Claude Haiku 5.5 Enters Text Arena: Haiku Jumped from 0% to 39.2% on Terminal-Bench 4.0 in One Generation
Arena.ai added Claude Haiku 5.5 (High) to its Text Arena leaderboard on October 22, 2026, three weeks after the model launched on October 7. The Text Arena evaluates human preference through blind side-by-side comparisons and assigns an ELO-style rating; rankings are pending as evaluations accumulate.
The reason the addition is worth noting is the underlying benchmark record. Claude Haiku 4.5, the previous small-model anchor in Anthropic’s lineup, scored 0.0% on Terminal-Bench 4.0 — a zero-shot failure on agentic terminal tasks. Claude Haiku 5.5 scores 39.2% on the same benchmark. That is not a marginal gain. It is the difference between a model that cannot complete terminal-agent tasks and one that succeeds at more than a third of them.
Key Benchmark Numbers
| Benchmark | Score |
|---|---|
| Terminal-Bench 4.0 | 39.2% |
| FrontierCode 1.1 Main | 46.4% |
| OSWorld 2.1 (offline subset) | 72.4% |
| HLE (no tools) | 45.9% |
| HLE (with tools) | 57.4% |
| GDPval-AA v2.1 | 1620 |
The 57.4% HLE-with-tools score is notable for a small model — the benchmark probes graduate-level reasoning across science, medicine, and engineering. Adding tool access closes 11.5 percentage points of the gap over the no-tools baseline, suggesting the model is actively using retrieval and computation rather than just recognizing it has access to them.
Against GPT-6 Luna
Anthropic’s launch materials put two direct comparisons on record: Haiku 5.5 scores 39.2% on Terminal-Bench 4.0 versus GPT-6 Luna’s 16.4%, and 72.4% on OSWorld 2.1 (offline) versus GPT-6 Luna’s 48.9%. GPT-6 Luna is OpenAI’s lightweight tier equivalent. On both benchmarks, Haiku 5.5 outperforms it by 22+ points.
Effort Scaling
Haiku 5.5 is the first model in the Haiku tier to support an adjustable effort setting. The “(High)” designation in its Text Arena entry means evaluations are running at the high effort level — not the model’s minimum or default. Anthropic also offers low, medium, and extra-high tiers. The Terminal-Bench and HLE scores above are from the high effort setting unless otherwise specified in source materials.
Cost
Haiku 5.5 is 75% cheaper than Haiku 4.5 on average, per Anthropic’s launch announcement. That reduction, combined with the agentic capability jump, repositions the Haiku tier from latency-optimized chatbot work toward lightweight agent pipelines where cost-per-task matters more than raw intelligence.
GDPval-AA v2.1 of 1620 places it well below Fable 5.1 (1853) and Opus 5.5, but above older models that have since been deprecated. For tasks where Fable-class intelligence is unnecessary, the cost curve is now favorable enough that Haiku 5.5 warrants a direct evaluation rather than a default-to-larger-model choice.
Text Arena ELO results will be visible as the model accumulates head-to-head comparisons in the coming weeks.