Six LLM pricing traps that turn a cheap model into an expensive bill

Nobody gets a shocking API bill because the headline price was high. They get it because of the fine print. These are the six traps we watch when we verify prices — every one of them is in a provider’s current pricing, dated October 2026.

1. The 100K cliff (Claude Haiku 5.5)

Haiku 5.5 is Anthropic’s cheapest model at $0.10/$0.50 per million tokens — for prompts up to 100,000 tokens. Cross that line and the entire prompt is billed at $0.50/$2.50, five times the rate. A RAG pipeline that usually retrieves 80K tokens but occasionally pulls 120K does not have a predictable bill; it has a lottery. We flag this on the Haiku model page and in the calculator itself.

2. The long-context segment (GPT-6.1 Sol)

Above roughly 272K input tokens, Sol enters a long-context pricing segment where input and cached rates roughly double and output rises by about half. Load-testing with short prompts will never reveal it. Your production conversations will.

3. The calendar doubling (Gemini 4 Argon)

Argon’s introductory $2/$10 pricing expires November 30, 2026; from December 1 the standard rate is $4/$20. Any budget, quote or unit-economics model built on Argon in October is wrong by 2x in December. Promotional pricing is a loan, not a price — Gemini’s Flash models carry the same pattern, doubling on January 1, 2027.

4. Thinking tokens are output tokens (Gemini)

Gemini’s output price includes the model’s internal reasoning tokens. A “500-token answer” from a reasoning-heavy request can bill several thousand output tokens. Estimates built from visible response length alone undercount, sometimes badly.

5. Speed tiers multiply everything (OpenAI)

OpenAI’s Fast and Ultrafast service tiers multiply the standard rates — Ultrafast GPT-6 Astra is $60/$300 per million, six times its standard tier. Tiers get enabled in consoles, inherited from example code, and forgotten. Audit which tier your keys actually call.

6. Gateway prices wearing an official costume

Resellers such as OpenRouter list the same models at their own rates, sometimes higher, sometimes lower, occasionally stale. Calculators that mix gateway and official prices without labeling them produce numbers that match nobody’s invoice. Check that any price you rely on names its source and its date — ours do, on every model page.

The defense

None of these traps is hidden; all of them are published. They bite because estimates are usually made once, from headline prices, and never revisited. Re-run your numbers when a model changes price, when your prompts grow, and when the calendar crosses a promotional expiry. That is the entire discipline — and it is why every figure on this site carries a verification date.