The Six Lines That Matter
| Line | What it means | Where it hides | Impact |
|---|---|---|---|
| Input price | Cost of tokens you send | The headline number everyone compares | Low — often <20% of real spend |
| Output price | Cost of tokens the model writes | Buried below input on most pages | Dominant — 60-90% of spend |
| Cached input | Discount for repeated prefixes | A separate line or fine print | Up to 97% off input |
| Reasoning tokens | Thinking billed as output | Not shown; appears on the invoice | Can triple the output line |
| Batch / off-peak | Discounts for async or night work | Separate section, easy to miss | 50% or more when applicable |
| Long-context surcharge | Multiplier past a token threshold | Footnotes on premium tiers | 2x on big-context workloads |
The order matters: the first line is marketing, the second is math. GPT-5.6 Terra and Claude Sonnet 5 both show $2/1M input — identical headline. Terra's $12 output vs Sonnet's $10 flips the comparison at 4K/1K, and vs Gemini 3.7 Flash's $6 output the gap is 40%. You cannot compare pricing pages by their first line.
The Worked Comparison: 4K In / 1K Out
Same workload on three pages, verified 2026-08-28 rates, 1M requests a month:
Then add the hidden lines: 50% cacheable input drops all three by roughly a third (Flash to ~$0.0085/req), and a batch/off-peak schedule halves the rest. The pricing page never shows the number you'll actually pay — you have to compute it from your mix.
The Five-Minute Page Read
Find the output price first — it decides the bill. If it's not on the first screen, the page is hiding it.
Check for a cached-input line. No line = no discount = repeated prefixes bill at full price.
Scan for off-peak and batch sections. Half-price for night work is the easiest money in AI pricing.
Read the footnotes for context thresholds (e.g., GPT-5.6 bills 2x past 272K tokens).
Search for 'reasoning' or 'thinking' — if the model has it, output costs more than the page implies.
Where We Do This For You
Every record in the model catalog carries all six lines where published — input, output, cached, tier notes, batch flags, context surcharges — with the source link and verification date on every page. The prompt cost calculator applies your mix automatically.
Frequently Asked Questions
What do input and output prices mean on AI pricing pages?
Input is what you send (prompt, history, documents); output is what the model writes. Output typically costs 3-25x more because generating tokens is compute-heavy. Prices are almost always quoted per 1M tokens — not per request.
Why does my bill never match the pricing page?
Four hidden lines: cached input (repeated prefixes bill at a fraction), reasoning tokens (billed as output on thinking models), long-context surcharges (some models 2x past a token threshold), and off-peak/batch discounts that don't apply automatically. The pricing page shows the base rate; the bill shows the mix.
Is the input price or output price more important?
Output — it is 60-90% of real spend on most workloads and the spread between models is widest there. Two models can share an input rate and differ 5x on output.
How do I compare two pricing pages fairly?
Same workload, same mix: 4K input / 1K output per request, 50% cacheable, 1M requests. Price both models on that identical shape — the per-request number is the only honest comparison.
What is a cached input price?
The rate for input tokens that match a previously seen prefix. Providers discount it 50-97% because the prefix's attention state is reused. It only applies to stable, repeated text — not unique per-user content.