Anthropic launched Claude Haiku 5.5 on October 7, 2026, completing the company’s Claude 5.5 generation roughly two weeks after the original rollout began. The release follows Claude Opus 5.5 on September 22 and Claude Sonnet 5.5 on September 28, positioning the smallest and most economical model in the family as a high-volume workhorse for production agentic workloads. Anthropic is marketing the launch as the most aggressive pricing reset in the Haiku line’s history, betting that low per-token economics will pull workloads previously handled by larger competitors onto its infrastructure ahead of a potential initial public offering.
Claude Haiku 5.5: Pricing: A 90% Cut for the Bulk of Haiku Workloads
For prompts under 100,000 tokens, Claude Haiku 5.5 is priced at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic says that threshold covers roughly 90% of the historical Haiku workload on its platform. Against Haiku 4.5’s previous rates of $1 per million input and $5 per million output, the new list price amounts to a 90% reduction for the high-volume short-prompt segment. When accounting for the model’s new tokenizer, which packs information more densely, Anthropic states the average effective cost reduction is closer to 75%. For longer prompts exceeding 100,000 tokens, customers receive a 50% discount relative to Haiku 4.5. In a parallel move, Anthropic trimmed Sonnet 5.5’s cache-read price by 50%, to $0.10 per million tokens, sharpening the company’s competitive position on retrieval-heavy agentic systems.
Adjustable Reasoning Effort Debuts on Haiku
Claude Haiku 5.5 is the first Haiku-tier model to ship with an adjustable effort setting, giving developers a runtime dial that trades latency for cost by tuning reasoning depth. The control allows teams to keep Haiku on low-effort settings for routine routing, classification, and tool-calling tasks while escalating to higher-effort reasoning only on the prompts that benefit from it. Anthropic framed the feature as a response to enterprise buyers who want fine-grained control over compute spend without spinning up a separate model for each tier of difficulty. The setting arrives as rivals including OpenAI and Google have also begun exposing reasoning budgets as a first-class parameter in their mid-tier models.
Benchmark Lead Against GPT-6 Luna
Anthropic published benchmark numbers placing Claude Haiku 5.5 well ahead of OpenAI’s GPT-6 Luna on two agent evaluations. On OSWorld 2.1, which measures computer-use agents operating a desktop environment, Haiku 5.5 scored 72.4% versus Luna’s 48.9%. On Terminal-Bench 4.0, an agentic coding benchmark, Haiku 5.5 reached 39.2% against Luna’s 16.4%. Haiku 4.5, by comparison, scored 0% on Terminal-Bench 4.0, illustrating how much the agentic ceiling has moved between generations. Anthropic has not released comparable scores against Google’s Gemini 3 family or xAI’s Grok 4, leaving the broader competitive picture incomplete for now.
Enterprise Pilots Report Latency and Throughput Gains
Three enterprise customers — Asana, Box, and HubSpot — tested Claude Haiku 5.5 during the pre-release window. Asana, the first to publish operational numbers, said task-completion latency for its internal agents dropped by more than 30%, while per-agent-turn inference ran up to 2.5x faster than on its previous Haiku deployment. Box and HubSpot reported qualitative improvements in tool-use reliability and lower infrastructure overhead, though neither disclosed specific throughput figures. Anthropic is leaning on these reference customers to seed the rollout, betting that documented latency wins in production agent stacks will drive faster adoption than headline benchmarks alone.
Profitability, Compute Mix, and IPO Timing
The launch lands against an unusually busy financial backdrop for Anthropic. The company closed a Series H round that brought cumulative funding to $65 billion at a $965 billion valuation, and reporting has suggested Anthropic is preparing for a landmark IPO as early as next month. The Financial Times has cited figures showing Anthropic was profitable in the second quarter with gross margins above 80% on annualized revenue of approximately $11.5 billion. A separate SemiAnalysis estimate noted that subscription products consume more than 40% of Anthropic inference compute while generating only around 10% of revenue, highlighting why low-priced, high-volume API tiers like Claude Haiku 5.5 matter strategically. To soften the impact on individual builders, Anthropic said Max and Team subscribers will receive $100 to $500 in monthly API credits to experiment with the new model.
Cybersecurity Stance and What to Watch Next
Claude Haiku 5.5 ships with expanded cybersecurity safeguards that allow more defensive security work than Sonnet 5.5 permits, though offensive penetration testing remains blocked. Anthropic simultaneously expanded its Cyber Verification Program to cover a broader set of defensive use cases, opening a path for security vendors to onboard the new model under vetted terms. For now, analysts are watching three signals — sustained agent-benchmark dominance against GPT-6 and Gemini 3, the rate of Haiku 5.5 migration among existing Claude customers, and the timing of any S-1 filing from Anthropic — to gauge whether the pricing reset translates into the kind of API revenue growth that would justify the company’s pre-IPO valuation. Claude Haiku 5.5 is the clearest expression yet of Anthropic’s strategy to monetize the long tail of inference traffic at near-cost economics.
Source: Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read prices via CoinCustard, 2026-10-10 evening.