Claude Haiku 5.5 Hands-On: 90% Cheaper, Small Model Performance That Blows You Away

Anthropic is sprinting toward its potential IPO, and the newly released Claude Haiku 5.5 puts enormous pressure on competitors. Positioned as the lowest-tier model in Anthropic's lineup, it delivers a surprisingly dramatic performance leap — not only are running costs significantly lower, but the overall capability improvements are hard to believe coming from a "small model."

From Haiku 4.5 to 5.5: A Generational Performance Jump

To appreciate what this update means, look back at Haiku 4.5 from eleven months ago. Its macOS clone interface wasn't terrible, but quality was clearly incomparable to today's output. Haiku 5.5 can generate a fully functional Minecraft clone — complete with cave systems, water physics, and features that actually work. That level of detail was once aspirational for any model; now it's being delivered by Anthropic's most affordable offering.

Haiku 5.5 is designed for high-concurrency, cost-sensitive workloads including summarization, classification, customer support, browser automation, and coding sub-agents. It can also collaborate with more powerful models like Opus 5.5 and Sonnet 5.5, handling smaller tasks so you're not burning expensive large models on everything.

Benchmarks: 3x to 5x Improvements

The benchmark data is the most eye-catching part of this release. Haiku 5.5 achieves 3x to 5x — or greater — improvements over 4.5 across multiple dimensions:

• OS World (evaluating computer operation capability): jumped from 15.7% to 72.4%
• Humanity's Last Exam: without tools from 10.2% to 45.9%, with tools from 18.7% to 57.4%
• Chartography: from 6.4% to 46.4%

For a model positioned at the smaller, cheaper end of the spectrum, these gains are remarkable. According to the author's analysis, Haiku 5.5 outperforms GPT-6 Luna on multiple benchmarks including computer use, knowledge work, frontier coding, and visual reasoning. Benchmarks don't tell the whole story, of course, but they do reflect the capability level this model has reached.

OS World is an open academic benchmark that evaluates an AI model's ability to complete tasks in real computer operating system environments — covering file management, app operation, web browsing, and more. It's considered a key measure of "computer use" capability. Humanity's Last Exam (HLE) is an extremely difficult knowledge test released by the Center for AI Safety and other organizations, featuring graduate-level questions from mathematics, science, law, and other top disciplines. Even human experts average below 6% correct, making it a measure of how close a model approaches the cognitive frontier of human expertise. Substantial jumps on both of these recognized high-difficulty benchmarks mean Haiku 5.5 is not only more practical for everyday tasks — its reasoning depth and general knowledge ceiling have improved qualitatively.

Pricing Strategy: 90% Cost Reduction

Pricing is undoubtedly one of the biggest highlights of this release. For prompts under 100K tokens:

• Input: just $0.10 per million tokens
• Output: just $0.50 per million tokens

Compared to Haiku 4.5 ($1.00/million input, $5.00/million output), this is a 90% reduction in base rates. Even for prompts exceeding 100K tokens ($0.50/million input, $2.50/million output), it's still about 50% cheaper than 4.5.

Anthropic notes that roughly 90% of Haiku requests previously fell under 100K tokens, meaning the vast majority of existing workloads will benefit from the lower pricing tier. Cache pricing is also extremely affordable.

One detail worth noting: under high reasoning intensity, Haiku 5.5 is fairly token-hungry. According to Artificial Analysis data, at max effort mode, completing each intelligence index task consumes roughly 162 output tokens — about 3x what GPT-6 Luna uses in equivalent mode. So while each token is extremely cheap, actual task costs still depend on token consumption. That said, this class of small model isn't intended for maximum reasoning intensity tasks anyway.

Anthropic also cut Sonnet 5.5's cache read pricing in half, down to just $0.10 per million tokens, making it roughly 20% cheaper for many long-running workloads — particularly beneficial for coding agents that repeatedly reuse large amounts of context.

Adjustable Reasoning Intensity: Coming to Haiku for the First Time

Haiku 5.5 is the first Haiku model to support adjustable effort. Developers can now choose how much reasoning resources the model should invest based on task difficulty: simple requests can prioritize speed and cost, while complex tasks can dial up effort for better results. This is especially useful for agentic workloads — not every task needs the same level of intelligence.

On the World of AI benchmark platform, Haiku 5.5 currently ranks 10th overall, even surpassing Grok 4.7. It performs particularly well in debugging, agentic tasks, and frontend generation.

Adjustable effort (also called extended thinking) essentially lets the model engage in an internal chain-of-thought before giving a final answer. Higher effort levels mean more tokens consumed and deeper reasoning steps, but also higher latency and cost. This mechanism first appeared in OpenAI's o-series and Anthropic's larger models; the Haiku lineup had previously focused on speed and cost and didn't support it. Haiku 5.5 bringing this to a small model means developers building agentic workflows can now dynamically balance "fast and cheap" versus "deep reasoning" at a granular level — for example, having sub-agents use low-effort mode for simple routing in multi-step tasks, then temporarily switching to high-effort mode for complex judgment calls, without needing to invoke more expensive large models.

Real-World Testing: Gaming, Frontend, and Motion Design

The author ran a series of real-world tests. On the Vibe Coding platform in ultra code mode, Haiku 5.5 was prompted to create a Call of Duty zombies mode clone using Three.js. The results were impressive: lighting, shaders, and overall atmosphere were well-executed, along with complete gameplay mechanics and sound effects — running, throwing grenades, rebuilding barriers, mystery weapon boxes, and a stamina system. Critically, this barely made a dent in the ultra code token budget.

On the frontend side, Haiku 5.5 shows clear improvement — fast generation, high quality, and excellent pricing. While it's not the top choice for SVGs, its overall performance is strong, with isometric rooms and shader machine simulations demonstrating a distinctive design sensibility.

A particularly representative comparison pitted Haiku 5.5 against Grok 4.7: Haiku 5.5 completed the task in about 20 seconds, while Grok 4.7 took roughly 35 minutes, at about one-fifth the cost — with visual quality the author deemed superior. The one caveat is that they were tested in different coding environments (Haiku in Claude Code, Grok in Grok Build), so toolchain experience differs.

In a comparison with GPT-6 Luna, both models were asked to generate "a zebra running in Africa." Haiku 5.5 clearly outperformed GPT-6 Luna in visual quality and depth.

Stand-out community examples have also emerged: X user Leon used Haiku 5.5 to clone a polished Minecraft Dungeons-style isometric action-adventure game in under an hour, complete with detailed environments, melee combat, dodge mechanics, animations, particles, lighting, and environmental interactions. Most remarkably, this reportedly consumed only 1% of Leon's five-hour Claude usage allocation. In motion design, Haiku completed in just 15 minutes what Fable, GPT-6 Astra, and even Opus 5.5 needed over 40 minutes to produce — at a fraction of the cost.

The sub-agents mentioned throughout refer to an AI workflow architecture pattern: an "orchestrator" large model breaks down complex tasks and distributes subtasks to multiple smaller agents specialized in specific functions running in parallel. Haiku 5.5's low cost and high throughput make it a natural fit for sub-agent roles in this architecture — for example, in a code review pipeline where Opus handles understanding overall requirements while Haiku scans files for syntax errors or generates unit tests. Vibe Coding is a development paradigm that uses natural language as the primary interaction mode for rapid iterative generation of runnable code, typically via AI platform tools like Claude Code or Cursor, emphasizing the experiential flow of the process over strict engineering standards. The real-world tests in this article were all conducted in such environments.

Conclusion: A Strong Candidate for Everyday Use

Overall, Anthropic has delivered an impressive result here. For such a small and affordable model, the level of intelligence and capability is remarkable — from building polished games and motion design to agentic tasks, Haiku 5.5's performance in many scenarios approaches that of Sonnet.

What's most compelling is the value proposition. Compared to Grok and Luna, it completes tasks faster, and while it's not cheaper than Luna per token, the quality is better. The author considers Haiku 5.5 likely to become one of the best "everyday workhorse" models for coding, automation, and daily productivity: fast, capable, and affordable. Seeing this kind of performance from Anthropic's smallest model is genuinely remarkable.