Claude Haiku 5.5 vs Sonnet 5.5: Coding, Cost and Speed

Use Claude Haiku 5.5 for high-volume, well-specified work: it costs a twentieth of Sonnet 5.5 for prompts up to 100K tokens and a quarter above that. Use Sonnet 5.5 for long, multi-step agentic coding, where Anthropic's Terminal-Bench 4.0 scores are 39.2% against 70.6%. On our six coding tasks on October 27, 2026, both passed every check.

Both models have a 1M-token context window, 128K output and the same published rate limits, so the choice comes down to price, coding depth and speed.

Haiku 5.5 vs Sonnet 5.5 at a glance

Anthropic released Haiku 5.5 on October 7, 2026 and Sonnet 5.5 on September 28, 2026. Both are on OmniaKey: Claude Haiku 5.5 and Claude Sonnet 5.5. Each model also has its own write-up, the Haiku 5.5 review and the Sonnet 5.5 review.

Anthropic list prices per million tokens, checked October 27, 2026.

Price: 20 times apart, until the prompt passes 100K tokens

Only Haiku 5.5 is priced by prompt length. A request whose prompt is over 100,000 tokens pays five times the input and output rates, and the prompt length counts cache reads and writes, so a long cached context pushes a request into the dearer band. Sonnet 5.5 charges one price at any length.

Cache reads are where the two come closest. Sonnet 5.5 charges 0.05 times its input price for them, $0.10 per million tokens, down from $0.20, while Haiku 5.5 charges the usual 0.1 times, $0.01. The Batch API halves input and output prices on both models, so the plain input and output ratios below hold there too.

Anthropic says prompts up to 100,000 tokens made up about 90% of requests to its previous Haiku model, which is the case the Haiku 5.5 tier is built around. A reader asked the same question in the Hacker News launch discussion: "why would I not pick Sonnet 5.5 instead of Haiku 5.5?" Five request shapes answer it.

Per 1,000 requests at list prices:

What the rows show:

• K1 and K2: plain input and output are 20 times dearer on Sonnet 5.5 while the prompt stays at or under 100K tokens. A classifier call and a single coding step land in the same band.
• K3: over 100K tokens Haiku's input and output prices go up five times, so the gap falls to 4 times.
• K4: cache reads are only 10 times dearer on Sonnet, $0.10 against $0.01, because Sonnet 5.5's cache reads cost 0.05 times its input price while Haiku's cost the usual 0.1 times. A cache-heavy agent turn is therefore about 15 times dearer.
• K5: cached tokens count toward the 100K line, so a long cached context pushes Haiku into the higher band and the gap falls to under 3 times.

Our own six tasks cost $0.0022 on Haiku 5.5 and $0.0467 on Sonnet 5.5 at list prices, because both models wrote about as many tokens. The Haiku 5.5 pricing breakdown has the cache writes and the Batch rows.

Coding: what Anthropic's benchmarks say

Anthropic's Haiku 5.5 announcement prints Sonnet 5.5 in the same comparison table, marked "For reference", so both columns come from one page dated October 7, 2026.

Anthropic prints no effort level next to either Terminal-Bench score, so read that row as each model's headline figure rather than a matched pair.

Anthropic's own guidance is plain: "Sonnet 5.5 and Opus 5.5 remain better choices for complex agentic coding tasks like those measured by Terminal-Bench 4.0." The same announcement puts Haiku 5.5 on narrowly scoped, high-volume work such as compaction and subagents, which matches the shape of the numbers above.

Neither launch page nor model page lists a SWE-bench Verified score, so the comparison rests on the rows above.

Our test: six coding tasks on both models

• Both passed all six tasks, every check on every task.
• Haiku 5.5 was faster on 4 of 6 tasks and in total, 38.6 s against 54.9 s. Sonnet 5.5 was slightly faster on the bug fix and the first SQL query.
• The six tasks cost $0.0022 on Haiku 5.5 against $0.0467 on Sonnet 5.5, about 21 times, because both models wrote about as many tokens: 4,031 against 4,237 output tokens in total.

What the test shows: on short, clearly specified, single-message coding and data tasks with exact checks, Haiku 5.5 at its default effort matched Sonnet 5.5 at its default effort, at about a twentieth of the cost. What it does not show: long multi-step agent work such as editing many files, running tools and recovering from errors, where Anthropic's Terminal-Bench 4.0 row is widest. Our tasks were small enough that both models passed, and one run each makes the times a single sample.

OmniaKey sells API access to both models, so we ran the test through it. Prices here are Anthropic's list prices, not ours.

Speed

Anthropic calls Haiku 5.5 its fastest model at each model's standard speed, and says Sonnet 5.5 generates output 30%+ faster than Sonnet 5. Artificial Analysis (third party) measured 234.4 and 141.3 output tokens per second for Haiku 5.5 and Sonnet 5.5 at max effort, which is not the API default. Our own numbers are whole requests, one run each: 38.6 s against 54.9 s over six tasks, with Haiku ahead on four.

The default effort differs too, medium on Haiku 5.5 against high on Sonnet 5.5. In our run, Sonnet 5.5 returned thinking blocks on four of the six tasks and Haiku 5.5 on two.

Which one for which job

For the hardest work Anthropic offers Opus 5.5 above both, and the Claude Opus 5.5 vs Sonnet 5.5 comparison shows what the next price step buys.

Switching between them

Both models take the same Messages API request, so switching is the model field.

• Set the effort yourself when you compare them. The API defaults differ, medium for Haiku 5.5 and high for Sonnet 5.5, and Anthropic's Sonnet 5.5 announcement adds: "In Claude Code and our apps, the default effort is set to Medium, while the Claude Platform defaults to High."
• Haiku 5.5 accepts thinking: {"type": "disabled"} at high effort or below, and returns a 400 error at xhigh or max.
• Sonnet 5.5's lowest thinking setting is between_tools, which turns off up-front thinking at high effort or below.
• Both 5.5 models use Anthropic's newer tokenizer, so the same text counts about 30% more tokens than on Sonnet 4.6 or Haiku 4.5.

curl https://api.omniakey.com/v1/messages \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model": "claude-haiku-5-5", "max_tokens": 1024, "messages": [{"role": "user", "content": "Classify this ticket as billing, bug or other: I was charged twice this month."}]}'

Change the model to claude-sonnet-5-5 and the same request runs on Sonnet 5.5.

export ANTHROPIC_BASE_URL="https://api.omniakey.com"
export ANTHROPIC_AUTH_TOKEN="your-omniakey-api-key"

Those two variables point Claude Code at OmniaKey, and our Claude Code setup guide covers the rest. The model pages are Claude Haiku 5.5 and Claude Sonnet 5.5.

Common questions

How much cheaper is Haiku 5.5 than Sonnet 5.5?

Its input and output prices are one twentieth of Sonnet 5.5 for prompts up to 100K tokens, and one quarter above that. Cache reads are 10 times cheaper, $0.01 against $0.10 per million tokens, and the gap on a cache-heavy agent turn works out at about 15 times.

Is Haiku 5.5 good enough for coding?

On six short coding and data tasks with exact checks, it passed everything Sonnet 5.5 passed, at about a twentieth of the cost. Anthropic's Terminal-Bench 4.0 figure for agentic coding is 39.2% against 70.6%, and Anthropic still points to Sonnet 5.5 for complex agentic coding.

Do Haiku 5.5 and Sonnet 5.5 have the same context window?

Yes: 1M tokens of context and 128K of output for both. On the Batch API either model goes to 300K output tokens with the output-300k-2026-03-24 beta header.

Which is faster, Haiku 5.5 or Sonnet 5.5?

Anthropic calls Haiku 5.5 its fastest model at standard speed, and Artificial Analysis (third party) measured 234.4 output tokens per second for Haiku 5.5 against 141.3 for Sonnet 5.5 at max effort. In our six tasks Haiku 5.5 finished four of them first and the whole set in 38.6 s against 54.9 s.

Is Haiku 5.5 a replacement for Sonnet 5?

That is a different pairing, and we compared Haiku 5.5 vs Sonnet 5 separately. Sonnet 5 is on Anthropic's legacy model list, while Sonnet 5.5 is current and the model compared above.

Sources

• Anthropic: Introducing Claude Haiku 5.5, Claude Haiku 5.5 model overview
• Anthropic: Introducing Claude Sonnet 5.5, Claude Sonnet 5.5 model overview
• Anthropic: models overview, pricing, effort, rate limits
• Artificial Analysis, third party: Claude Haiku 5.5, Claude Sonnet 5.5

All checked October 27, 2026.

Update log

• October 28, 2026: first version, with six coding tasks run on both models on October 27, 2026.