Tech • AI • Robotics • Game
Claude Haiku 5.5 Is a Game Changer! 90% Cheaper & Insane Performance! (Fully Tested)
9.2/10TL;DR
Anthropic has launched Claude Haiku 5.5, a much cheaper and far more capable small model aimed at coding, automation and high-volume enterprise tasks, intensifying competition with OpenAI.
KEY POINTS
Major upgrade at lower cost
Haiku 5.5 is positioned as Anthropic’s cheapest, fastest and most capable small model so far. It is designed for summarization, classification, customer support, browser automation and coding sub-agents, and can offload routine tasks from larger models such as Opus 5.5 and Sonnet 5.5. The company says average usage is about 75% cheaper than Haiku 4.5 while delivering sharply better results.
Benchmarks show large jumps
Reported benchmark gains are unusually large for a lower-tier model. On OSWorld, which measures computer-use ability, performance rose from 15.7% to 72.4%. On Humanity’s Last Exam, scores improved from 10.2% to 45.9% without tools and from 18.7% to 57.4% with tools, while Chartography increased from 6.4% to 46.4%.
Pricing is aggressively reduced
For prompts under 100,000 tokens, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens. That compares with Haiku 4.5 pricing of $1 per million input tokens and $5 per million output tokens, a 90% cut in base rates. For prompts above 100,000 tokens, pricing rises to $0.50 per million input tokens and $2.50 per million output tokens, still about 50% cheaper than the earlier model.
Most workloads should benefit
Anthropic says roughly 90% of requests to the previous Haiku model were below 100,000 tokens, suggesting most existing workloads qualify for the lowest pricing tier. Prompt caching is also pitched as inexpensive. The company separately cut cached-read prices for Claude Sonnet 5.5 in half to $0.10 per million input tokens, a move likely to help coding agents and other context-heavy systems.
Built for agents and adjustable reasoning
Haiku 5.5 is the first Haiku model with adjustable effort, allowing developers to trade off speed and cost against deeper reasoning. That feature is particularly useful for agentic systems where some steps are simple and others require more deliberation. The model is also available in Claude Code, where its low token pricing makes it attractive for workflows involving dozens or hundreds of small actions.
Capability claims extend beyond benchmarks
Demonstrations highlighted strong code generation and front-end output, including a Minecraft-like game with cave systems and water physics, plus a Call of Duty Zombies-style browser game with stamina, rebuildable barriers, sound effects and weapon mechanics. Another reported test produced a polished Minecraft Dungeons-inspired isometric action game in about one hour, allegedly using only 1% of a five-hour usage limit. These examples suggest the small model is now viable for surprisingly complex interactive builds.
Competitive pressure on rivals
The model is described as outperforming GPT-6 Luna on several benchmarks tied to computer use, knowledge work, frontier coding and visual reasoning. In one coding comparison against Grok 4.7, Haiku 5.5 reportedly finished in about 27 minutes versus 35 minutes, while costing roughly one-fifth as much, though the tests were run in different coding environments. On the World of AI benchmark platform, it was said to rank 10th overall, ahead of Grok 4.7.
Cheap per token, but usage still matters
One caveat is token consumption at high reasoning settings. An external analysis cited Haiku 5.5 at maximum effort using about 162 output tokens per intelligence-index task, roughly three times the level of GPT-6 Luna at maximum effort. That means real-world task cost will depend not only on token rates but also on how much reasoning developers ask the model to perform.
CONCLUSION
Haiku 5.5 signals an aggressive push by Anthropic to make smaller AI models far more useful without premium pricing. If the reported gains hold up in production, the release could make low-cost coding and automation agents significantly more practical at scale.
Ask a question