Claude Haiku 5.5: Anthropic's Cheapest and Fastest Model Reshapes the Small-Model Race

Anthropic is living through the most consequential weeks in its history: according to QZ.com, the company plans to start its investor roadshow in the week of November 9 and list before Thanksgiving at a valuation between $1.8 and $2 trillion. Claude Haiku 5.5 landed right in the middle of that picture, released on Wednesday, October 7. For Anthropic , the timing is no accident: in an intensifying race with OpenAI, the company wants its smallest and fastest model to capture developers' everyday workloads.

The launch completes the 5.5 family trilogy: Opus 5.5 arrived on September 22, Sonnet 5.5 on September 28. As Decrypt summarized it, Haiku 5.5 is positioned as the cheapest and fastest model the company has ever shipped. The presenter makes no secret of his surprise that the refresh finally came: Haiku 4.5 had gone almost a year without an update and was lagging behind its rivals.

Pricing: a steep cut with a 100k-token threshold

The pricing is aggressive: $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. According to Anthropic's announcement, that list sits 90% below Haiku 4.5's $1 and $5 rates; above 100,000 tokens the discount narrows to 50% ($0.50 and $2.50). The company's blended figure puts running costs about 75% lower , since roughly 90% of the old model's requests stayed under the threshold. SimonWillison flags a subtle catch: the new tokenizer splits the same text into roughly 1.25x more tokens, so the real-world discount is softer than the list suggests.

The price lands exactly on top of OpenAI's small model GPT-6 Luna, released September 22 at the same $0.10/$0.50 rates. Under 100,000 tokens the two models cost the same; above the line the picture flips, because Luna holds its $0.10 and $0.50 rates up to 272,000 tokens and only then steps up to $0.20 and $0.75. The benchmark rows reported by Decrypt justify the money: Haiku 5.5 leads Luna on all six published tests. SimonWillison's conclusion is crisp: Haiku wins on short context, Luna on long context.

Benchmarks: a big leap for a small model

The table in the official announcement shows the scale of the jump: on OSWorld 2.1, which measures real computer use, Haiku 5.5 scores 72.4% while Haiku 4.5 sits at 15.7% and Luna at 48.9%. The agentic coding test Terminal-Bench 4.0 is starker still: 39.2% versus 0.0% , with Luna at 16.4%. Knowledge-work scores read 1620 versus 735 on GDPval-AA (Luna 1437), 46.4% versus 42.4% on FrontierCode, and 46.4% versus 6.4% on the visual-reasoning test Chartography (Luna 29.1%). The full compilation by Technology Org shows the one strong rival left out of the comparison, Sonnet 5.5 (70.6% on Terminal-Bench, for instance), still playing a league above.

Numbers from early-access customers back up the table: according to SiliconAngle, Asana saw task-completion latency in its agent product drop by more than 30% , with inference running up to 2.5x faster per agent turn. HubSpot recorded 92.8 on its CRM test suite, the best score it has seen on that battery. The Box report is more granular: 60 versus 49 on complex-work evaluation, about 27% fewer tokens, and roughly half the wall-clock time, with the gap widening to 19 points on report-drafting work that demands numerical derivation.

Adjustable effort and where to get it

Haiku 5.5 is the first Haiku model with an adjustable effort setting : developers pick low, medium, high, or maximum reasoning for the job at hand; reasoning cannot be switched off entirely, and the default is medium. Anthropic's published curves show accuracy climbing with cost as effort rises. SimonWillison's playful experiment makes the trade concrete: a pelican sketch at low effort took 7 seconds and cost 0.09 cents, while the maximum-effort version took 5 minutes 9 seconds and cost 3.38 cents.

Availability is broad: the model is live in the Claude chatbot, inside Claude Code, through the API, and on Amazon, Google, and Microsoft clouds, with the US government cloud on the list too. SiliconAngle notes two further discounts: Sonnet 5.5's cache-read price is halved from $0.20 to $0.10, taking roughly 20% off most agentic work on that model. Monthly API credits sweeten subscriptions as well: $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team plans. Decrypt points out the credits match the subscription fee and do not roll over.

Hands-on tests: from games to design

The presenter's own tests go beyond the numbers: the plain Mac clone Haiku 4.5 produced 11 months ago sits next to the Minecraft clone Haiku 5.5 generated, complete with cave systems and water physics. The most extreme trial yields a Call of Duty-style zombie game skeleton; an isometric room drawing earns praise for texture and lighting. The standout community example comes from a user named Theo, who built a dungeon crawler with shifting atmosphere in a single hour. Motion design impresses most: an animation that takes frontier models 40 minutes comes out of Haiku 5.5 in about 15 minutes .

The presenter's verdict sits squarely on price-performance: the model costs as little as Luna yet clearly outscores it, and it finishes the job faster in head-to-head runs against Grok. Near-Sonnet behavior on simple tasks makes it a natural subagent inside Claude Code: the large model draws the plan while the small model takes the repetitive subtasks. The daily-driver claim makes sense in that architecture — not a model without limits, but a cheap teammate that keeps working.