Haiku 5.5 pricing and the real cost of your workflow

Recount the same text before comparing prices

Anthropic’s migration guide says the same input text produces approximately 30% more tokens on Haiku 5.5 than on Haiku 4.5. The exact increase depends on the content.

That means yesterday’s token counts are not a reliable input to tomorrow’s estimate. Count representative prompts with the new model selected, including instructions, conversation history, retrieved documents, and tool definitions.

Consider a deliberately simplified input-only example. A prompt counted as 60,000 tokens on Haiku 4.5 costs $0.060 at its published input rate. If the same text counts as 78,000 tokens on Haiku 5.5—assuming the documented approximate 30% increase—it costs $0.0078 at the lower tier.

That is an illustrative 87% reduction in the input component. It is not a measured workflow saving: output, caching, thinking, tool use, retries, and review are excluded. The assumed token increase also needs replacing with an actual count.

Near the 100,000-token boundary, recounting becomes particularly important. An 80,000-token input would become approximately 104,000 tokens under the same assumption. The lower-tier rate would no longer be the right basis for the estimate.

Budget for the work the model actually performs

The visible answer does not tell you everything about a request’s resource use.

Haiku 5.5 uses adaptive thinking by default. The migration guide says thinking tokens count toward the output limit, and a limit carried over from Haiku 4.5 can stop a response before it produces answer text. Effort settings therefore belong in the evaluation alongside model choice.

Record actual billed usage and completion status. A cheap request that returns an incomplete result can trigger a second attempt, an escalation, or manual repair. Those costs belong to the same customer task.

Caching also changes the comparison. Repeated context may qualify for cache reads, while new or changed context may require writes or ordinary input processing. Estimate each category separately from observed usage. Do not assume every repeated prompt receives the lowest cache rate. For broader optimization options, see ModelScale’s guide to lowering LLM API costs with caching, batch jobs, and output caps.

How Haiku compares with other cost-efficient models

Haiku 5.5 belongs on a shortlist alongside OpenAI’s GPT-6 Luna, DeepSeek V4.1 Flash, and Google’s Gemini 3.1 Flash-Lite. Open the four-model comparison in ModelScale to inspect their pricing, capability scores, context windows, and runtime evidence side by side. Flash-Lite is the budget variant used here; Gemini 3.1 Pro is listed separately at $2.00 input and $12.00 output per million tokens.

The comparison’s rates below were observed on 8 October 2026. Prices are US dollars per million tokens; the final column is a calculated example.

Example assumptions: each request bills 10,000 uncached input tokens and 2,000 output tokens, with one call per request. Cost equals 1,000 × [(10,000 ÷ 1,000,000 × input rate) + (2,000 ÷ 1,000,000 × output rate)]. These are equal-token estimates, not equal-text measurements. They exclude cache writes, tools, retries, fallback calls, and review; each model’s tokenizer and actual billed output must replace the assumptions for a workflow forecast. Haiku’s higher prompt-length tier also changes the comparison.

For short, uncached requests, Haiku and Luna therefore tie on these base rates. DeepSeek and Flash-Lite cost about 2.7 and 2.75 times as much in the example. Repeated context can change that ordering: DeepSeek’s listed cache-read rate is lower than either Haiku’s or Luna’s. Check cache eligibility and write charges before treating the read rate as a saving.

Capability and speed help decide which models deserve a trial. The comparison reports reasoning scores of 79.2 for Haiku and 79.6 for DeepSeek, while Luna has a reported coding score of 63.6 versus 56.5 for DeepSeek and 54 for Flash-Lite. Haiku’s coding score is unavailable. Flash-Lite’s reported throughput is 284 tokens per second versus DeepSeek’s 227, but those source measurements do not establish response times for your workload. Missing measurements remain unknown.

Use these figures to choose candidates, then measure cost per accepted result on your own cases. The comparison has no published pair-derived decision deltas or cost-per-successful-task values for this selection, so it does not establish a workflow winner. ModelScale’s comparison links each available figure to its pricing, benchmark, or runtime source.

Compare cost per accepted result

A useful operating measure is:

Cost per accepted result = total workflow cost ÷ number of results that meet your acceptance criteria.

Total workflow cost should include the initial model calls, retries, fallback calls, tool charges, and human review attributable to the evaluated tasks. Keep one-time migration work separate so you can see both ongoing savings and the time needed to recover that investment.

Define acceptance before testing. For an extraction task, that might mean required fields are correct and supported by the source. For a support workflow, it might require a correct answer, compliance with policy, and appropriate escalation.

Run the current and proposed configurations against the same representative cases. Include long inputs, ambiguous requests, missing information, and cases that previously required intervention. Track accepted results, total cost, latency, and escalation rate together. A cheaper configuration is useful only if it meets the product’s quality and response-time requirements.

Make the migration a measured decision

Start with one bounded, high-volume task. Establish its current cost and acceptance rate, recount its inputs for Haiku 5.5, and test an effort setting suited to the work.

Also check integration compatibility. The migration guide documents changes to thinking configuration, sampling parameters, assistant prefills, and response handling. A model-name swap alone can leave an application sending rejected requests or misreading responses.

Expand only after the comparison shows a worthwhile benefit. Anthropic positions Haiku 5.5 for narrowly scoped work and continues to recommend larger models for more complex agentic coding. That is a useful starting hypothesis for workload selection, not a substitute for your own evaluation.

Use ModelScale’s model workbench to build a shortlist, then test the task your users need completed. Before your next migration, take a sample of real requests and answer three questions: which pricing tier will they enter, how often will the results pass review, and what will each accepted result cost?