Anthropic Cuts Claude Haiku 5.5 API Costs Up to 90 Percent

Anthropic has repriced the bottom of its model stack, and the move lands squarely on the budgets of software teams that buy AI by the token. The company on October 7 released Claude Haiku 5.5, its smallest and cheapest model, cutting per-token API prices by 90 percent for requests under 100,000 tokens and by 50 percent above that threshold, with average running costs falling roughly 75 percent compared to Haiku 4.5 once a new tokenizer and a mix of request sizes are factored in2118.

The new rates — $0.10 per million input tokens and $0.50 per million output tokens for shorter prompts — put Haiku 5.5 at a tenth of Haiku 4.5's $1/$5 rate card, and match OpenAI's GPT-6 Luna at the same price point1122. Cache reads drop to $0.01 per million tokens in the lower tier, and the Batch API takes a further 50 percent off input and output charges221. For any engineering organization whose monthly AI invoice is dominated by classification, extraction, summarization, and other high-volume work, this is one of the steepest list-price cuts in the current model market.

The tiered pricing that vendors don't put in the headline

The launch's fine print matters enormously for licensing cost planning. Haiku 5.5 is priced in two tiers by prompt length: prompts of 100,000 tokens or fewer get the advertised rates, while longer prompts pay $0.50 input and $2.50 output per million tokens — a fivefold jump on input the moment a request crosses the line146. Because Anthropic's pricing documentation lists two full price rows rather than a surcharge on the excess, the reasonable reading is that the entire request moves to the higher rate once it exceeds 100,000 tokens6.

Anthropic says roughly 90 percent of requests to Haiku 4.5 fell under the threshold, so most historical usage patterns will land in the cheap tier2117. But there is a catch that several analysts flagged: Haiku 5.5 uses an updated tokenizer — the mechanism that splits text into billable units — that converts the same text into roughly 30 percent more tokens than Haiku 4.5's tokenizer did. Independent calculations suggest the real-world saving for identical text is closer to 87 percent below the line, and that long-prompt workloads see something closer to 35 percent in practice rather than the 50 percent on paper67. That gap between the headline discount and the effective discount is precisely the kind of detail that should appear in any build-versus-buy or vendor-switching analysis.

The competitive framing sharpens the licensing question. OpenAI's GPT-6 Luna charges the same $0.10/$0.50 rates, but its premium tier only kicks in above 272,000 input tokens, far higher than Haiku's 100,000-token boundary. Simon Willison's assessment is blunt: if your workloads stay short, Haiku matches Luna on price while posting higher benchmark scores, but above the threshold Luna is the better deal2211. Per-token price alone also can't settle which model is cheaper to complete a job — token consumption per task, retry rates, and accuracy all enter the total cost of ownership11.

Subscription customers get API credits they never had before

The more structural change for software subscription customers arrived alongside the model launch. Anthropic is rolling out a monthly API credit for Claude Max and Team subscribers: Max 5x users get $100 per month, Max 20x users get $200, and Team subscriptions receive up to $500 pooled across their users, spendable on any Claude model through the Anthropic platform2125.

This blurs a line that has defined the industry's commercial model since the ChatGPT era — the split between seat-based consumer subscriptions and metered API billing. Willison notes that the credits roughly match the cost of the subscriptions themselves, calling the scheme "really generous," and observes that claiming them is as simple as selecting an API organization in the billing settings. One operational caveat: the credits do not roll over month to month — unused value is forfeited22. The design also lets subscribers disable auto-reload so API requests simply stop when the balance runs out, avoiding surprise overage bills — a budget-control feature that matters to small teams without dedicated finance oversight22.

For Team-plan customers in particular, a pooled $500 monthly credit turns an existing subscription into a development subsidy for building agents and applications on the Claude Platform, which is exactly how Anthropic frames it2123. VentureBeat's coverage adds that the company has not publicly detailed how Team allocations are divided among users or what rollover conditions apply beyond the launch materials11.

Sonnet cache cuts compound the savings on agentic work

The second pricing change may be the quiet winner for software customers. Anthropic halved the price of cache reads on Claude Sonnet 5.5 — its $2/$10 mid-tier model and the recommended default for most production workloads — from $0.20 to $0.10 per million tokens218. Because cached tokens account for a large share of what agents consume in agentic pipelines, Anthropic estimates this single change makes Sonnet 5.5 roughly 20 percent cheaper on most agentic work2118.

For agent architectures that repeatedly send the same system prompts and document context, prompt caching is where the economics live. Haiku 5.5's own cache-read price of $0.01 per million tokens means that even the five-minute cache write premium breaks even after roughly two reads, making caching effectively free for any high-volume pipeline2. The combined effect is that Anthropic cut prices across two tiers of its stack simultaneously — the cheap model and the cache economics of the mid-tier workhorse — in a single announcement15.

Where the model fits in a licensing strategy

Anthropic is explicit that Haiku 5.5 is not aimed at complex coding or reasoning work; Sonnet 5.5 and Opus 5.5 remain the recommended models for those tasks, with Haiku positioned as a subagent handling summarization, compaction, classification, and quick lookups within larger pipelines19. It is also the first Haiku-class model with an adjustable effort setting — low through max — letting a single API contract trade cost against intelligence per request rather than forcing a model-tier switch2113. That flexibility has direct licensing implications: one model ID can serve both cheap bulk jobs and more demanding ones, simplifying vendor management.

Early customer metrics quoted in the announcement are vendor-selected but concrete. AlphaSense reported a statistically significant quality improvement over Haiku 4.5 across 400 production-style queries on a document question-answering workload that runs about 8 million calls a week — precisely the kind of high-volume feature where a 75 percent unit-cost drop transforms a cost center. Asana reported over 30 percent latency reduction and up to 2.5x faster inference per agent turn, HubSpot reported the best score it had seen on its simulated CRM suite at 92.8 percent, and Box cited an 11-point improvement at about half the latency11.

On benchmarks, Anthropic reports Haiku 5.5 ahead of GPT-6 Luna on all six tests where both have scores, including 72.4 percent against Luna's 48.9 percent on OSWorld and 39.2 percent against 16.4 percent on Terminal-Bench 4.0 — though these are vendor-reported figures, not independent verification11. Notably, Haiku 5.5's public SWE-bench number had not been confirmed in the coverage, meaning rivals can still dispute raw capability claims even where price is settled17.

The strategic read: usage growth before a listing

The timing context is hard to ignore. Haiku 5.5 completes a trio of 5.5-family launches inside a single month, following Opus 5.5 and Sonnet 5.5, and the releases come as Anthropic prepares for an anticipated initial public offering while managing enormous infrastructure expenses — one cited report put 2025 revenue at $4.6 billion against $42 billion in losses1415. Aggressive price cuts on the volume tier are the classic lever for driving usage growth ahead of a public debut, and matching OpenAI's Luna rate card point-for-point is a competitive pricing response, not a unilateral discount1522.

For software buyers, the practical takeaway is threefold. First, if your Haiku workloads stay under 100,000 tokens — as roughly nine in ten historically have — your effective unit cost just fell by about three-quarters, and batch-mode jobs cost half again that212. Second, if you're a Max or Team subscriber, there is now free API spend arriving monthly that did not exist last week; claim it before it expires22. Third, before migrating anything, measure your real prompts against the new tokenizer, because a workload that quietly sits at 105,000 tokens pays five times the advertised input rate, and that is the number that belongs in your licensing budget, not the one on the launch page62.