Just now, Claude Haiku 5.5 was released! The price plummeted by 90%

Just now, Anthropic released Claude Haiku 5.5, priced at only 1/10 of the previous generation Haiku 4.5, 1/20 of Sonnet 5.5, and the same price as OpenAI's GPT-6 Luna.

In the official benchmark table, Luna lost all six categories to it, with several categories closely competing with Sonnet.

The OSWorld for operating computers jumped from 15.7% in the previous generation to 72.4%, just 11.5 percentage points behind Sonnet 5.5.

For terminal programming, Terminal-Bench 4.0 scored 39.2%, up from zero in the previous generation.

The GDPval-AA for knowledge work scored 1620 points, more than double the previous generation.

The "Final Exam for Humanity" HLE, without tools, increased from 10.2% to 45.9%.

How Much Can You Save

The price is $0.10 per million tokens for input and $0.50 for output, which is the price for prompts under 100,000 tokens; for over 100,000, it is calculated at $0.50 for input and $2.50 for output.

Cache reading costs $0.01 per million tokens, and batch processing gets an additional 50% off.

According to official calculations, the same tasks average about 75% savings compared to Haiku 4.5, allowing approximately four times the number of calls for the same budget.

The savings are not as high as 90% because requests over 100,000 tokens are only half the price; fortunately, about 90% of requests in the previous generation Haiku were under 100,000 tokens. Additionally, the new tokenizer will increase the number of tokens for the same passage by about 30%.

Compared to GPT-6 Luna, the input, output, and cache prices for both are exactly the same for under 100,000 tokens.

When dealing with long contexts, Luna is cheaper, as it only raises prices around 270,000 tokens, and even then, it costs $0.20 for input and $0.75 for output.

Sonnet 5.5's cache reading also gets halved starting today, dropping from $0.20 per million tokens to $0.10, with most Agent tasks estimated to be about 20% cheaper.

Max and Team subscriptions will start providing monthly API credits this week, with Max 5x giving $100, Max 20x giving $200, and Team up to $500, shared among the whole team, usable on any model on the Claude Platform.

Scored Higher than Luna but Still Lags Behind Sonnet in Complex Programming

This generation of Haiku has made significant improvements in scoring compared to the previous generation.

The OSWorld for operating computers jumped from 15.7% in the previous generation to 72.4%, just 11.5 percentage points behind Sonnet 5.5.

For terminal programming, Terminal-Bench 4.0 scored 39.2%, up from zero in the previous generation.

The GDPval-AA for knowledge work scored 1620 points, more than double the previous generation.

The "Final Exam for Humanity" HLE, without tools, increased from 10.2% to 45.9%.

Computer operation is the area with the most significant improvement this time, with Luna at the same price only scoring 48.9%, while Haiku 5.5 outperformed it by 23.5 percentage points.

Considering costs, it is even more impressive; it exceeded Luna's full score while running at the Medium setting, costing even less.

In knowledge work, the GDPval-AA scored 1620 points, surpassing Luna's 1437 points, but still trailing Sonnet 5.5's 1840 points, even at the default Medium setting, it scored 1277 points.

For long-term projects on AA-Briefcase, it rose from 614 points in the previous generation to 1578 points, while Luna scored 1336 points.

In reasoning and image reading, when HLE uses search and code tools, Haiku 5.5 scored 57.4%, while Sonnet 5.5 scored 64.5%. For reading professional charts, Chartography increased from 6.4% in the previous generation to 46.4%, while Luna scored 29.1% and Sonnet 5.5 scored 61.6%.

The biggest gap with Sonnet is in programming; on Terminal-Bench 4.0, its 39.2% is more than double Luna's 16.4%, but Sonnet 5.5 scored 70.6%.

On FrontierCode, it scored 46.4%, also beating Luna's 42.4%, but still nearly 6 percentage points behind Sonnet 5.5's best score of 52.1%.

The official statement also clearly indicates that complex Agent programming should still be left to Sonnet 5.5 and Opus 5.5.

What It Is Suitable For

The official positioning for Haiku 5.5 is for tasks that require high volume, speed, and cost-effectiveness, such as writing summaries, compressing contexts, querying databases, and classification.

It is also currently the fastest Claude model, only slightly slower than Opus when running in fast mode (which is very expensive!), making it suitable for real-time customer service and browser operations.

The official feedback from several star AI application companies has also been provided.

Asana found that task completion delays were reduced by more than 30% compared to their current model, while Box measured its delay to be about half that of Haiku 4.5.

Another main use is as a sub-agent for Opus or Sonnet, where the large model breaks down tasks and makes decisions, while Haiku runs them.

Cognition's Devin Fusion uses Opus 5.5 as the main model and Haiku 5.5 as the assistant, maintaining a top level of 66.2 points on FrontierCode, with reduced costs and delays.

The financial AI company Rogo uses the large model to create PPTs, with Haiku 5.5 diving into 10-K reports to extract revenue data by division.

The Claude developer account provided a demo, where the task was to design an egg holder using a bunch of household items, with the goal of not breaking when dropped from 32 meters.

Opus 5.5 took 3 minutes and 37 seconds, tried 25 solutions, and cost $0.47, while using 10 Haiku 5.5 sub-agents only took 58 seconds, tried 86 solutions, and cost $0.14.

Haiku 5.5 User Guide

Haiku 5.5 is now live on the Claude client, Claude Code, and platforms like AWS, Google Cloud, Microsoft Azure, Cursor, OpenRouter, etc. The model ID is claude-haiku-5-5, with a context of 1 million tokens and a maximum output of 128,000 tokens per request.

It is the first Haiku that can adjust the thinking mode, with five levels from Low to Max, and the API defaults to Medium.

The Python and TypeScript SDKs have also added computer operation and browser operation, currently in Beta.

When migrating from Haiku 4.5, the same text tokens will increase by about 30%, and max_tokens and cost estimates need to be recalculated.

The budgettokens have been replaced with adaptive thinking and Effort levels, while temperature, topp, and top_k need to be removed, and pre-filled responses can no longer be used; format requirements have changed to structured output.

Computer operations need to switch to the new toolset computertoolset20260801, and the Priority Tier is not yet supported.

If you find it troublesome, you can run a command in Claude Code:

Let it modify itself.

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

For simple questions, there's no need to bother Fable, Opus, or Sonnet. Even with low inference intensity, Haiku 5.5 can accurately provide answers in seconds, perfectly replacing search engines.

Go try it out!

References:

https://www.anthropic.com/claude-haiku-5-5