Tool Suggestions
Claude Haiku 5.5 Pricing: The Cheapest Capable AI Model in 2026 — and the 100K Token Trap
Which cheap AI model should you actually build on in 2026? On October 7, Anthropic released Claude Haiku 5.5 — the first Haiku update in about a year — priced at $0.10 per million input tokens and $0.50 per million output tokens, exactly matching OpenAI's GPT-6 Luna. Anthropic calls it "the cheapest, fastest, and most capable small model we've ever released," and says it costs around 75% less to run than Haiku 4.5 on average. The developer response was immediate: the launch topped Hacker News with 938 points and 442 comments, and Cursor, GitHub Copilot and Devin all added the model on launch day. This article lays out the full pricing, the benchmark results, and the three costs that sit underneath the headline price.
Claude Haiku 5.5 pricing, in full
The headline rates apply to prompts up to 100,000 tokens: $0.10 input, $0.50 output per million tokens, inside a 1-million-token context window. The same launch also cut Sonnet 5.5's cache-read price by half, and introduced monthly API credits for Claude Max and Team subscribers, reported at $100 to $500 depending on plan. Haiku 5.5 is the first model in its tier with Anthropic's effort settings — Low through Max, with adaptive thinking — so the same model can be run cheap and shallow for sorting work, or harder for tasks that need judgement. That flexibility is the point of the release: one model string, claude-haiku-5-5, covering the high-volume jobs that used to force a choice between a weak cheap model and an expensive strong one.
What you get: benchmarks against GPT-6 Luna
Haiku 5.5 scores 43 on the Artificial Analysis Intelligence Index — 26 points above its predecessor a year ago, and in the range of frontier models from late 2024. On Anthropic's published comparisons it beats GPT-6 Luna on FrontierCode 1.1 (46.4% against 42.4%) and by a wide margin on OSWorld computer use (72.4% against 48.9%). Third-party signals are unusually strong for a launch week: Cognition reported 58.4% on FrontierCode 1.1 at roughly an eighth of Sonnet 5's cost per task, and GitHub reported Haiku 5.5 matching Sonnet 5 on many coding tasks while taking fewer steps. The intended roles are subagents — classification, extraction, summarising, context compaction — running under a larger model, and high-volume production traffic where the per-request price decides the margin.
The three costs hiding under the headline price
First, the 100K cliff: prompts above 100,000 tokens are billed at five times the headline rate. A workload that drifts over that line — a growing agent context, a long document pipeline — does not get slightly more expensive; it changes tier entirely, and several developers flagged exactly this in the launch discussion. Second, the tokenizer: Haiku 5.5's new tokenizer produces roughly 25–30% more tokens for the same text, so a per-token price cut is partly handed back on every request. Third, effort: at maximum effort the model generates substantially more output tokens — Artificial Analysis measured roughly three times GPT-6 Luna's output token counts — so "cost per token" and "cost per task" can move in opposite directions. None of these are reasons to avoid the model. All three are reasons to measure your real cost per completed task before migrating a workload on the strength of the $0.10 figure.
Haiku 5.5 vs GPT-6 Luna: which cheap model wins?
The two models are priced to the cent identically, so the decision rests on behaviour, not tariff. On the published evidence, Haiku 5.5 leads on coding benchmarks and computer use, and its effort settings give a control Luna's tier does not emphasise. Luna's counter-arguments are ecosystem and predictability: it sits inside the OpenAI stack many teams already run, and its pricing has no 100K step. The third option worth pricing before you commit is Inception's Mercury 3, whose launch discount undercuts both on output — covered in our Mercury 3 analysis. For most teams the practical answer is to route short, parallel, well-defined jobs to the cheap tier — whichever vendor — and keep a frontier model for the steps that fail cheaply nowhere.
The takeaway
Claude Haiku 5.5 is the strongest answer yet to the cheapest-capable-model question: frontier-2024-level ability, a 1M context, effort controls, and a $0.10 headline that survived its first week of independent testing. Its pricing is also genuinely three prices — the headline, the over-100K tier, and the effort-driven token bill — and the teams reporting the best results are the ones measuring per-task cost rather than per-token price. At these rates, the small-model tier is no longer a compromise engine. It is where most production AI work should probably start.