Claude Haiku 5.5 Matches Luna and Wins on Computer Use

Anthropic released Claude Haiku 5.5 on October 7, 2026, at $0.10 per million input tokens, matching OpenAI’s GPT-6 Luna on the public sticker. Output is $0.50 per million tokens on prompts up to 100,000 tokens, and the model id is claude-haiku-5-5.

Anthropic says the new small model costs around 75% less to run than Haiku 4.5. That average already folds in a heavier tokenizer and a fivefold jump once a prompt clears 100,000 tokens.

Haiku 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.

Read more: https://t.co/InzZXcEcRP

— Claude (@claudeai) October 7, 2026

Haiku 5.5 Matches GPT-6 Luna Dollar for Dollar

OpenAI put GPT-6 Luna at $0.10 and $0.50 on September 22, 2026, after reducing API prices for Sol and Luna by 50% from GPT-5.6. Fifteen days later Anthropic printed the same two numbers on Haiku 5.5 for prompts up to 100,000 tokens.

The benches Anthropic chose for the launch are not a tie. On computer use, Haiku 5.5 scores 72.4% on the OSWorld 2.1 offline subset, against 48.9% for Luna and 15.7% for Haiku 4.5. On Terminal-Bench 4.0, it scores 39.2%, against 16.4% for Luna and 0.0% for Haiku 4.5.

HAIKU 5.5 VERSUS LUNA AND HAIKU 4.5

Sonnet 5.5 still sits above the new Haiku on the same sheet, at 83.9% on OSWorld 2.1 and 70.6% on Terminal-Bench 4.0. Anthropic says those larger models remain the better pick for hard agentic coding, and that Haiku 5.5 is for narrower jobs that used to be too dear to run at volume.

Why the 75% Savings Is Not a 90% Cut

The short-prompt sticker is a 90% cut from Haiku 4.5’s $1.00 and $5.00. Prompts over 100,000 tokens pay $0.50 and $2.50, which is 50% below Haiku 4.5, not 90%. Anthropic says about 90% of requests to the previous Haiku fell in the cheap bucket, and that the blended saving is around 75 percent less to run.

The other haircut is hiding in the tokenizer. Platform docs say Haiku 5.5 uses the newer tokenizer from Claude 4.7 and later, so the same text counts as approximately 30 percent more tokens than on Haiku 4.5. A naive swap of the old bill onto the new rates will overstate the cash you keep.

PRICE PER MILLION TOKENS

That over-100k step is a fivefold jump on Haiku 5.5’s own low tier, and it bills on the prompt length of the request. Teams that stuff long histories, full repos, or uncompacted browser traces into one call will watch the cheap headline vanish. Cache reads at $0.01 still give high-volume loops a way to stay in the low band if the prompt is stable.

Context also grew. Haiku 5.5 takes 1 million tokens in and up to 128,000 tokens out, against 200,000 and 64,000 on Haiku 4.5. The window is large. The cheap price is not, once you cross 100,000.

THE RUN-UP TO HAIKU 5.5

1. October 15, 2025: Anthropic ships Claude Haiku 4.5 at $1.00 and $5.00 per million tokens.
2. September 22, 2026: OpenAI prices GPT-6 Luna at $0.10 and $0.50 per million tokens.
3. October 7, 2026: Anthropic ships Haiku 5.5 at that sticker and cuts Sonnet 5.5 cache reads in half.
4. October 8, 2026: Monthly API credits become available on Max and Team plans.

Haiku 5.5 is also the first Haiku with an effort setting. Adaptive thinking is on by default, and the default effort is medium, so a classification call and a desktop task do not have to spend the same thinking budget.

Computer Use Jumps From 15.7% to 72.4%

The OSWorld figure is a partial-credit score on an offline subset, not a claim that the model finishes every desktop task. Even on that yardstick, 72.4% against 15.7% is the gap that turns Haiku from a summarizer into something you can point at a browser.

Anthropic is shipping computer use and browser use in beta in its Python and TypeScript SDKs, and it names live support and browser agents as the speed case. A footnote on the launch page adds a limit: Haiku 5.5 is the fastest Claude at each model’s standard speed, and it still trails Opus models in Fast Mode.

Claude Haiku 5.5 on Amazon Bedrock is live through US, EU, AU, JP, and Global cross-region profiles, with model id anthropic.claude-haiku-5-5. Google Cloud and Microsoft Foundry use claude-haiku-5-5. Free, Pro, Max, Team, and Enterprise users can pick it in the Claude apps, and it is in Claude Code.

Cyber safeguards are tighter than Haiku 4.5’s and looser than Sonnet 5.5’s. Anthropic says they still block penetration testing and other attacker-side techniques, while leaving more room for defensive work. Biology rules match Sonnet 5, Sonnet 5.5, and Opus 5. The system card says the model does not cross new Responsible Scaling Policy thresholds and stays broadly less capable than Claude Opus 5.

A Subagent That Pulls Segment Revenue From a 10-K

The launch quotes are not about chat. They are about routing cheap calls under a larger model, or replacing a slower small one in a product that already runs millions of times a week.

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’s Ze’ev Klapow, a distinguished software engineer, said Haiku 5.5 scored 92.8% averaged over three runs on an internal CRM suite built around simulated portals, the best score the company had seen on that set, and that it was the fastest on a stale-record audit with the highest hit rate and the lowest false-positive rate.

Ask in Document is one of our big sources of spend, doing about 8M calls a week in production. It answers very specific questions on top of one or a few documents. We ran 400 queries, and Claude Haiku 5.5 was a statistically significant improvement over Haiku 4.5: 0.84 vs. 0.76.

Daniel Campos, Distinguished Engineer, AlphaSense

Box’s Yashodha Bhavnani, VP of AI Products, said early tests scored 11 points higher than Haiku 4.5 at about half the latency, for analytical work at scale. Rogo’s Alex Wang described the split that Anthropic wants: a bigger model builds the deck, and a Haiku 5.5 subagent goes into the 10-K and pulls the segment revenue line. Cognition’s Walden Yan, co-founder and CPO, said Haiku 5.5 joins Devin Fusion as a sidekick and that Fusion holds a FrontierCode score of 66.2 in that setup, which is a harness result, not the 46.4% Haiku posts on FrontierCode 1.1 alone.

Cache Reads on Sonnet 5.5 Fall by Half

The same morning, Anthropic cut Claude Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens. It says cache reads are a large share of token use on agent jobs, and that the change makes Sonnet 5.5 about 20% cheaper on most of that work.

Sonnet 5.5 still costs $2.00 input and $10.00 output. The cache cut is the lever for loops that reread the same repo, ticket, or tool schema. Haiku 5.5’s own cache read is $0.01 on short prompts, so a stack that uses Haiku for compaction and Sonnet for the hard turn now has a lower bill on both sides of the handoff.

Humanity’s Last Exam still shows the gap those larger models keep. Haiku 5.5 scores 45.9% with no tools and 57.4% with tools. Sonnet 5.5 scores 56.9% and 64.5% on the same two settings. Luna has no score on that row in Anthropic’s table.

Who Gets the Monthly API Credits?

On October 8, 2026, Anthropic’s developer account said the new monthly API credits for Max and Team were available. They work on any Claude model, including Haiku 5.5, in a team’s own code or in third-party harnesses. That last clause puts the credits inside tools Anthropic does not own, so a Max subscriber can burn the grant in the editor they already use, then keep paying on the same API rail once the grant is gone.

MONTHLY CREDIT BY PLAN

• Max 5x: $100 in Claude Platform credits each month.
• Max 20x: $200 in credits each month.
• Team pool: $20 per Standard seat and $100 per Premium seat, capped at $500 and shared.

Free, Pro, and Enterprise plans are out. Credits cover the Claude API, the Console playground, Managed Agents, and the Agent SDK. They do not cover interactive Claude Code sessions, extra usage in the Claude apps, or models reached through other clouds. New subscribers wait seven days on an eligible plan, then link a Console organization in billing settings.

Haiku 5.5 is live as claude-haiku-5-5, with a migration guide for teams leaving Haiku 4.5. Anthropic says it will not retire the model before October 7, 2027. The cheap line holds if you keep prompts under 100,000 tokens and remeasure token counts on the new tokenizer. Cross that line, and Luna’s matching sticker is no longer the whole comparison.

Frequently Asked Questions

What Is the Claude Haiku 5.5 Model ID on Each Cloud?

The Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS all use claude-haiku-5-5. Amazon Bedrock uses anthropic.claude-haiku-5-5, and AWS sample calls also show a global.anthropic.claude-haiku-5-5 inference profile on bedrock-runtime.

When Does Anthropic Plan to Retire Claude Haiku 5.5?

Platform docs list the model as Active (latest) and set retirement as not sooner than October 7, 2027, a one-year floor from the October 7, 2026 release rather than a date on the calendar.

Can You Set Temperature on Claude Haiku 5.5?

No. Docs say to omit temperature, top_p, and top_k, because a non-default value for any of them returns a 400 error. Steer depth with the effort parameter instead, which defaults to medium.

Who Is Eligible for the Monthly Claude API Credits?

Only active Max and Team plans in good standing. A Team with three Standard seats and two Premium seats receives $260 that month, because 3 × $20 plus 2 × $100 equals $260, then the pool stops at $500. Owners claim the balance; no Platform payment card is required for that grant.

What Is Claude Haiku 5.5’s Knowledge Cutoff?

The system card and the model overview both set the reliable knowledge cutoff at June 2026, which is also the training-data cutoff listed in the specs table. Batch API calls take 50% off input and output, and a one-hour cache write is $0.20 per million tokens on prompts up to 100,000 tokens, or $1.00 above that line.