic fallback concealed a refusal.
The final sheet only needs task completion, test result, maintainer rework time, unrelated-file count, and real cost. That evidence is more useful for adoption than a headline about one benchmark win.
How to decide now
Teams already using Fable 5 have a strong reason to run an identical-task replacement test, but not to switch everything immediately. CursorBench suggests the two are extremely close on demanding coding work. Frontier-Bench and OSWorld point to task-specific upside for Opus 5. The lower price makes evaluation attractive; the repository's own tests and bill should make the decision.