Claude Opus 5.5

The model is stronger in the Full/Agentic lane than in the SWE lane; the overall score is therefore shown with both component lanes visible. Hard Intelligence score is 96.10 and contributes to the overall score alongside Full/Agentic and SWE. Runs on the Claude Max subscription; the transport is the official Claude Code CLI rather than a metered API key, so the subscription bills no marginal per-token charge. Reasoning effort is extra-high on every lane, matching the published extra-high baselines. Costed at the model’s published Standard API rates ($4.00/1M input, $20.00/1M output, $0.20/1M cache reads, $5.00/1M cache writes) as a comparable basis, not an invoice. The prompt prefix is cached by the subscription route, so pricing only the uncached remainder would understate each call. Thinking tokens are spent and billed on this route but are not returned as text, so reasoning content is absent while its token cost is counted. Hard Intelligence is a public diagnostic, not a hidden official score. Token median and P90 per scored item are intentionally omitted: their derivation could not be reproduced on the same basis as the published rows. Cache writes are priced at the one-hour duration, which is the TTL this route writes with. The first publication priced every write at the short five-minute duration. That duration was reached through the records carrying an explicit cache breakdown, which is the Full lane's direct prompts, the SWE lane, and the Hard diagnostic. The Full lane's agentic prompts store no per-turn cache breakdown and priced their input at the write rate as a whole, so the first correction did not reach them and they stayed on the short rate. Correcting all three lanes moves the total from $2.269694 to $3.119018, with Full carrying $0.73 of that. The overall goes from 81.0355 to 80.8655; the overall rank stays at 10, and every sub-rank is unchanged. Agentic prompts re-measured with a real tool loop. The earlier records were single-turn refusals that scored zero on capability while reporting the speed of the refusal as performance and efficiency; both are now measured from a real loop, so capability rises and the latency and cost terms fall. Net position on this table is unchanged.