TECHNOLOGY · AI · CYBERSECURITY · THE SYSTEMS SHAPING BUSINESS
Anthropic released Claude Haiku 5.5 Wednesday, pricing its new small model up to 90% below Haiku 4.5 and giving a Haiku model an adjustable effort setting for the first time, VentureBeat reported. The company also halved cache-read prices for Sonnet 5.5, the midsize model it shipped Sept. 28.
What Happened
Haiku 5.5 uses two price tiers split at 100,000 tokens. For prompts under that line, it costs $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.5. Above the threshold, the rates rise to $0.50 and $2.50, half of what Haiku 4.5 charged. Cache reads cost $0.01 per million tokens and cache writes $0.125 in the lower tier, and $0.05 and $0.625 above it.
Anthropic estimates that about 90% of Haiku 4.5 requests fall below 100,000 tokens. After accounting for a new tokenizer, it puts the average saving on existing workloads at roughly 75%.
RelatedThe model ships with five effort levels, from low through medium, high and xhigh to max, with medium as the default. The setting changes both cost and results. On Terminal-Bench 4.0, Haiku 5.5 scores about 20% at medium effort and 39.2% at max, according to Anthropic's figures as reported by VentureBeat.
Other vendor-reported results show a large jump over Haiku 4.5: 72.4% on the OSWorld 2.1 computer-use benchmark, up from 15.7%, and an Elo score of 1,620 on GDPval-AA v2.1, up from 735. The model is available under the API identifier claude-haiku-5-5 on Anthropic's Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. Claude Max and Team subscribers also get new monthly API credits, ranging from $100 for Max 5x plans to as much as $500 shared across a Team subscription.
Why It Matters
The lower-tier price is identical to that of OpenAI's GPT-6 Luna, which makes small models a straight price fight between the two labs. The tiers are not the same, though: Luna's long-context surcharge begins above 272,000 input tokens, while Haiku's starts at 100,000, so jobs between those sizes cost more on Anthropic's model. On benchmarks Anthropic chose to publish, Haiku 5.5 leads Luna, scoring 72.4% to Luna's 48.9% on OSWorld 2.1 and 39.2% to 16.4% on Terminal-Bench 4.0.
Anthropic is aiming Haiku 5.5 at high-volume work such as summaries, classification, database queries, live customer support, browser automation and subagent tasks handed off by larger models. In agent systems that split one job across many smaller calls, the per-token price of the helper model drives most of the bill. Alex Wang, who works on applied AI at Rogo, told VentureBeat the model is "fast and cheap enough that we can run it a lot."
Early customers reported gains in their own tests. Box said it saw an 11-point improvement at about half the latency of Haiku 4.5, Asana said task-completion latency improved by more than 30%, and HubSpot reported a 92.8% average on a CRM evaluation.
Sonnet Cache Cut and Safeguards
The Sonnet 5.5 change takes cache reads from $0.20 to $0.10 per million tokens. Anthropic says that trims the cost of a typical agentic workload by about 20%, because such workloads reread the same cached context many times.
On security, Anthropic says Haiku 5.5's cyber safeguards are stricter than Haiku 4.5's but less restrictive than those on its other recent models, allowing a wider range of defensive tasks than Sonnet 5.5 while still blocking penetration testing. Organizations can apply for broader access through Anthropic's verification programs.
What to Watch
Haiku 5.5 completes Anthropic's 5.5 family after Opus 5.5 on Sept. 22 and Sonnet 5.5 on Sept. 28. All of the benchmark figures so far come from Anthropic, so independent tests against GPT-6 Luna are the next check on its claims. Developers will also be watching how much running at high or max effort eats into the headline savings, and whether OpenAI or Google answers with price cuts of its own.
