Claude Haiku 5.5: Count the cost of a finished task

Anthropic’s standard rates, checked 11 November 2026. The worked request costs $0.002 before tools, retries or taxes. Original data graphic by PING! Magazine. Pricing data: Anthropic.

The smallest number on an API price card is not the cost of getting a job done. Anthropic's 7 October launch of Claude Haiku 5.5 makes that distinction worth revisiting: the new small model is aimed at high-volume work, but the useful comparison is between acceptable completed tasks, not between isolated tokens. Launch and price card

The short version

• Read the prompt-length tier before estimating the bill.
• Count output, retries and human corrections as well as input.
• Do not turn a vendor benchmark into a result for your own business.

What the published numbers say

For prompts up to 100,000 tokens, Anthropic lists $0.10 per million input tokens and $0.50 per million output tokens. Above that prompt length, its card lists $0.50 input and $2.50 output. Anthropic counts all input towards the threshold, including cache reads and writes, which have separate rates. These are standard list rates in US dollars, rechecked on 11 November 2026. Check the price for your exact provider and account.

Worked example: 10,000 input tokens at the lower tier cost $0.001; 2,000 output tokens cost another $0.001. That makes $0.002 for the stated request—before tools, retries, taxes or other services. One million identical requests would make $2,000 under those simplified assumptions. This is arithmetic from the card, not measured operating spend.

Decide what a successful task means

Before choosing a model, write down the acceptance rule. A classification might need the correct category and valid output format. A customer-support draft might need a verified policy reference and no invented promise. A summary might need to preserve quantities and uncertainty. Define these checks before comparing the answers, rather than letting fluent wording become the pass condition.

Keep a small, representative evaluation set that includes awkward cases, missing information and work in the languages you actually serve. Record both the first answer and the correction required. A request that needs another call or an employee's intervention is not the same outcome as one that passes immediately. This is PING's proposed evaluation method, not a benchmark we have run.

Separate the routing decision from the marketing claim

Anthropic positions Haiku for narrower workloads and describes its larger models as better choices for complex agentic coding. That is a vendor description, not our comparative verdict. We are not claiming a universal latency, coding-accuracy advantage or winner over Gemini. Vendor's task distinction

A sensible trial starts with one bounded, reversible task and a clear fallback. Keep the previous path available while reviewing failures. Cheap calls should not gain extra account permissions simply because they can run more frequently. Production authority and model selection are different decisions.

Budget in the currency your team pays

For teams in London, Lagos or Nairobi, keep the dollar price, billing currency and conversion assumption separate. Ask who owns the budget when retries increase and where a human review is required. Do not infer local availability, data-location suitability or affordability from the global launch sentence alone.

Haiku 5.5 changes the price discussion. It does not remove the need to know what the system is doing. The useful number to carry into the next buying meeting is the total cost of a result your team would accept, accompanied by the checks that define acceptance. Find more documented guides on PING's Innovation desk.