3.11.0: Claude Opus 5.0

Models

Model	PR	Average generation time	Cost	Average JSON size	Max JSON size
Claude Opus 5.0	#68	32m 10.2s (1930.2s)	$89.97 for 15 builds	91.00 MiB	369.57 MiB (worldtree)
Claude Fable 5	#50	18m 04s (1,084.4s)	$54.93 for 15 builds	30.65 MiB	97.39 MiB (worldtree)

(Note: Claude Fable 5 was added in 3.7.0 and is shown here only for comparison)

Notes

Opus 5.0 seems to be a massive jump from Opus 4.8. In fact, it seems quite clear that the model is at or above Fable in this benchmark, so comparing it to Opus 4.8 would honestly be a disservice. Though that doesn't indicate Opus 5.0 would be better than Fable for everyday uses like coding; it seemed that Fable was a lot more conservative in its interpretation of the system prompt, whereas Opus was a lot more liberal, and created scenes in each of its builds; I don't have enough experience with using Opus 5.0 for coding to weigh in on whether it likes to over-engineer solutions or how it interprets the user's prompts, but the model is impressive regardless.

However, like other Anthropic releases, Opus 5.0 is horribly token inefficient. The sole reason the benchmarking cost for Opus 5.0 was so high is because at maximum reasoning effort, the model reaches the token output cap before it can finish its JSON response. Not because the JSON outputs it generates are large, but primarily because it uses so many tokens in its internal CoT process that it ends up having to truncate it's JSON output, causing us to have to reattempt the benchmark builds.

What's Changed

Models

• Added Claude Opus 5.0 with native Anthropic routing, OpenRouter fallback, a 1,000,000-token context window by default, adaptive thinking at max effort, strict structured output, and the 128,000-token synchronous Messages API output cap. #68
• Omitted unsupported sampling controls and added the opus-5 batch/upload slug, model details, and provider request-contract coverage. #68

Generation diagnostics

• Persisted every completed provider attempt as a numbered RAW artifact before extraction and validation so failed responses remain available for debugging. #68
• Kept the unnumbered canonical RAW artifact synchronized with the current run's successful or final failed response. #68

Benchmark automation

• Tracked provider calls, completed responses, rejected responses, failed attempts, failed runs, and interrupted runs per job, accumulated across runs that failed, were interrupted, or resumed later. #68
• Published a cumulative counter only when every job in the cohort tracked it, so a partially tracked cohort reports the counter as untracked instead of a total that excludes earlier runs. #68
• Recorded the effective request settings on each benchmark sample so implicit and explicit provider defaults no longer split a cohort. #68
• Reported provider calls, rejected responses, and failed calls in the batch generation summary. #68

Model details

• Led the model details popover with Statistics, added Total attempts, and revealed cost per attempt on hover or keyboard focus. #68
• Divided benchmark cost by whichever denominator was measured, so models predating attempt tracking report cost per build rather than presenting a build cost as a per-attempt figure. #68

Internals

• Centralized Claude release capabilities — effort ladder, sampling policy, thinking mode, output ceiling, long-context header, and effort environment variable — for both the direct and OpenRouter routes, with no capability inherited from an earlier release. #68
• Centralized per-model output ceilings and sampling policy across every provider, collapsing the separate request and cap tables into one ceiling per model. #68
• Verified identical output budget and sampling decisions across all catalogued model IDs, and capability parity between every Anthropic model ID and its OpenRouter form. #68

Documentation

• Added a guide covering every surface a new model touches, from the catalog entry and request configuration through cohort generation and tests. #68

Changelog

Full Changelog: 3.10.0...3.11.0

• #68 (feat) add Claude Opus 5 and generalize model capability and benchmark telemetry @Ammaar-Alam