Claude Haiku 5.5 Ships at $0.10/$0.50, With a Pricing Cliff at 100K Tokens
Anthropic's new small model is far cheaper on paper. The tier cliff, tokenizer change, and API breaks decide what you actually save.
Jump to 6 sections
Claude Haiku 5.5 launched October 7, 2026, at $0.10 per million input tokens and $0.50 per million output for prompts up to 100K tokens. Above 100K, the whole request costs five times as much. Anthropic calls it about 75% cheaper to run than Haiku 4.5, but the tokenizer and the tier cliff change the math.
Small models do the unglamorous work: routing, classification, extraction, and the first pass of every agent. A new small model from Anthropic arrives nine days after Sonnet 5.5 and replaces Haiku 4.5. Model Drop tracks launches as they happen, and this brief covers what shipped, what the price table really says, and what breaks if you migrate.
It is written for developers who run Haiku-class models in production and need to decide whether to switch, and when.
What Anthropic Shipped With Haiku 5.5
According to ComputingForGeeks' launch breakdown, Haiku 5.5 replaces Haiku 4.5 as the small, fast tier of the Claude family. The API model ID is claude-haiku-5-5. Haiku 4.5 retires no sooner than October 15, 2026, and Haiku 5.5 retires no sooner than October 7, 2027.
The spec jump is large. The table below compares the two models using the figures in that breakdown.
| Spec | Haiku 5.5 | Haiku 4.5 |
|---|---|---|
| Context window | 1M tokens | 200K tokens |
| Max output | 128K tokens | 64K tokens |
| Knowledge cutoff | June 2026 | February 2025 |
| Thinking | Adaptive, on by default | Manual budget |
| Effort levels | Low to max, default medium | Not supported |
Haiku 5.5 is the first Haiku with an adjustable effort setting. Inputs are text and images, and output is text. It is available in the Claude apps, the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and GitHub Copilot, the same breakdown reports.
Model Drop has not run Haiku 5.5 independently. Where this brief cites tests, it names who ran them.
How Much Does Haiku 5.5 Actually Cost?
Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 100K tokens. Above 100K, the rates rise to $0.50 and $2.50 and apply to the entire request, not just the extra tokens.
A Braindetox analysis of the price table lays out the tiers. Cache reads cost $0.01 per million at the short tier and $0.05 above 100K. Haiku 4.5 listed at $1 input and $5 output. Sonnet 5.5 lists at $2 and $10.
That makes the headline cut 90% for short prompts and 50% for long ones. Anthropic's blended figure, about 75% lower, accounts for traffic mix and the tokenizer change. The analysis notes that the per-request sizes behind that figure are not published, so it cannot be reproduced.
The tier cliff deserves a worked example. A 120,000-token prompt with a 2,000-token reply costs about $0.065 at the long tier: $0.06 for input and $0.005 for output. The same request at the short rate would have cost about $0.013. One extra token past 100K multiplies the bill by five.
Haiku 5.5 is the only current Claude model with this long-context tier, the breakdown says. Teams that routinely send big documents should watch their prompt sizes closely. Our guide to prompt caching and inference costs shows how caching softens the bill for repeated prefixes.
Does the New Tokenizer Eat the Savings?
The new tokenizer eats part of the savings. Anthropic documents about 30% more tokens for the same text compared with Haiku 4.5, and the exact amount depends on the content. Braindetox measured 1.11 to 1.36 times on prose and code, and 1.39 times on one coding test.
The effect is more than a cost multiplier. It can push a prompt over the 100K line. A prompt of 80,000 tokens on Haiku 4.5 becomes about 104,000 tokens on Haiku 5.5. That crosses into the long tier, and the whole request is billed at $0.50 per million.
The math: on Haiku 4.5, 80,000 input tokens at $1.00 per million cost $0.08. On Haiku 5.5, 104,000 tokens at $0.50 per million cost about $0.052. The saving is about 35%, not 90%.
The fix is to recount your tokens on the new model and update every cost estimate and max_tokens limit. Do not carry over numbers from Haiku 4.5.
Effort level changes cost as well. In the ComputingForGeeks coding test of eight tasks, Haiku 5.5 passed all eight at low, medium, and high effort, costing $0.0034, $0.0043, and $0.0056 in total. At max effort it passed seven and cost $0.0664. The test was small and informal, but it shows the cost range inside one model.
What Do the Benchmark Numbers Say?
Anthropic's published numbers show Haiku 5.5 far ahead of Haiku 4.5 and ahead of GPT-6 Luna on every row where both have scores. They also show it well behind Sonnet 5.5. All figures below are Anthropic's, at max effort.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| OSWorld 2.1 (offline subset) | 72.4% | 15.7% | 48.9% | 83.9% |
| Terminal-Bench 4.0 | 39.2% | 0.0% | 16.4% | 70.6% |
| Humanity's Last Exam, no tools | 45.9% | 10.2% | n/a | 56.9% |
| SWE-bench Multilingual | 83.7% | 67.4% | n/a | 90.3% |
Treat the max-effort caveat seriously. Most production calls run at medium or lower. The same breakdown shows one score, GDPval-AA, dropping to 1277 at medium effort from 1620 at max.
Independent scoring is thinner. Artificial Analysis gave Haiku 5.5 an Intelligence Index score of 43, versus 41 for Gemini 3.8 Flash, according to the launch coverage. It also measured Terminal-Bench at 33%, below Anthropic's 39.2%. A gap of that size between vendor and independent numbers is normal, and it is why we wait for third-party data.
Cost per task is the other catch. At max effort, the Braindetox analysis finds Haiku 5.5 costs more per attempt than Haiku 4.5 on GDPval-AA and Humanity's Last Exam, even though its list price is lower. See our explainer on which AI benchmarks still mean anything for how to read these tables.
What Breaks if You Migrate From Haiku 4.5?
Several things break, and all of them fail loudly with an HTTP 400 error. Migration is a code change, not a model ID swap.
The ComputingForGeeks breakdown lists the main rejections:
- The temperature, top_p, and top_k sampling parameters are rejected.
- Manual thinking with a token budget is rejected, so use adaptive thinking with an effort setting.
- Assistant prefill is rejected.
- Effort set to max with thinking disabled is rejected.
- The older computer-use tool version is rejected in favor of a newer toolset.
One more trap: thinking tokens count toward max_tokens. In the breakdown, a hard prompt capped at 16,000 tokens returned only a thinking block and no text. Raise your limits when you turn thinking on.
Defaults can bite too. In Claude Code, Haiku 5.5 is the default haiku model on the Anthropic API, but Bedrock, Google Cloud, and Foundry still resolve haiku to Haiku 4.5 unless you pin the version. Pin explicitly if you want predictable behavior.
Should You Switch Now?
Switch if your traffic is short prompts, high volume, and tolerant of a smaller model. Wait if you rely on sampling parameters, send prompts near 100K tokens, or need strong shell and terminal work.
The safety notes in the system card, as summarized in the breakdown, are mixed. Prompt-injection attack success on Gray Swan coding scenarios fell from 58.4% to 0.08%. But Haiku 5.5 over-refused more than any other model tested in the automated audit, and its hallucination rate on AA-Omniscience was 40%, similar to Haiku 4.5.
That mix matters for production. Fewer successful injections help agents that read untrusted text. More refusals can break workflows that depend on the model answering.
A sensible rollout runs a shadow test: send a copy of live traffic to both models for a week, compare pass rate and cost per completed task, and only then move traffic. For context on how this launch fits the month, see our look at four labs, three price bands, and one pattern from September.
At Model Drop, our editors will revisit Haiku 5.5 once more independent data arrives. Your next step is to recount your prompts on the new tokenizer and see how many land above 100K.
