Haiku 5.5 costs $0.10 per million input tokens up to 100K, one tenth of Haiku 4.5

First seen on X 1 hour ago@claudeai ♥ 32,238

Anthropic launched Claude Haiku 5.5 on October 7. Haiku 4.5 cost $1 per million input tokens and $5 output. For prompts up to 100K tokens, Haiku 5.5 charges $0.10 input and $0.50 output, one tenth of the old price on both.

Prompts over 100K tokens rise to $0.50 input and $2.50 output. The company says average run cost is about 75% lower, and notes that a new tokenizer uses slightly more tokens per task, which is included in that figure.

More from today

5 stories

Haiku 5.5 beats GPT-6 Luna on OSWorld, 72.4% to 48.9%
First seen on X 1 hour ago@claudeai ♥ 32,238
In Anthropic's own table, Haiku 5.5 scored 72.4% on OSWorld 2.1 (offline subset). GPT-6 Luna scored 48.9%, and Haiku 4.5 scored 15.7%. On the Terminal-Bench 4.0 coding test, Haiku 5.5 scored 39.2%, ahead of Luna's 16.4%.
Sonnet 5.5 is still higher, at 83.9% on OSWorld and 70.6% on Terminal-Bench. Every score here is the company's own number.
Further reading
- Introducing Claude Haiku 5.5anthropic.com ↗
1 source
- @claudeaiX♥ 32K
Ten Haiku 5.5 subagents under Opus 5.5 finished in the same 13 seconds for $0.04, against $0.05 for Opus alone
First seen on X 1 hour ago@ClaudeDevs ♥ 8,451
This is an egg-drop test published by Anthropic's developer account. The task is to design a carrier from household items that keeps an egg intact when dropped from height.
Opus 5.5 alone produced 3 designs in 13 seconds for $0.05. Opus 5.5 orchestrating ten Haiku 5.5 subagents produced 5 designs in 13 seconds for $0.04. The company says the combination is faster and cheaper. It is one demo on one task, so check it on your own work before relying on it.
Further reading
- ClaudeDevs: Haiku 5.5 is now available in the Claude Platform and Claude Codex.com ↗
1 source
- @ClaudeDevsX♥ 8.5K
Summaries, conversation compaction and database queries are the jobs to give Haiku 5.5
First seen on X 1 hour ago@ClaudeDevs ♥ 8,451
Anthropic's developer account recommends running Haiku 5.5 as a subagent under Opus 5.5 or Sonnet 5.5. It suggests high-volume, cost-sensitive work such as summaries, compaction and database queries.
Boris Cherny of Claude Code said that with a 100K-token context, Haiku 5.5 is 10 times cheaper than Haiku 4.5. For long material, splitting it to stay under 100K tokens per call keeps the bill down.
Further reading
- bcherny: Haiku 5.5 is a good Haikux.com ↗
- ClaudeDevs: Haiku 5.5 is now available in the Claude Platform and Claude Codex.com ↗
2 sources
- @bchernyX♥ 1.4K
- @ClaudeDevsX♥ 8.5K
Haiku 5.5 is on the Claude Platform, Claude Code, three major clouds and OpenCode
First seen on X 1 hour ago@ClaudeDevs ♥ 8,451
The model ID is claude-haiku-5-5. Per the launch post it is available on the Claude Platform and also on Amazon Web Services, Google Cloud and Microsoft Azure. It can be picked in Claude Code, and it arrived the same day in the open-source coding tool OpenCode and its Go plan.
One practical split is to let Opus 5.5 or Sonnet 5.5 plan and have Haiku 5.5 carry out each defined step, which cuts cost per task.
Further reading
- OpenCode: Claude 5.5 Haiku now available in OpenCode and Gox.com ↗
- Introducing Claude Haiku 5.5anthropic.com ↗
2 sources
- @opencodeX♥ 1.3K
- @ClaudeDevsX♥ 8.5K
Sonnet 5.5 cache reads were halved to $0.10 per million tokens
According to SiliconANGLE, Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens. Haiku 5.5 cache reads cost $0.01.
Caching pays off when the same long prompt, such as system instructions or a document, is sent again and again. Writing to the cache costs $0.125 per million tokens on Haiku 5.5.
Further reading
- Anthropic releases Claude Haiku 5.5 small model and halves Sonnet 5.5 cache read pricessiliconangle.com ↗
- Introducing Claude Haiku 5.5anthropic.com ↗

Event · User reviews

Haiku 5.5: what people who tried it said

Hands-on reviews posted on X and Threads in the first day. Posts from Anthropic's official accounts and staff are left out, and both praise and criticism are included.

Positive@kickingkeys · X ↗
Surprisingly good at painting with code
They had Haiku 5.5, GPT-6 Luna and Sonnet 5.5 draw the same picture in code and posted the results side by side. At about a twentieth of Sonnet's price, the result was better than expected.
Positive@RLanceMartin · X ↗
Good for running many subagents at once
Under 100K input tokens, it is 20 times cheaper than Sonnet 5.5 and fast, they said. They also shared a map demo in which many Haiku 5.5 subagents stream drawing commands to one canvas.
Positive@addyosmani · X ↗
The team uses it as Opus 5.5's helper
Developer Addy Osmani said his team likes Haiku 5.5 as a subagent alongside Opus 5.5, and noted that its effort level can be adjusted.
Positive@aykenonchain · X ↗
Better than expected, will start most work on Haiku
Citing official tests where it beat GPT-6 Luna on knowledge work, computer use, coding and visual reasoning, they called it better than expected. They plan to start with Haiku and use a bigger model only when needed.
Mixed@larrabee.kor · Threads ↗
Measured at about 200 tokens per second
Timed in Claude Code, Haiku 5.5 was fastest at roughly 200 tokens per second, against about 100 for Sonnet and Opus. On the GPT side, Luna came close to 100 and the others ran at 20 to 30. Results may vary with the method, they added.
Mixed@LuminaBench · X ↗
Outside benchmarks call it 'not bad'
They reported 43.40 on the Artificial Analysis intelligence index, around Kimi K3, and ahead of Luna in the same class.
Negative@dhruvtwt_ · X ↗
Some call it dumb
After trying it, one user posted that it is 'insanely dumb', which drew about 150 likes. They did not say what task it failed.
Negative@cyapstaye · Threads ↗
Cheap, light models already exist
In a post titled 'Haiku 5.5, honest review', they asked why anyone needs it when cheap, light models are already plentiful.
Negative@xoaanya · X ↗
Has anyone actually used Haiku?
They posted that they still don't know anyone who uses Haiku for anything, a skeptical take on how much small models get used.

Other AI news

Picked from news outside X and Threads, published in the past 24 hours

US
California bars firing workers on AI alone, a person must review the decision
Governor Newsom signed SB 947, the No Robo Bosses Act, on September 30. It bars employers from using AI as the sole basis for firing or disciplining a worker. A person must review and verify the outcome, and employers must tell the worker in writing when an automated system was part of the decision.
It takes effect July 1, 2027, and is enforced by the Labor Commissioner, the state attorney general and local district attorneys. Newsom vetoed a broader earlier version, SB 7, in 2025.
Further reading
- Governor signs No Robo Bosses billstaffingindustry.com ↗
Policy
The FCC takes comment on letting AI-voice political robocalls go without consent, and advocates object
The Federal Communications Commission has opened public comment on a petition from Club for Growth. If granted, AI-generated political robocalls could reach cell phones without prior consent.
The petition caps calls at three per 30 days, but the National Consumer Law Center argues most political groups could avoid that limit. The FCC already ruled in 2024 that AI voice-cloning robocalls are illegal.
Further reading
- Stop efforts to greenlight AI robocalls ahead of the midtermsnclc.org ↗
Copyright
A US appeals court held that training AI on Westlaw headnotes is not fair use
The Third Circuit affirmed the lower court in Thomson Reuters' suit against Ross Intelligence. The court found Westlaw's headnotes copyrightable, and that Ross's use of them to train an AI tool was not fair use.
Ross's tool is not generative AI, so it is unclear whether the reasoning will carry over to generative AI cases such as the New York Times lawsuit.
Further reading
- US appeals court upholds Thomson Reuters' landmark win in AI training lawsuitthestar.com.my ↗
- 3rd Circuit rules for Thomson Reuters in its copyright fight against Rosslawnext.com ↗
Copyright & creative work
A Tokyo court said a human voice can be protected under publicity rights
On September 30 the Tokyo District Court ruled in a suit by voice actor Kenjiro Tsuda that a voice can fall under publicity rights. Using someone's voice without permission for commercial gain can be infringement.
The account in question had already closed, so the court rejected the request to delete the videos. In April Japan's Justice Ministry formed an expert panel on AI voice misuse.
Further reading
- Tokyo court grants legal protection to human voices in AI clone casekhaosodenglish.com ↗

View all 5 collected posts

* Summaries and publishing on this page are done by an automated agent. Check the facts against each story's original links.

Haiku 5.5 costs $0.10 per million input tokens up to 100K, one tenth of Haiku 4.5

More from today

Haiku 5.5 beats GPT-6 Luna on OSWorld, 72.4% to 48.9%

Ten Haiku 5.5 subagents under Opus 5.5 finished in the same 13 seconds for $0.04, against $0.05 for Opus alone

Summaries, conversation compaction and database queries are the jobs to give Haiku 5.5

Haiku 5.5 is on the Claude Platform, Claude Code, three major clouds and OpenCode

Sonnet 5.5 cache reads were halved to $0.10 per million tokens

Haiku 5.5: what people who tried it said

Surprisingly good at painting with code

Good for running many subagents at once

The team uses it as Opus 5.5's helper

Better than expected, will start most work on Haiku

Measured at about 200 tokens per second

Outside benchmarks call it 'not bad'

Some call it dumb

Cheap, light models already exist

Has anyone actually used Haiku?

Other AI news

California bars firing workers on AI alone, a person must review the decision

The FCC takes comment on letting AI-voice political robocalls go without consent, and advocates object

A US appeals court held that training AI on Westlaw headnotes is not fair use

A Tokyo court said a human voice can be protected under publicity rights