Anthropic Prices Claude Haiku 5.5 as a Subagent Worker

Anthropic priced Claude Haiku 5.5 at $0.10 per million input tokens on October 7, 2026, matching OpenAI’s GPT-6 Luna on the short-prompt rate card. The company says the model costs around 75% less to run than Haiku 4.5 and is built to sit under Opus 5.5 and Sonnet 5.5. That is a worker SKU, sold with a Sonnet cache cut and new Max and Team API credits on the same day.

Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released.

On average, it costs around 75% less to run than Claude Haiku 4.5. pic.twitter.com/jm06cZJfkV

— Claude (@claudeai) October 7, 2026

Haiku 5.5 Is Priced to Run Under Sonnet

Anthropic’s launch note is blunt about the job. Claude Haiku 5.5 pairs as a subagent on coding work and takes the repetitive load that used to be too dear to fire at volume. It is the fastest Claude at each model’s standard speed, though Opus still wins in Fast Mode.

The pitch is a split stack: Sonnet or Opus plans the work, Haiku runs the loops. Rogo’s applied AI group already describes that split in production language, with a larger model building a deck while Haiku opens a 10-K. Cognition is putting the same idea into Devin Fusion, with Opus 5.5 as the lead and Haiku 5.5 as the sidekick.

JOBS ANTHROPIC LISTS FOR HAIKU 5.5

• Short text: Summaries, classification, extraction, routing, and database queries at high volume.
• Agent chores: Compaction, subagent calls, and other narrowly scoped steps inside a larger run.
• Live speed: Customer support and browser use, plus computer-use and browser-use beta in the Python and TypeScript SDKs.

Free, Pro, Max, Team, and Enterprise users can pick the model on Claude.ai, and it is in Claude Code. Developers call claude-haiku-5-5 on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure.

What $0.10 per Million Tokens Buys

The cheap band is for prompts up to 100,000 tokens, which Anthropic says covered around 90% of requests to Haiku 4.5. Input is $0.10 per million tokens and output is $0.50. Cache reads are $0.01 and five-minute cache writes are $0.125. Cross 100,000 tokens and every line in that table rises fivefold.

HAIKU 5.5 API RATES PER MILLION TOKENS

Those short-prompt input and output figures match GPT-6 Luna at the same list rates. OpenAI cut Luna from $0.20 and $1.20 when it launched the model on September 22, 2026. Anthropic’s footnote is the 90% list cut versus Haiku 4.5 for the cheap band, and a 50% list cut above 100,000 tokens. The 75% “less to run” figure is a different number: it folds in request mix and a new tokenizer.

AlphaSense distinguished engineer Daniel Campos said Ask in Document does about 8 million calls a week, and that on 400 queries Haiku 5.5 scored 0.84 against 0.76 for Haiku 4.5. That is the traffic this rate card is for. Batch processing takes another 50% off input and output.

The 100,000-Token Line Changes the Bill

Haiku 5.5 keeps a 1 million token context window and up to 128,000 output tokens, but it does not keep one price across that window. Other recent Claude models bill the same per-token rate at 9,000 tokens and at 900,000. Haiku 5.5 is the exception: a prompt over 100,000 tokens pays $0.50 input and $2.50 output.

The tokenizer moved too. Platform docs say Haiku 5.5 uses the newer tokenizer from Claude 4.7 onward, so the same text counts as about 30 percent more tokens than on Haiku 4.5. Anthropic already bakes that into the 75% average. A shop that still thinks in Haiku 4.5 token counts will see a fatter bill than the 90% headline.

Thinking is on by default and cannot be switched off. Effort defaults to medium, and a non-default temperature, top_p, or top_k returns a 400 error. That is why a matching rate card can still lose a long, fussy job. One public API test of a voxel scene billed $24.96 on Haiku 5.5 and $1.91 on GPT-6 Luna, even though both models list $0.10 and $0.50. Effort is now part of the price, not a quality toggle you ignore.

Computer Use Jumps, Coding Still Belongs to Sonnet

On Anthropic’s own board, Haiku 5.5 is a different animal from Haiku 4.5 on computer use and a much smaller step on hard agentic coding. OSWorld 2.1 (offline subset) is the computer-use test: Haiku 5.5 scores 72.4%, against 15.7% for Haiku 4.5, 48.9% for GPT-6 Luna, and 83.9% for Sonnet 5.5. Terminal-Bench 4.0, which grades multi-step command-line work, tells the other half: 39.2% for Haiku 5.5, 0.0% for Haiku 4.5, 16.4% for Luna, and 70.6% for Sonnet 5.5.

ANTHROPIC’S HAIKU 5.5 SCOREBOARD

Anthropic says Sonnet 5.5 and Opus 5.5 remain the better choice for the class of work Terminal-Bench 4.0 measures. Haiku 5.5 is for tasks that were too costly on older Claude SKUs: compaction, summarization, subagent steps. Cognition co-founder and CPO Walden Yan said Fusion with Haiku 5.5 as sidekick holds a FrontierCode score of 66.2, which is a product score, not the 46.4% Haiku posts alone on FrontierCode 1.1 Main.

Knowledge-work Elo-style scores sit in the same pattern. GDPval-AA v2.1 is 1620 for Haiku 5.5, 735 for Haiku 4.5, 1437 for Luna, and 1840 for Sonnet 5.5. AA-Briefcase v1.1 is 1578, 614, 1336, and 1824. Chartography, with no tools, is 46.4% against 6.4%, 29.1%, and 61.6%. Humanity’s Last Exam with tools is 57.4% for Haiku 5.5 and 64.5% for Sonnet 5.5. The small model closed a lot of ground. It did not take the hard coding bench.

A Subagent Pulls the 10-K Line

The customer notes Anthropic published are not “it writes better poems.” They are latency, hit rate, and who holds the expensive model. Asana ran Haiku 5.5 through evals for AI Teammates, covering bug triage, project setup, and portfolio search.

We’re very impressed with Claude Haiku 5.5, particularly its speed. We ran it through our eval suite for AI Teammates, our AI agent product, covering use cases like triaging bugs, setting up projects, and searching large portfolios to surface high-risk or overdue work. Compared with the model we use today, we saw over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn. It’s a noticeably snappier experience.

Aaron Vinh, Staff Software Engineer, Asana

HubSpot distinguished software engineer Ze’ev Klapow said Haiku 5.5 posted 92.8% averaged over three runs on a CRM suite of simulated portals, the best score that shop has seen on the set, and that it was fastest on a stale-record audit with the highest hit rate and the lowest false positive rate. Box VP of AI Products Yashodha Bhavnani said early tests scored 11 points higher than Haiku 4.5 at about half the latency.

The short and high-volume work is where Claude Haiku 5.5 fits for us, like quick lookups, subagents, and summaries. While a bigger model builds the deck, a Haiku 5.5 subagent goes into the 10-K and pulls the segment revenue line the deck needs. It’s accurate enough that we’d trust it there, and fast and cheap enough that we can run it a lot.

Alex Wang, Applied AI, Rogo

That 10-K line is the product. If the subagent is cheap enough, the planner can stay expensive.

Sonnet Cache Reads Fall to $0.10

Haiku 5.5 closes a 15-day run of Claude 5.5 SKUs: Opus 5.5 at $4 and $20 per million tokens on September 22, 2026, Sonnet 5.5 at $2 and $10 on September 28, then Haiku. Sonnet’s other rates did not move. Cache reads did. They now cost $0.10 per million tokens rather than $0.20, which Anthropic says cuts the cost of Sonnet 5.5 on most agentic tasks by around 20%, because agents keep replaying the same system prompt, tools, and history.

On the same announcement, Anthropic said it would roll out monthly API credits for Max plans and Team seats on the Claude Platform, usable on any of its models, to let subscribers try tools, apps, and agents that call the API.

MONTHLY CLAUDE PLATFORM API CREDITS

• Max 5x: $100 in credits each month.
• Max 20x: $200 in credits each month.
• Team: Up to $500 pooled across users on the plan.

The credits seed the habit of calling the API from a subscription that used to stop at chat. They do not change Terminal-Bench. A Max user can now afford to wire Haiku as the worker and keep Sonnet on the cache-cheaper planner path without opening a surprise invoice on day one.

Cyber Rules Tighten for a Faster Model

Haiku 5.5 is the first Haiku with an adjustable effort setting, and the safety package moved with the capability. Anthropic says the model shows fewer cases of misaligned behavior than Haiku 4.5 and less willingness to help with misuse; those alignment tests relative to Haiku 4.5 sit in the system card. Cyber safeguards are stricter than Haiku 4.5 and looser than Sonnet 5.5. Defensive work has more room. The model still blocks penetration testing and other attacker-side methods, a tighter gate after open-source pentest agents in card theft.

Biology safeguards match Sonnet 5, Sonnet 5.5, and Opus 5: research questions can pass, requests judged likely to cause harm cannot. Labs that need a wider cyber or biology range can apply to Anthropic’s Cyber Verification Program and Life Sciences Verification Program.

Claude Haiku 5.5 is live as claude-haiku-5-5, with computer use and browser use exposed in beta in the Python and TypeScript SDKs. That is the work this price is meant to buy, so long as the prompt stays in the cheap band and effort does not silently eat the discount.

Frequently Asked Questions

What Is the Claude Haiku 5.5 Model ID?

Developers call it claude-haiku-5-5 on the Claude Platform. Anthropic lists the model as active, with a retirement date no sooner than October 7, 2027, and says thinking blocks from Haiku 5.5 work only in the account that produced them, or in an account linked to it.

Does Claude Haiku 5.5 Accept Images?

Yes. The documented input is text and images, and the output is text only. The reliable knowledge cutoff and the training-data cutoff are both June 2026.

Can You Turn Off Thinking on Haiku 5.5?

No. Adaptive thinking is on by default, depth is steered with the effort parameter, and the documented default effort is medium. Omit temperature, top_p, and top_k; a non-default value for any of them returns a 400 error.

How Does the Batch API Change Haiku 5.5 Prices?

The Batch API takes 50% off input and output, so the cheap band is $0.05 and $0.25 per million tokens and the long-prompt band is $0.25 and $1.25. One-hour cache writes, which are separate from batch, are $0.20 per million tokens up to 100,000 prompt tokens and $1.00 above that. With the output-300k-2026-03-24 beta header, Message Batches can return up to 300,000 output tokens.

Does US-Only Inference Cost More on Haiku 5.5?

Yes. For Claude 4.6 and later, including Haiku 5.5, setting US-only inference applies a 1.1× multiplier across token price categories. Global routing, the default, uses the standard rates.

The rate card is the easy part. The bill still depends on prompt length, the heavier tokenizer, and how much thinking you leave switched on.