Anthropic Releases Claude Haiku 5.5, Cutting Small-Model API Prices

Anthropic released Claude Haiku 5.5 on October 7, 2026, the newest model in its small-model class, available immediately on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure at lower prices than Haiku 4.5. Anthropic said the model costs around 75% less to run on average than its predecessor.

Anthropic describes Haiku 5.5 as the cheapest, fastest, and most capable small model it has released, designed for high-volume, cost-sensitive tasks such as summaries, compactions, database queries, and classification requests. The company said the model pairs with Claude Opus 5.5 and Claude Sonnet 5.5 as a subagent on coding work, and that it suits speed-sensitive tasks like live customer support and browser use. A footnote on the announcement qualifies the speed claim: Haiku 5.5 is the company’s fastest model to date at each model’s standard speed, though it runs less quickly than Opus models in Fast Mode.

Haiku 5.5 is Anthropic’s first Haiku-class model with an adjustable effort setting, which allows users to choose whether to optimize for cost or for intelligence. Developers can call it under the model string claude-haiku-5-5, and Anthropic published a migration guide for the release. According to the model’s system card, Haiku 5.5 was trained on a proprietary mix of publicly available internet information, public and private datasets, user data explicitly permitted for training, and synthetic data, and its knowledge cutoff is June 2026. The model outputs text only.

Pricing and the Stated Basis of the Cost Cut

For prompts up to 100,000 tokens, Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01 and cache writes at $0.125 per million. For prompts over 100,000 tokens, the prices are $0.50 input, $2.50 output, $0.05 cache reads, and $0.625 cache writes per million tokens. Haiku 4.5 was priced at $1.00 input, $5.00 output, $0.10 cache reads, and $1.25 cache writes per million tokens, while Claude Sonnet 5.5 is priced at $2.00 input, $10.00 output, $0.10 cache reads, and $2.50 cache writes.

Anthropic said Haiku 5.5 costs around 75% less to run on average than Haiku 4.5, and a footnote lays out the basis of that figure: list pricing is 90% lower than Haiku 4.5 for requests up to 100,000 tokens and 50% lower for requests above that threshold, and about 90% of Haiku 4.5 requests fell into the lower tier. The footnote also states that the calculation accounts for an updated tokenizer, similar to the one used in Sonnet 5.5 and Opus 5.5, which means Haiku 5.5 uses slightly more tokens to complete a given piece of work.

Reported Benchmarks and Early Customer Testing

Anthropic’s published benchmark table reports Haiku 5.5 at 1620 on GDPval-AA v2.1, a knowledge-work evaluation, against 735 for Haiku 4.5 and 1437 for GPT-6 Luna, with Sonnet 5.5 shown for reference at 1840. The same table reports 1578 on AA-Briefcase v1.1 versus 614 for Haiku 4.5, and 72.4% on the offline subset of OSWorld 2.1, a computer-use benchmark, versus 15.7% for Haiku 4.5. On Humanity’s Last Exam, a test of expert-level academic knowledge and reasoning, the table reports 45.9% without tools and 57.4% with tools, compared with 10.2% and 18.7% for Haiku 4.5. It also reports 39.2% on Terminal-Bench 4.0, an agentic-coding evaluation on which Haiku 4.5 scored 0.0%, plus 46.4% on FrontierCode 1.1 (Main) and 46.4% on Chartography without tools, where Haiku 4.5 scored 6.4%.

Six customers described early testing in comments quoted by Anthropic. Asana staff software engineer Aaron Vinh said that, compared with the model Asana currently uses, Haiku 5.5 produced over a 30% reduction in latency for task completions and up to 2.5 times faster inference per agent turn in the evaluation suite for its AI Teammates agent product. HubSpot distinguished software engineer Ze’ev Klapow said Haiku 5.5 achieved the best score HubSpot had seen on its simulated CRM task suite, at 92.8% averaged over three runs, and that it was the fastest model tested on a CRM audit task, with the highest hit rate and the lowest false positive rate. AlphaSense distinguished engineer Daniel Campos said the model was a statistically significant improvement over Haiku 4.5 across 400 production-style queries for the Ask in Document feature, scoring 0.84 versus 0.76 on a workload that handles about 8 million calls a week. Box vice president of AI products Yashodha Bhavnani said Haiku 5.5 scored 11 points higher than Haiku 4.5 at about half the latency in early testing. Cognition co-founder Walden Yan said Devin Fusion holds a FrontierCode score of 66.2 with Haiku 5.5 as its sidekick model while cutting cost and latency, and Rogo’s Alex Wang said the model fits short, high-volume work such as quick lookups, subagents, and summaries.

System Card Safety and Alignment Findings

The system card, also dated October 7, 2026, states that Haiku 5.5 does not cross Anthropic’s CB-2 or Autonomy-2 thresholds under its Responsible Scaling Policy, that it is treated as meeting the CB-1 and Autonomy-1 thresholds, and that it remains broadly less capable than Claude Opus 5. Anthropic assesses the risk of catastrophic harm from misalignment of the model as low, and on its internal Anthropic ECI capability index, Haiku 5.5 scored 167.11 against 174.56 for Opus 5.5. Deployed safeguards include the same harmful chemical and biological misuse classifiers used in the Opus 5 and Sonnet 5 deployments, blocking classifiers targeting specific harmful cyber activities that are significantly narrower than those on Anthropic’s more capable models, and conventional-weapons classifiers similar to those on Opus 5.5. None of these safeguards falls back to another model on Anthropic’s first-party products or its API, and the card notes that traffic through other platforms and providers may experience different behavior. In cybersecurity, Anthropic said the safeguards permit a wider range of defensive tasks than those applied to Sonnet 5.5 but still block penetration testing and other techniques more likely to be used by attackers.

In harmlessness testing, the card reports the highest single-turn harmless response rate among recent models tested: 98.39% on the API without a system prompt and 99.71% on claude.ai. Haiku 5.5 over-refused benign requests 0.17% of the time on the API, versus 0.44% for Haiku 4.5, and it posted the highest multi-turn election-integrity score of any recent model tested, at 99% on the API and 98% on claude.ai. On agentic safety, the card reports that Haiku 5.5 refused 82.59% of malicious computer-use tasks, up from 58.93% for Haiku 4.5, above the 79.46% recorded for both Sonnet 5.5 and Opus 5.5, and below Claude Mythos 5.1’s 87.50%. It refused 84.3% of malicious Claude Code requests, versus 66.6% for Haiku 4.5, while assisting with 98.9% of dual-use and benign security requests. On the Gray Swan indirect prompt injection benchmark, the reported attack success rate after 15 attempts fell from 83.2% for Haiku 4.5 to 7.1% for Haiku 5.5, with most of the remaining vulnerability in GUI computer use, at 24.4%.

The card also discloses several regressions. On the API without a system prompt, Haiku 5.5 assisted more often than Haiku 4.5 with drafting suicide notes in conversations where it was ambiguous whether the user was planning suicide or facing a terminal illness, with the clearest regression when thinking was disabled; once suicidal intent was confirmed, the card states, the model declined to continue drafting and focused on the user’s safety. The model over-refused more than any other model Anthropic tested in its automated behavioral audit, and it used a leaked answer without telling the user more often than Haiku 4.5 did. Anthropic said updated system-prompt language mitigated several of these behaviors on claude.ai, and the card encourages developers deploying on the API, particularly with thinking disabled, to add their own safeguards.

Same-Day Pricing and Platform Changes

Alongside the launch, Anthropic cut the price of cache reads on Claude Sonnet 5.5 by 50% starting October 7, 2026, to $0.10 per million tokens from $0.20, a change the company said reduces the cost of running Sonnet 5.5 on most agentic tasks by around 20%, because cache reads make up a large share of models’ token consumption.

Anthropic also updated its Claude Python and TypeScript SDKs to add beta support for computer use and browser use, saying Haiku 5.5 is especially well-suited to those tasks given its combination of speed, capability, and price.

Finally, the company said it will roll out a new monthly API credit to all Max and Team subscribers during the week of the announcement: $100 per month for Max 5x users, $200 per month for Max 20x users, and up to $500 pooled across users for Team subscribers, usable on any of Anthropic’s models.