Claude Haiku 5.5 has just been released!

  It's the first time in the history of the world.

Recently, Anthropic has released Claude Haiku 5.5, one-tenth of the previous generation of Haiku 4.5, one-twentieth of Sonnet 5.5, and one price for OpenAI's GPT-6 Luna.

In the official run-off, Luna lost all six of her scores, and Sonnet was bitten by several others.

• OSWorld's operating computer jumped from 15.7% of the previous generation to 72.4%, only 11.5 percentage points behind Sonnet 5.5.
• Terminal-Bench 4.0 for terminal programming, the previous generation was zero, taking 39.2% this time.
• Gross Domestic Product (GDPval-AA) of knowledge work is down 1620 points, more than double the previous generation
• "Last Human Exam" HLE without tools, from 10.2% to 45.9%

The price is $0.10 per million Token input and $0.50 for output, which suggests a price within 100,000 Tokens, over 100,000 for input and $0.50 for output and $2.50 for output.

The cache reads $0.01 per million of tokens and goes through the batch processing five times.

According to official estimates, the same live average is about 75% higher than in the province of Haiku 4.5, with the same budget running about four times the number of calls.

It was not saved until 9%, because requests for more than 100,000 tokens are only half as cheap, so that in the previous generation of Haiku, about 90% of requests were within 100,000, and the new terminology will give the same sentence about 30% more tokens.

GPT-6 Luna has the same input, output, and cache price as 100,000 tokens.

The Luna is cheaper in the above context, with a price tag of around 270,000 tokens, $0.20 for input and $0.75 for output.

The caching reading of Sonnet 5.5 has also been cut in half as of today, with tokens dropping from $0.20 to $0.10 per million, officially estimating that most Agent tasks will be about two percent cheaper.

Max and Team subscribe for a monthly API allowance this week, Max 5x for $100, Max 20x for $200, Team for up to $500, and the entire team can share any model on the Claude Platform.

This generation of Haiku has made tremendous progress in running compared to the previous generation.

• OSWorld's operating computer jumped from 15.7% of the previous generation to 72.4%, only 11.5 percentage points behind Sonnet 5.5.
• Terminal-Bench 4.0 for terminal programming failed the previous generation, taking 39.2% this time.
• Gross Domestic Product (GDPval-AA) of knowledge work is down 1620 points, more than double the previous generation
• "Last Human Exam" HLE without tools, from 10.2% to 45.9%

Computer operation was one of the strongest advances, with Luna at only 48.9%, while Haiku 5.5 opened it up by 23.5 percentage points.

If the cost is more exaggerated, it goes to the Medium grade, which is higher than Luna's full grade, and costs less.

On the GDPval-AA of knowledge work, Haiku 5.5's 1620 points outperformed Luna's 1437 points, slightly less than Sonnet 5.5's 1840 points, and only the default Medium rating has 1277 points.

On the AA-Briefcase of the long-range running project, it jumped from 614 points to 1578 points from the previous generation, and Luna was 1336 points.

In reasoning and reading charts, the Haiku 5.5 is 57.4%, and the Sonnet 5.5 is 64.5% when the HLE comes with search and code tools. Chartography is 64.5% read on professional charts, up from 6.4% in the previous generation to 46.4%, Luna is 29.1%, and Sonnet 5.5 is 61.6%.

The biggest difference with Sonnet is programming, with its 39.2% on Terminal-Bench 4.0 being more than twice as much as Luna's (16.4%), but Sonnet 5.5 being 70.6%.

Its 46.4% on FrontierCode also won Luna's 42.4% and 52.1% off Sonnet 5.5's best score by almost 6 points.

Officials also made it clear that the sophisticated Agent programming was still handed over to Sonnet 5.5 and Opus 5.5.

The official ranking of Haiku 5.5 is large, fast, and cost-effective, such as summary writing, context compression, database checking, and sorting.

It is also currently the fastest Claude model, and is slower than the Opus, which has only a fast mode (the price is very expensive!) and is suitable for real-time customer service and browser operation.

Officials also gave feedback from a number of star AI app companies.

Asana estimates that the delay in completing the task is more than three percent lower than their current model, and Box estimates that the delay is only about half that of Haiku 4.5.

Another major use is to give Opus or Sonnet as a sub-agent, to dismantle the big model tasks, to take the idea, and to run the Haiku division.

With Cognition's Devin Fusion, Opus 5.5 is powered by Haiku 5.5 while its successor, FrontierCode, stays at its highest level of 66.2 points, with both costs and delays decreasing.

Rogo, a financial AI company, has made a big model of PPT, Haiku 5.5 digging into the revenue data of 10-K earnings per share.

Claude's developer account gave Demo the task of designing a shelf containing eggs from a pile of household waste, with the goal of throwing them from 32 meters.

The Opus 5.5 alone took 3 minutes and 37 seconds, tested 25 programs, cost $0.47, and with 10 Haiku 5.5s, Agent only took 58 seconds, tested 86 programs, cost $0.14.

Haiku 5.5 is already available on platforms such as Claude Client, Claude Code and AWS, Google Cloud, Microsoft Azure, Cursor, OpenRouter, and the model ID is claude-haiku-5-5, with 1 million tokens and a maximum output of 128,000 tokens at a time.

It was the first Haiku to adjust thinking grades, from Low to Max, with the default API being Medium.

Python and the TypeScript SDK also incorporate computer and browser operations, currently in Beta.

From the Haiku 4.5 migration, the same text Token will be increased by about 30%, max_tokens and cost estimates will be recalculated.

Budget_tokens were replaced by adaptive thinking plus Effort grades, temperature, top_p, top_k were removed, pre-filled replies could not be used, and the format required a change in structured output.

Priority Tier is not currently supported for computer_toolset_20260801.

It's easy to get into trouble with a claude code:

/claude-api migrate this project to claude-haiku-5-5

Let it change itself.

Of course, for non-developers, you can now talk directly to Haiku 5.5 in the client!

The simple question is, don't bother Fable, Opus, Sonnet, Haiku 5.5 is a perfect substitute for search engines.

Go ahead and try it!