LLM API Invoice Traps: Thinking Tokens, Cache Charges, Context Multipliers
6 traps, all in official docs the number that matters here
By Stan K.
All three majors officially bill reasoning tokens you never see, OpenAI quietly added a cache write charge, Gemini doubles input price past 200K, and Anthropic adds 10% for US-region inference. Six verified fine-print traps that add 30–200% to a naïve token estimate.
Every trap below is a direct quote or figure from official provider documentation — verified, not community lore. None of them is hidden; all of them are invisible until the invoice arrives. Together they routinely add 30–200% to a back-of-napkin token estimate.
1. You pay for thinking you never see
The single biggest line-item surprise in modern APIs:
- Anthropic: tokens Claude spends reasoning are “billed as output tokens, even when the thinking text isn’t returned” — you pay full thinking, not the summary, and a response with thinking omitted costs the same.
- OpenAI: “reasoning tokens are not visible via the API… and are billed as output tokens” — and on an
incompletefinish you can pay input + reasoning with no visible answer at all. - Google: “response pricing is the sum of output tokens and thinking tokens” — if
max_output_tokenscuts the response off mid-reasoning, the thinking tokens are still billed. The control isthinking_level.
A “short answer” from a reasoning model is not a cheap answer. The visible 200-token reply may ride on 10,000 billed reasoning tokens at output rates.
2. Caching now costs money to fill
The classic advice — “prompt caching is free savings” — is outdated at OpenAI:
- OpenAI: GPT-5.6+ models now charge a cache write of 1.25× input (30-minute TTL) or 2.5× (12-hour). Reads drop to 0.1× (0.05× on GPT-6.1 Sol/Terra) — the discount is real, but filling the cache is no longer free. Minimum cacheable prompt: 1,024 tokens.
- Anthropic: cache write 1.25× (5-min TTL) or 2× (1-hour); reads at 0.1× input (as low as 0.025× on the newest models). Below the minimum (512–4,096 tokens depending on model) caching silently doesn’t apply — no error, no discount.
- Gemini: implicit caching on by default; minimums 2,048–4,096 tokens; explicit caching has disappeared from the newest APIs.
The trap inside the trap: prompt caching amplifies the wrong pattern. Cache a prompt you’ll never re-read and you paid a premium (1.25–2×) for nothing.
3. Long context has a multiplier — at some vendors
- Gemini: past 200K tokens, input price doubles (2×) and output rises 1.5×; cache rates double too. A 3.1 Pro prompt that was $2/1M becomes $4/1M when you cross the line.
- Anthropic: the old 1M-context beta 2× multiplier is gone — 1M context on Claude 4.6+ bills at standard price. Exception: Haiku 5.5 above 100K jumps to 5× input ($0.10 → $0.50) — the sleeper cliff in the cheap model.
- OpenAI: no documented context multiplier.
4. Region surcharges
Anthropic charges 1.1× on all usage when inference geography is pinned to the US (inference_geo). Compliance-driven US-region pinning is a flat 10% on the whole bill.
5. Batch discounts are real — and stack
The honest counterweight: all three majors offer −50% for batch (non-urgent jobs, up to 24-hour turnaround), and at Anthropic batch stacks on top of prompt-caching discounts. Anything that runs overnight — evaluation sweeps, enrichment passes, bulk classification — belongs in Batch; leaving it interactive is a 2× overpayment on purpose.
6. Brokers: the fee is not where you think
OpenRouter passes through inference pricing with no markup — but charges 5.5% (minimum $0.80) when you buy credits, and 5% of the OpenRouter price if you bring your own provider key (with a free pay-as-you-go allowance of $25K/month on BYOK). The trap is comparing sticker prices: a broker’s model price can equal the provider’s while your effective cost carries the payment fee — or, with BYOK, be cheaper than the provider’s list for routing alone.
Every API bill multiplies a number you can’t observe at request time — hidden reasoning depth, cache fill vs cache hit, the token count that crosses 200K, the region flag set for compliance — by a rate you read on a different page. Meter the four numbers yourself (reasoning tokens, cache writes vs reads, context length, geo) before you trust any estimate.
The fix list
- Set thinking budgets explicitly (
thinking_level, reasoning effort) on every production call — default “let it think” is a blank check. - Measure cache hit rate, not just cache existence: writes at 1.25–2.5× only pay off above roughly 2–3 re-reads.
- Watch the 200K line (Gemini) and the 100K Haiku cliff — chunk or summarize before crossing.
- Move everything non-interactive to Batch — the only trap working in your favor.
- Price the payment fee, not the model fee, when choosing a broker.
For the other side of this market — when a flat coding plan absorbs these mechanics entirely — see Coding Plan vs API. The full plan-by-plan value ranking is in the Coding Plans hub.