The fiercest price war in AI history is playing out right now in lightweight, budget language models—the "small workhorses." These models handle the lion's share of production server traffic: triaging incoming emails, extracting entities, filtering logs, routing tasks, and powering lightweight RAG pipelines.
Until recently, Anthropic had been losing that niche badly: the older Haiku 4.5 cost noticeably more than ultra-cheap rivals from OpenAI (the GPT-4o-mini and GPT-6 Luna families).
Claude Haiku 5.5 was pitched as a knockout counterpunch: Anthropic touted a huge leap on coding and reasoning benchmarks alongside eye-catching banner pricing—just $0.10 per million input tokens and $0.50 per million output tokens.
Looked like an easy win. Then independent researcher Simon Willison—creator of Datasette and the llm CLI tool—ran a thorough technical audit and uncovered several hidden catches the marketing banners left out. Let's check the real math.
The 100K Token Cliff: A 5x Price Jump
The first surprise hides in the API pricing fine print. That advertised "ten cents per million" rate applies strictly as long as your total context stays under 100,000 tokens.
Cross that 100K-token threshold—say, to analyze an annual financial report or a large code repo—and a 5x surcharge kicks in immediately:
- Input tokens jump from $0.10 to $0.50 per million;
- Output tokens leap from $0.50 to $2.50 per million.
If a long multi-volume PDF accidentally slips through your RAG queue, the cost of a single call multiplies fivefold out of nowhere. Compared to rivals holding flat pricing across the whole context window, tiered pricing makes Haiku 5.5 a risky bet for large-document workloads.
A 25% Hungrier Tokenizer
The second hidden tax sits deeper, inside how text gets chopped into tokens. Simon Willison ran the exact same English test corpus through Anthropic's official Claude Token Counter on both Haiku 4.5 and the new Haiku 5.5.
Turns out the 5.5 generation's tokenizer vocabulary was heavily reworked. On identical text, Haiku 5.5 counts an average of 1.25x more tokens (+25%) than its predecessor.
What does that mean for your wallet? Even when the nominal price per million tokens looks cheap on paper, your actual bill for processing the exact same body of text quietly climbs by a quarter simply because it's sliced into smaller pieces.
Forced Reasoning: Goodbye, Instant Responses
The third architectural change is mandatory reasoning mode (Forced Reasoning). In Haiku 5.5, thinking effort is locked at medium effort, and developers can't turn off hidden thinking-token generation via an API flag.
Testing simple SVG vector generation ("a pelican riding a bicycle") in the llm CLI, the model spends noticeable time on internal reflection before outputting anything. That pushes up Time to First Token (TTFT). Fine for background batch jobs, but if you're building a voice bot or instant search autocomplete, an extra second of dead air ruins the user experience.
Haiku 5.5 genuinely is much sharper at coding and strict logic. Just don't deploy it blindly: if your pipeline pushes past 100K tokens or lives and dies by millisecond latency, competing models remain far more predictable on cost.
