Claude Haiku 5.5: What Is New and What It Means for Business

Anthropic has launched Claude Haiku 5.5, the small model in its 5.5 family, built for high-volume, cost-sensitive work. It is available today from $0.10 per million input tokens, and Anthropic says it costs around 75% less to run on average than Haiku 4.5.

The release completes the 5.5 generation, following Claude Opus 5.5 on 22 September and Sonnet 5.5 on 28 September. There was no Haiku in the Claude 5 generation, so for organisations running small-model workloads this is the first new Haiku in roughly a year, and the jump from Haiku 4.5 is a large one.

Haiku 5.5 is not being switched on by default across Owlpen at launch. More on that below. The rest of this article covers what has changed, how the new pricing works in practice, and what we would check before moving production work onto it.

What has changed from Haiku 4.5

Anthropic positions Haiku 5.5 for quick, repetitive workloads: summaries, conversation compaction, database queries, and classification, as well as customer support, browser automation, and work as a subagent beneath a larger model. The changes cluster around price, capacity, and capability.

A two-tier price based on prompt length

For prompts up to 100,000 tokens, Haiku 5.5 is listed at $0.10 per million input tokens and $0.50 per million output tokens, against $1 and $5 for Haiku 4.5. For prompts over 100,000 tokens, the rates rise to $0.50 and $2.50. Anthropic describes this as 90% lower than Haiku 4.5 for shorter requests and 50% lower for longer ones.

A larger context window and longer output

The context window grows from 200,000 to 1 million tokens, and maximum output from 64,000 to 128,000 tokens. The knowledge cutoff is June 2026, the same as the rest of the current Claude lineup.

Adaptive thinking and effort settings

Haiku 5.5 decides for itself when and how much to reason, steered by an effort setting that defaults to medium. Higher effort generally buys better answers at the cost of speed and tokens, so effort becomes a cost lever in its own right alongside the rate card.

Published benchmark results

Anthropic's figures show a large step up on every benchmark it reports. On OSWorld 2.1, which tests computer use over long multi-step tasks, it reports 72.4% against 15.7% for Haiku 4.5 and 83.9% for Sonnet 5.5. On GDPval-AA v2.1, which covers professional work across 44 occupations, it reports 1620 Elo against 735 and 1840. On Terminal-Bench 4.0 for command-line work, Haiku 5.5 scores 39.2% against 70.6% for Sonnet 5.5, which is a useful reminder of where the larger models still lead.

Who can use it

Haiku 5.5 is available through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Anthropic commits not to retire it before 7 October 2027 on its own platforms. Alongside the launch, Anthropic is rolling out a monthly API credit for Max and Team subscribers this week ($100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team), intended for building tools and agents against the API. It has also cut the price of cache reads on Sonnet 5.5 by half, to $0.10 per million tokens.

What it means in practice

The pricing threshold is the detail that matters most. Most classification, extraction, and routing calls are short, so they will sit under 100,000 tokens and see the deepest reduction. Long-document work will not. A pipeline that feeds whole contracts or case files into each request pays the higher tier, so the saving there is closer to half than to 90%.

The tokenizer also changes the arithmetic. Anthropic's documentation states that the same text counts as roughly 30% more tokens on Haiku 5.5 than on Haiku 4.5, which is part of why the average saving Anthropic quotes is around 75% rather than the 90% cut in the headline rate. Any budget, rate limit, or prompt-length check written in tokens for Haiku 4.5 should be recounted before migration, and a prompt that sat just under 100,000 tokens before may now sit above it.

There are code changes too. Haiku 5.5 rejects non-default temperature settings, pre-filled assistant responses, and the older manual thinking budget, and responses can now begin with a thinking block, so code that reads the first block as the answer needs updating. For agentic workflows that pair a large model with small helpers, Haiku 5.5 is a strong candidate for the helper role, but Anthropic itself says Sonnet 5.5 and Opus 5.5 remain the better choice for complex agentic coding.

Finally, cheaper inference tends to increase usage. A 90% lower unit rate can make previously uneconomic volumes viable, which is the point, but total spend still needs per-workflow limits rather than an assumption that a cheaper model means a smaller bill.

Safety, safeguards, and stated limits

Anthropic reports improved alignment results for Haiku 5.5 and has published a system card. Its cybersecurity safeguards are more restrictive than those on Haiku 4.5, though Anthropic says they are somewhat less restrictive than on its other recent models, and its biology safeguards match those on Sonnet 5, Sonnet 5.5, and Opus 5. A new safety classifier can also decline a request outright, and Anthropic's documentation notes that there is no server-side fallback, so applications need to handle that response.

For organisations with AI governance obligations, the practical point is that high-volume automation multiplies the effect of any error. A small model running thousands of classifications a day needs sampling, review, and escalation paths, because the risk of hallucination is reduced by better models, not removed by them.

Owlpen and Claude Haiku 5.5

Claude models are routed into configured Owlpen workflows through Anthropic API access, as described when Claude Sonnet 5 became available. Haiku 5.5 is not being enabled by default across Owlpen at launch. Coaley Peak can assess it for eligible client deployments where Anthropic model access is approved, and where the cost and latency gains hold on the client's own material.

Where clients ask us to review routing, the questions we would expect to answer are which steps fall under the 100,000 token threshold, whether accuracy on a representative sample matches the model it would replace, how the new tokenizer affects existing limits, and whether data handling terms are unchanged for the relevant deployment channel.

Sources used

This article is based on Anthropic's Claude Haiku 5.5 launch post for the cost claim, benchmark figures, safeguards, and subscriber credits, and on Anthropic's model documentation and release notes for the rate card, context and output limits, platform availability, tokenizer change, and breaking changes. All figures are those reported by Anthropic unless stated otherwise.

If you would like to discuss model selection, AI cost control, or the Owlpen platform in the context of your own workloads, contact us at enquiries@coaleypeak.co.uk or read more about the Owlpen platform.