Haiku 5.5 vs Opus 5.5: costs and task selection

Which tasks belong on each model?

Anthropic positions Haiku 5.5 for summaries, compaction, classification and bounded subagent work. Its launch guidance recommends larger models for complex agentic coding. The Opus 5.5 announcement focuses on long coding jobs and knowledge work. Those are provider claims, not independent measurements from this site.

For a support queue, try Haiku on category assignment and extracting order fields into a fixed schema. Define the acceptable labels, validate the fields and route uncertain cases for review. A fluent answer is a poor success criterion when the task is to extract an exact account number.

For a code migration spanning several services, evaluate Opus on the complete change: build success, behavior preservation and the review effort required. Counting generated lines or accepting a patch that merely compiles would miss the expensive failures. Keep focused subtasks separate if they can be verified without repeating the whole migration.

A mixed workflow can use Haiku to summarize documents or inspect individual files, then ask Opus to make the final decision. Budget every subtask, the handoff context and any escalation. Compressing away a constraint can make the final stage fail, so test the combined workflow against direct Opus calls.

Worked budget: 20K input, 2K output, 10K requests

Assume each request contains 20,000 fresh input tokens and produces 2,000 billable output tokens. Neither model uses caching. Ten thousand such requests produce:

Haiku's arithmetic is (20,000 × $0.10 + 2,000 × $0.50) ÷ 1,000,000 = $0.003. Opus Standard is (20,000 × $4 + 2,000 × $20) ÷ 1,000,000 = $0.12. These estimates exclude cache writes, Batch, paid tools, tax and residency premiums.

Open these exact inputs with Haiku and Opus selected. The calculator uses Standard rates. Its caching toggle models already-populated cache reads; it excludes writes and mode adjustments. You can switch back to all models or share the selected comparison.

Prompt length and caching change the comparison

Haiku's higher band begins at 100,001 prompt tokens and applies to the whole request, including output. Count fresh input, cache reads and cache creation together. Opus Standard keeps the same per-token rates throughout its 1M context window. See the official billing rules.

With 120,000 uncached input and 2,000 output tokens, Haiku costs $0.065 and Opus costs $0.52. At 100,000 input and the same output, Haiku costs $0.011; at 100,001 it costs $0.0550005. Leave room for history and tool definitions near the boundary.

If 18,000 of the original 20,000 input tokens are actual cache reads, the short request costs $0.00138 on Haiku or $0.0516 on Opus. Writing or refreshing the cache adds a separate charge. Read the Haiku 5.5 pricing guide for exact boundary and cache-write examples.

Batch, Fast and subscription billing

Eligible Batch requests halve token prices: the uncached example becomes $15 for Haiku or $600 for Opus across 10,000 requests. Opus Fast uses $8 input and $40 output per million tokens and cannot combine with Batch. US-only inference adds 10%; partner-cloud regional bills can differ. These settings are outside the linked calculator's estimate.

Use the API invoice's billable output, including thinking, rather than only the answer displayed to users. The model overviews list adaptive thinking for both models, always on for Opus. Claude subscription fees and included credits are separate from these API unit prices.

Choose by cost per accepted result

Take a representative sample from your own workload and set acceptance criteria before testing either model. Include routine cases and failures that would require a human to intervene. Record total spend over every attempt, accepted results and time to an acceptable result.

Calculate cost per accepted result = total workflow spend ÷ accepted results. If ten Haiku attempts cost $0.03 and one Opus attempt costs $0.12 at the example's fixed usage, Haiku still has a lower token bill. That arithmetic says nothing about whether ten attempts can solve the task or whether the delay is acceptable.

For a routing plan, add the failed Haiku attempt before each Opus escalation. A hypothetical 90% Haiku-only completion rate with 10% escalation costs $150 for 10,000 initial requests at the example's usage: $30 in Haiku calls plus $120 in Opus calls. That assumes one call per stage; it is a planning scenario, not an observed success rate.

Model identity and sources

The official overviews list claude-haiku-5-5 (released October 7, 2026) and claude-opus-5-5 (September 22, 2026) as Active. Both have 1M context and 128K standard output limits. We did not run a paid inference benchmark or carry over Haiku 4.5 results.

Continue with the two-model cost calculator, the full model catalog, or the Opus 5.5 pricing page. AI API Prices editors checked the prices and arithmetic on 2026-10-08. Report corrections to [email protected].