Claude Haiku 5.5 Beats GPT-6 Luna on Computer Use — at a Fraction of the Cost
Anthropic released Claude Haiku 5.5 on October 7, calling it the cheapest, fastest, and most capable small model it has ever shipped. The benchmark numbers support that framing — but the more striking detail is what Haiku 5.5 does to GPT-6 Luna on agentic tasks.
On Terminal-Bench 4.0, Haiku 5.5 scores 39.2%. Luna scores 16.4%. Haiku 4.5 scored 0.0%. On OSWorld 2.1’s offline subset, Haiku 5.5 reaches 72.4%. Luna hits 48.9%. This is a small model comfortably above the midpoint between its predecessor and Anthropic’s own Sonnet 5.5.
The Pricing Shift
Anthropic says Haiku 5.5 runs about 75% cheaper on average than Haiku 4.5. The company notes that prompts under 100,000 tokens represent approximately 90% of requests to the previous model, making short-context tasks the sweet spot for cost savings.
Alongside the Haiku 5.5 launch, Anthropic is halving cache-read prices for Sonnet 5.5, which it says will reduce cost on most agentic workloads by around 20%. Claude Max and Team subscribers also receive a new monthly API credit.
Effort Levels Come to Haiku
This is the first Haiku-class model with an adjustable effort setting. Like Opus 5.5 and Sonnet 5.5, users can now slide Haiku 5.5 between cost-optimised and intelligence-optimised modes, trading tokens for accuracy. Anthropic published OSWorld 2.1 accuracy-versus-cost curves for each effort setting on its model page.
Benchmark Comparison
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 | 1620 | 735 | 1437 | 1840 |
| OSWorld 2.1 (offline) | 72.4% | 15.7% | 48.9% | 83.9% |
| HLE (no tools) | 45.9% | 10.2% | — | 56.9% |
| HLE (with tools) | 57.4% | 18.7% | — | 64.5% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| FrontierCode 1.1 Main | 46.4% | — | 42.4% | 52.1%* |
*Sonnet 5.5 at Xhigh effort.
The HLE-with-tools result deserves attention: Haiku 5.5 at 57.4% approaches Sonnet 5.5’s 64.5%, while GDPval-AA at 1620 versus Sonnet 5.5’s 1840 tells a similar story. For tasks where throughput and cost dominate, this is now an easy substitution.
What This Means for Agentic Pipelines
Haiku 4.5 was functionally useless on Terminal-Bench — 0% reflects an inability to complete even basic shell-based agentic tasks. Haiku 5.5 at 39.2% clears GPT-6 Luna (16.4%) and approaches the level of models that ranked as frontier on the same benchmark twelve months ago.
Asana reported a 30% reduction in task latency and up to 2.5x faster inference per agent turn in early testing. The combination of lower cost, higher speed, and credible agentic capability makes Haiku 5.5 a practical subagent candidate for pipelines currently running Sonnet-class models.
Claude Haiku 5.5 is available now through the Anthropic API.