Anthropic, PBC released Claude Haiku 5.5 on October 7, 2026, a small model with a 1 million token context window priced at $0.10 per million input tokens and $0.50 per million output tokens [3][4]. That is one tenth of what Claude Haiku 4.5 charged and one twentieth of Claude Sonnet 5.5, for prompts up to 100,000 tokens. Haiku 5.5 is on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and every Claude plan. The cheap rate has a catch: a prompt over 100,000 tokens is billed at five times the base price.
- Claude Haiku 5.5 is Anthropic's fastest model and the first Haiku with adjustable effort, with a 1M-token context window.
- It costs $0.10 per million input tokens and $0.50 per million output tokens, rising fivefold for prompts over 100,000 tokens.
- Anthropic says it beats GPT-6 Luna on every shared benchmark, but the results are unverified and the API has breaking changes.
What is Claude Haiku 5.5 and what is new
Claude Haiku 5.5 (model id claude-haiku-5-5) is Anthropic's fastest current model, aimed at high-volume and latency-sensitive work such as classification, extraction and routing [3]. Anthropic also points it at live customer support and browser use [9]. It replaces Haiku 4.5 at the bottom of a lineup that now runs from Haiku 5.5 through Sonnet 5.5 at $2 and $10 to Opus 5.5 at $4 and $20 per million tokens [3].
The main change besides price is control. Haiku 5.5 is the first Haiku-class model with an adjustable effort setting, which runs from Low through Medium, High and Xhigh to Max, with adaptive thinking on by default and effort defaulting to medium [3][6][9]. Higher effort gives smarter answers and costs more to run. It takes text and images as input and produces text. Its reliable knowledge cutoff is June 2026 [3][6].
Anthropic frames the model as a worker rather than a planner. It recommends Sonnet 5.5 and Opus 5.5 for complex agentic coding, and its own comparison table has Sonnet 5.5 ahead of Haiku 5.5 on every row [6][9]. Depth Grid's Tech News Today, Oct 8 roundup covers the GPT-6 news from the same week.
Under the hood: context, output and the new tokenizer
Haiku 5.5 has a 1 million token context window and a maximum output of 128,000 tokens, or up to 300,000 tokens in the Batch API beta [3][6]. Computer use is supported through the SDK in beta, and Priority Tier is not supported [6][7]. Anthropic has not published a parameter count in the documentation I could read, and the system card [2] holds the safety evaluations.
"counts the same text as roughly 30% more tokens than Haiku 4.5"Anthropic, Claude Haiku 5.5 documentation, on the new tokenizer [6]
The tokenizer matters for the bill. The same text now measures about 30% longer, so a prompt of roughly 77,000 tokens on Haiku 4.5 can reach about 100,000 on Haiku 5.5 and cross the higher price tier sooner [7].
Benchmarks: Anthropic's table, with no outside check yet
Every score below was published by Anthropic, and no independent evaluator had confirmed them when this article was written [6][7]. The comparison set is Haiku 4.5, OpenAI's GPT-6 Luna and Sonnet 5.5.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna | Sonnet 5.5 |
|---|---|---|---|---|
| GDPval-AA v2.1 (Elo) | 1,620 | 735 | 1,437 | 1,840 |
| 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% |
| FrontierCode 1.1, Main | 46.4% | n/a | 42.4% | 52.1% (Xhigh) |
| Humanity's Last Exam, with tools | 57.4% | 18.7% | n/a | 64.5% |
| Chartography, no tools | 46.4% | 6.4% | 29.1% | 61.6% |
The jump over Haiku 4.5 is large. On Terminal-Bench 4.0 the old model scored 0.0% and the new one scores 39.2%, and on OSWorld 2.1 the offline subset rises from 15.7% to 72.4% [7]. Haiku 5.5 also beats GPT-6 Luna on every benchmark where both are reported, while Sonnet 5.5 beats Haiku 5.5 on all of them [7]. Without tools, Haiku 5.5 scores 45.9% on Humanity's Last Exam [6].
The comparison is selective. The New Stack notes that Anthropic's table leaves out cheaper rivals, and that GLM-5.3-Flash scores slightly above Haiku 5.5 on GDPval-AA v2.1 [7]. Treat the table as a statement of intent until outside evaluators publish their own runs.
Pricing, availability and how to access it
Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Cache reads cost $0.01, five-minute cache writes $0.125 and one-hour cache writes $0.20, while the Batch API halves the base rates to $0.05 and $0.25 [4]. Haiku 4.5 listed at $1 and $5, so the base price is 90% lower [7].
Above 100,000 tokens the rates are $0.50 input, $2.50 output and $0.05 for cache reads, with batch at $0.25 and $1.25 [4]. The rule is applied per request. Every input token counts toward the threshold, including cache reads and writes, and a request over the line is billed entirely at the higher rates even if most of the prompt is a cache hit [4]. Caching therefore saves less on long prompts than on short ones.
Developers can call the model through the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS using the id claude-haiku-5-5, and the Claude model documentation has the full spec sheet [3][7]. In the apps, Anthropic says every plan from the free tier to Enterprise can select the model on the web, iOS and Android [8]. The announcement itself is at anthropic.com [1].
Limitations, trade-offs and alternatives
Moving an existing app to Haiku 5.5 takes code changes. Non-default temperature, top_p or top_k values return an error, and so do assistant prefill, the older budget_tokens syntax and an older computer-use tool version. A new refusal stop reason has no server-side fallback [5][7]. Under standard settings the cybersecurity safeguards also block penetration testing, and wider access requires applying to Anthropic's verification programs [7].
Per-token price is not the whole cost. Thinking tokens are billed as output by default, retries add up, and the 30% tokenizer change plus the fivefold step above 100,000 tokens can raise the bill for long-prompt workloads [7]. A sensible pattern is to route short, simple calls to Haiku 5.5 and send hard agentic coding to Sonnet 5.5 or Opus 5.5, which is also Anthropic's own advice.
Cheaper models exist. Qwen3.7 Flash lists at 3 cents per million input tokens for prompts up to 32,000 tokens [7]. Among larger models, Meta's Muse Wonder matches the $4 and $20 price of Opus 5.5, as covered in Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks. Teams that need downloadable weights can wait for Mistral's open model, detailed in Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing, though it is far larger and priced at $1.36 and $4.18.
Sources
- Anthropic. "Claude Haiku 5.5," announcement, October 7, 2026. anthropic.com
- Anthropic. "Claude Haiku 5.5 System Card." anthropic.com
- Anthropic. "Models overview," Claude API documentation. platform.claude.com
- Anthropic. "Pricing," Claude API documentation. platform.claude.com
- Anthropic. "Claude Haiku 5.5 migration guide." platform.claude.com
- MarkTechPost. "Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input Tokens," October 7, 2026. marktechpost.com
- Let's Data Science. "Claude Haiku 5.5 cuts prices 90%, prompts over 100K pay 5x." letsdatascience.com
- Notebookcheck. "Claude Haiku 5.5: Anthropic's fastest model is also free to use." notebookcheck.net
- Android Authority. "Claude Haiku 5.5 launched."
Discussion (0)
Comments are disabled for this post.