How LLM API token pricing works: input, output, cache and context tiers
“Price per million tokens” is not a single number: the real cost depends on how many tokens you send versus receive, whether prompts are cached, and how long the context is. Here is the math on November 2026 list prices.
Input and output are priced differently
Prefill parallelises, decoding does not — so output costs several times input. GPT-5.5 is $5/$30 per 1M (6×), Claude Opus 5.5 is $4/$20 (5×), budget GPT-5.6 Luna is $0.2/$1.2 (6×). Comparing models on input price alone misleads: chat is input-heavy, code generation is output-heavy.
Blended cost for your workload
With R input tokens per output token, blended = (R×input + output)/(R+1). GPT-5.5 at an agentic 10:1 → (10×5+30)/11 ≈ $7.27 per 1M. At a Q&A 2:1 → (2×5+30)/3 ≈ $13.33. The same list price differs 2× across workloads — that is why our rankings normalise by token mix instead of quoting “in a vacuum” numbers.
A real request: 2,000 input, 500 output
That request costs $0.025 on GPT-5.5, $0.018 on Claude Opus 5.5, $0.009 on Claude Sonnet 5.5 and $0.00065 on Claude Haiku 5.5. At 1,000 requests a day it is $761, $548, $274 and $20 a month. Flagship-to-budget spread is nearly 40×, and no reseller discount closes it: pick a cheaper model before hunting a cheaper vendor.
Prompt caching: up to 90% off the input
Anthropic charges $0.20 per 1M for cached reads on Opus 5.5 versus $4 regular — 20× cheaper — while cache writes cost more than plain input ($5). With 70% cache hits the blended GPT-5.5 price on a 10:1 task falls from $7.27 to roughly $4.41. Cheap-input providers without caching lose on long prompts.
Context tiers and promo deadlines
OpenAI doubles prices above 272K context (GPT-5.5: $10/$45). Claude Haiku 5.5 has two tiers around a 100K prompt threshold. Google's promotional Gemini 3.8 Flash price ($0.75/$3.75) holds until December 31, 2026 — then $1.50/$7.50. A comparison that ignores tiers and dates is wrong; our model cards flag both.
Batch and wholesale schemes
Official batch modes are typically −50% on both sides but answer in hours. Aggregators shave 10–30% off synchronous prices via volume contracts and cache hits. Run the numbers at your blended cost: for offline analytics, batch beats any reseller.
Published 2026-11-11. Prices are vendor list prices on the publish date; live quotes are in the catalogue.