Launch report
Claude Haiku 5.5 launch report
ModelCap · generated from recorded data
Anthropic's Claude Haiku 5.5 was first listed on 7 October 2026. It entered the ModelCap Index at #60 of 350 ranked language models on 14 October 2026, 7 days after its listed release date, backed by 3 admitted benchmarks: ARC-AGI-2 Verified, ARC-AGI-3 Verified and LMArena Agent. As of 19 November 2026 it is #93 on the current board.
Position at entry
Entry is the first board position backed by at least two admitted benchmarks. Read from the board's last published snapshot of each day named below; these values are fixed and do not change with later refreshes.
- Listed release date
- 7 October 2026
- Entry · 14 October 2026
- #60 of 350 ranked language modelsBacked by ARC-AGI-2 Verified, ARC-AGI-3 Verified, LMArena Agent
- Neighbours at entry
- Directly above: Ornith 1.5 35B A3BDirectly below: Mistral Large 4
Benchmark results
Published results admitted to the Index, as of 19 November 2026. Each board links to ModelCap's page for it and each result to its source.
Against Claude Haiku 4.5
Claude Haiku 4.5 is the previous release in the same family. Current results for both, as of 19 November 2026.
- ModelCap Index
- Claude Haiku 5.5: #93 on the current board · score 58.3Claude Haiku 4.5: #173 across all versions · score 56.1
Where the tested setting differs between the two rows, the difference compares the published results as they stand, not the same setting.
ModelCap's estimate before measurement
Recorded in the launch ledger before any benchmark result for the model was admitted, and not edited afterwards.
- Estimate · 7 October 2026
- Index score 60.9 (80% interval 47.4–74.5) · estimated position #70
- Basis
- Family succession 78% · Publisher corpus prior 22%
- Measured outcome · 23 October 2026
- Index score 60.1, 0.8 points below the estimate and inside the 80% interval.
Every estimate and its outcome is published, hits and misses alike, in the launch accuracy record.
How this report is produced
ModelCap publishes a launch report when a newly released model first holds a measured position backed by at least two public benchmarks. The report is assembled automatically from ModelCap's own records: the board as it stood at entry, the results currently admitted to the Index and, where one exists, the estimate recorded before measurement. No text is written by hand, and no claim is made that the records do not contain. ModelCap does not run the benchmarks itself; each result links to the organisation that published it.