Anthropic launched Claude Haiku 5.5 on October 7, 2026, positioning it as the cheapest, fastest and most capable small model the company has ever shipped. The economics explain the fanfare: for prompts of up to 100,000 tokens — roughly 90% of the requests the Haiku line receives — input costs drop to $0.10 per million tokens and output to $0.50 per million tokens, about 90% below Haiku 4.5. Averaged across longer queries too, Anthropic puts the saving at around 75%. It is also the first Haiku to ship with adjustable effort settings, giving the compact tier the same operational dial as its larger siblings.
- Dramatic capability jump: Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset, against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna.
- Price parity, better scores: Anthropic matches OpenAI’s Luna pricing at $0.10/$0.50 per million input/output tokens for short prompts, but leads on every shared benchmark.
- Ecosystem sweeteners: Sonnet 5.5 cache reads are halved to $0.10 per million tokens, and Max and Team subscribers receive monthly API credits of up to $500.
From zero to competitive on agentic work
The most striking figure in Anthropic’s own results concerns the terminal. On Terminal-Bench 4.0, which measures complex command-line professional tasks, Haiku 4.5 scored exactly 0.0%. Haiku 5.5 reaches 39.2%. On FrontierCode 1.1, the agentic coding benchmark, it climbs to 46.4%, ahead of GPT-6 Luna’s 42.4%. On Humanity’s Last Exam, the maximum-difficulty question set assembled by researchers, the model lands at 45.9% without tools and 57.4% with tools.
Documentation fills in the rest of the specification sheet: a 1-million-token context window, up to 128,000 output tokens per answer, text and image inputs, and a knowledge cutoff of June 2026. Third-party evaluators reached with independent runs report a lower but still dramatic Terminal-Bench figure of around 33%, up from zero a year earlier, with GPT-6 Luna at 13%. Independent intelligence indices place Haiku 5.5 slightly ahead of GLM-5.3 Flash and Gemini 3.8 Flash, and comparable to the far larger open-weights model Kimi K3, while trailing Claude Sonnet 5.5. The same third-party testing flags a genuine weakness: Haiku 5.5 answered 36% of knowledge questions accurately with a 40% hallucination rate — lower accuracy than its rivals, but a far lower fabrication rate than GPT-6 Luna’s 77%.
“The cheapest, fastest, and most capable small model we’ve ever released.” — Anthropic
Tiered pricing and the tokenizer caveat
Anthropic’s headline rate applies only below the 100,000-token threshold. Above it, input rises to $0.50 and output to $2.50 per million tokens, with cache writes at $0.625 and cache reads at $0.05. GPT-6 Luna holds its $0.10 input rate up to 272,000 tokens before stepping up to $0.20, so teams pushing whole contracts or long transcripts through a compact model will find the advantage flipping toward OpenAI. A second caveat is the tokenizer: Haiku 5.5 shares the newer tokenizer used by Sonnet 5.5 and Opus 5.5, which counts roughly 30% more tokens for the same text — the reason Anthropic quotes a 75% average saving rather than 90%. On the baseline itself, the company’s launch page compares against Haiku 4.5 at $1/$5 per million tokens, while current pricing documentation lists Haiku 4.5 at $0.50/$2.50, so the precise discount depends on which reference point is used.
Effort control is the operational story. The default is medium; teams can drop to low for cost-sensitive bulk work or push to high and very high for reasoning-heavy jobs. Anthropic describes the canonical deployment as a larger model handling the main task while Haiku acts as a subagent for high-volume steps such as context compaction, automated summaries, database queries and classification.
What early adopters are reporting
Anthropic’s launch materials include customer figures that are more concrete than internal benchmarks. Box reports an 11-point score improvement over Haiku 4.5 at roughly half the latency in enterprise content analysis. Asana measured a reduction of more than 30% in task-completion latency and up to 2.5x inference throughput per agent turn for error triage and portfolio search. HubSpot averaged 92.8% across three runs of its simulated CRM portals — its best result on that test — with Haiku 5.5 also the fastest and the most accurate at flagging stale records. Cognition pairs it with Opus 5.5 in Devin Fusion, reporting a FrontierCode score of 66.2 for the combination.
Beyond Haiku: cheaper Sonnet and API credits
Two adjacent changes accompany the launch. Anthropic has halved Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens, which it says trims about 20% from the cost of typical agentic workloads. It is also introducing monthly Claude Platform API credits for subscribers: $100 (≈ €94) for Max 5x, $200 (≈ €188) for Max 20x and up to $500 (≈ €470) pooled per organisation for Team, usable across any model and in third-party harnesses. Computer-use and browser-use toolsets were added to the Python and TypeScript SDKs the same day, removing the need for developers to write the action loop themselves. Haiku 5.5 is available on Anthropic’s platform, AWS, Google Cloud and Microsoft Azure under the model identifier claude-haiku-5-5, with batch jobs at half price and retirement no earlier than October 7, 2027.
The question that will define Haiku 5.5 over the coming quarters is not benchmark performance but production durability — whether its agentic gains survive the messy variability of real enterprise data. There is a cost asterisk worth watching: at maximum effort the model consumes roughly 162,000 output tokens per index task, about three times GPT-6 Luna’s roughly 50,000, so verbosity rather than sticker price may determine the real bill. For teams already running a strong planner model with Haiku as the workhorse, though, this release rewrites the arithmetic of high-volume AI deployment.