Claude Haiku 5.5 Cuts Prices and Adds API Credits

⚡

Key Takeaways

1Haiku 5.5 reduces prices by up to 90% for invites ≤ 100,000 tokens and shows significant progress in benchmarks.

2Anthropic cuts the cache of Sonnet 5.5 by 50% ($0.20 → $0.10/M tokens) and creates monthly API credits.

3Strengthened safeguards, blocked penetration tests, availability on AWS, Google Cloud, and Azure.

💡Why it matters — these cost reductions and API credits aim to make high-volume workloads viable while framing sensitive uses.

Anthropic launches Claude Haiku 5.5 with significantly reduced token pricing compared to Haiku 4.5 and marked improvements in benchmarks. At the same time, the publisher cuts the cache reading cost of Sonnet 5.5 by 50% and creates monthly API credits. The model is now available on major clouds and comes with enhanced cybersecurity safeguards.

Anthropic Cuts Sonnet 5.5 Cache by 50% and Credits API

Anthropic has reduced the cache reading cost of Sonnet 5.5 by 50%, from $0.20 to $0.10 per million tokens. According to the company, this measure is expected to decrease the cost of most agentic tasks by about 20%. Anthropic is also introducing monthly API credits: $100 for Max-5x subscribers, $200 for Max-20x, and up to $500 per month for Team, which can be used to test tools, applications, and agents via the API. This price reduction is presented as a response to OpenAI's GPT-6.1 series. These announcements complement the introduction of API credits for subscribers.

Haiku 5.5 Reduces Query Costs by Up to 90%

On average, Haiku 5.5 costs about 75% less than Haiku 4.5. For queries of 100,000 tokens or less, which make up nearly 90% of previous Haiku queries, prices can drop by up to 90%. Queries exceeding 100,000 tokens are charged at a rate five times higher. The model employs an updated tokenizer that requires slightly more tokens per operation, a situation already observed with Opus 4.x, with an increase of about 30% due to the tokenizer modification. Therefore, the actual savings may be lower than what the token prices suggest. Haiku 5.5 is designed for high-volume, cost-sensitive tasks such as summarization, database queries, classification, or live customer support.

Benchmarks: From 735 to 1,620 on GDPval, 72.4% on OSWorld

On the GDPval-AA v2.1 benchmark, Haiku 5.5 achieves a score of 1,620, compared to 735 for Haiku 4.5. For standalone computer use, the model scores 72.4% on OSWorld-2.1, up from 15.7% previously. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2%, while Haiku 4.5 had no score. In the Humanity's Last Exam, Haiku 5.5 scores 45.9% without tools and 57.4% with tools, compared to 10.2% and 18.7% for its predecessor.

Facing GPT-6 Luna and Sonnet 5.5, a Targeted Model

Anthropic compares Haiku 5.5 to OpenAI's GPT-6 Luna economic model and claims it outperforms it in all tested areas, including agentic coding. Benchmark scores, however, show that Haiku 5.5 still falls short of Sonnet 5.5, Anthropic's larger model. Haiku 5.5 introduces the ability to adjust reasoning levels within the Haiku range. According to Anthropic, this model is better suited for specialized uses such as compaction, synthesis, or sub-agent tasks, while Sonnet 5.5 and Opus 5.5 are recommended for complex agentic coding.

Clouds, Enhanced Security, and New SDKs in Beta

Haiku 5.5 is available on AWS, Google Cloud, and Azure. Compared to the previous version, cybersecurity measures have been strengthened and allow for a broader range of defensive tasks than Sonnet 5.5, partly due to the model's lower overall capacity. Penetration testing remains prohibited. Organizations with broader needs can request access to Anthropic's verification programs for life sciences and cybersecurity. Additionally, the Python and TypeScript SDKs have been updated with beta support for computer and browser usage.