Haiku 5.5 vs Sonnet 5
Choose by task difficulty and the cost of a verified result.
Haiku 5.5 vs Sonnet 5 is a choice about task scope and the cost of getting a correct result. Start with Haiku 5.5 for classification, extraction, summaries and tightly scoped coding steps. Use Sonnet 5 as the starting candidate for sustained planning, debugging and changes across several files. These are routing recommendations to test, not a claim that either model always wins.
Both offer a 1M-token context window and 128K synchronous output. The big pricing difference is conditional: Sonnet 5's direct input and output rates are 20 times Haiku's at up to 100K prompt tokens, but four times Haiku's above 100K. That distinction matters when a short request grows into an agent conversation.
Fact-checked October 9, 2026. This is a comparison of Haiku 5.5 and Sonnet 5, not Sonnet 5.5. It uses official documentation and disclosed calculations; we did not run a paid head-to-head benchmark. Anthropic still lists Sonnet 5 as active, under legacy models, and recommends considering Sonnet 5.5 for new migrations.
Haiku 5.5 vs Sonnet 5 at a glance
| Specification | Haiku 5.5 | Sonnet 5 |
|---|---|---|
| Claude API model ID | claude-haiku-5-5 | claude-sonnet-5 |
| Context window | 1M tokens | 1M tokens |
| Maximum synchronous output | 128K tokens | 128K tokens |
| Input → output | Text and images → text | Text and images → text |
| Thinking | Adaptive | Adaptive |
| Default API effort | medium | high |
| Current lifecycle label | Active, latest Haiku | Active, legacy Sonnet |
Sources: Anthropic's Haiku 5.5 overview and Sonnet 5 overview. Batch output limits are a separate contract. Equal context capacity does not establish equal reasoning quality or reliable recall across a million tokens.
Pricing: the 100K threshold changes the comparison
All rates below are USD per million tokens on Anthropic's standard direct API, checked on the fact-check date.
| Billable category | Haiku 5.5: prompt ≤100K | Haiku 5.5: prompt >100K | Sonnet 5 |
|---|---|---|---|
| Uncached input | $0.10 | $0.50 | $2.00 |
| Output | $0.50 | $2.50 | $10.00 |
| Cache read | $0.01 | $0.05 | $0.20 |
The official pricing documentation counts all input tokens, including cache reads and writes, toward Haiku's 100K threshold. A request above that threshold uses the higher rates for that request; only charging the excess tokens at the higher rate would understate the bill. Earlier requests keep their original charges.
Sonnet 5 retains its standard rates across its full context window. Its cache-read rate is $0.20, not the $0.10 announced for Sonnet 5.5. Cache creation, Batch discounts, tools, taxes and regional charges require their own calculation. Our Haiku 5.5 API pricing guide covers those details.
What the same token budget costs
For an uncached request, calculate input tokens ÷ 1,000,000 × input rate + output tokens ÷ 1,000,000 × output rate.
| Illustrative request | Haiku 5.5 | Sonnet 5 |
|---|---|---|
| 50K input + 5K output | $0.0075 | $0.15 |
| 100K input + 20K output | $0.02 | $0.40 |
| 200K input + 20K output | $0.15 | $0.60 |
These are calculations, not observed workloads. For the last row, Haiku costs 0.2 × 0.50 + 0.02 × 2.50 = $0.15. Both its input and output use the higher band. The examples exclude caching, Batch, tools, retries and taxes, and assume equal billed tokens; real runs may use different amounts.
Routing can change the result again. Suppose 1,000 tasks each consume 100K input and 20K output. One Sonnet attempt per task costs $400. If one Haiku attempt passes validation on 80% of tasks and the remaining 200 each need one successful, separate Sonnet attempt with that same token budget, the total is $20 + $80 = $100. This hypothetical 75% saving assumes the validator catches failures and excludes its cost, extra context and human review. It is not a measured success rate or savings promise.
Coding: choose the scope before the model
Anthropic's Haiku announcement emphasizes high-volume, bounded work and coding subagents. Its Sonnet 5 announcement describes planning, tool use and autonomous work. Those roles suggest this evaluation order:
| Workload | First candidate | Acceptance check |
|---|---|---|
| Labels, structured extraction, document summaries | Haiku 5.5 | A fixed rubric, required fields and source accuracy |
| Test triage or one mechanical edit | Haiku 5.5 | Reproduction, passing tests and a bounded diff |
| A feature or bug spanning several files | Sonnet 5 | Correct behavior plus integration checks |
| Ambiguous architecture or a costly-to-reverse change | Sonnet 5, with human review | Explicit constraints and an independently reviewed plan |
The last row is an evaluation starting point, not certification for high-risk work. A useful boundary is whether a cheap, reliable check can reject a bad answer. Haiku is easier to trial when success is explicit. When the model must first discover what “correct” means, judge the whole plan and correction effort before selecting the cheaper route.
Do not reuse Sonnet 5.5 benchmark scores as Sonnet 5 scores. Haiku's launch comparison includes Sonnet 5.5, Haiku 4.5 and GPT-6 Luna. It does not establish a controlled Haiku 5.5 versus Sonnet 5 result. Our Haiku 5.5 review explains the published evaluations without inventing a winner for this pair.
Speed: a faster model can still take longer to finish
Anthropic calls Haiku 5.5 its fastest model at standard speed, explicitly excluding Opus Fast Mode from that claim. That supports trying it for interactive or high-volume steps; it does not provide a verified “twice as fast as Sonnet 5” number.
Measure time to first token, full response time and time to an accepted result separately. A quick incorrect patch followed by two retries can take longer than one slower correct patch. Provider route, queueing, prompt size, tool calls and effort also matter. Because default effort is medium on Haiku and high on Sonnet, record defaults as defaults; compare equal explicit effort in a separate run rather than silently mixing the two experiments.
Switching in Claude Code or an API client
Pin claude-haiku-5-5 or claude-sonnet-5 when evaluating. A moving alias can change the model under the comparison. Start fresh sessions from the same repository commit, with the same task, tools, permissions, acceptance checks and retry budget.
Both models use adaptive thinking. Set output_config.effort deliberately, omit legacy sampling parameters such as temperature, top_p and top_k, and parse response blocks by type rather than assuming the first block is text. The Haiku migration guide documents the client changes. Record billed tokens from each run instead of assuming identical text always means identical usage.
For OmniaKey, check the current model catalog and gateway prices before budgeting, then follow the Claude Code setup guide. The direct Anthropic prices above are not OmniaKey quotes. Catalog availability does not replace testing your intended route.
A practical selection rule
- Define what passes before making either call: a schema, answer rubric, test suite or reviewed change.
- Trial Haiku on bounded work and Sonnet on work needing sustained planning. Record failures as well as successes.
- Track total spend divided by accepted tasks, plus completion time and human correction minutes. Escalate failed Haiku tasks using an explicit rule.
- Recheck costs when prompt history crosses 100K. Consider Sonnet 5.5 separately if choosing a current Sonnet deployment.
Frequently asked questions
Is Haiku 5.5 better than Sonnet 5 for coding?
There is no universal winner established by the sources used here. Haiku is an attractive candidate for bounded, verifiable coding steps; Sonnet 5 is a reasonable starting point for sustained planning and debugging. Test both against the same acceptance criteria.
Is Haiku 5.5 really 20 times cheaper?
Its direct input and output rates are one twentieth of Sonnet 5's for prompts up to 100K tokens. Above 100K, they are one quarter. Completed-task savings depend on token use, retries, cache behavior and review costs.
Is Sonnet 5 the same as Sonnet 5.5?
No. They have different model IDs and release evidence. Sonnet 5 remains active as a legacy model. Check the exact version before using a benchmark score, cache price or migration recommendation.