o3 Reasoning Token Cost
Using o3? The thinking tokens you never see bill at its output rate. This page computes how much of your real cost is hidden reasoning — set your own ratio below.
For teams using reasoning models — computes the hidden thinking-token cost you never see in the answer but pay for at the output rate, driven by your own observed reasoning ratio, not a guess.
Model prices from OpenRouter · updated 2026-07-13
01 Your workload
$2/M in · $8/M out — reasoning tokens bill at the output rate
02 Visible answer vs real bill
Hidden reasoning is 77% of your bill — the thinking costs more than the answer you read.
$0.006000 visible + $0.02 hidden reasoning = $0.03/request
Real cost is 4.3× the visible-answer estimate
Related cost calculators
Prices from OpenRouter, snapshot 2026-07-13, synced daily. Reasoning tokens = ratio × visible output, billed at the model's output rate (how OpenAI o-series, Anthropic extended thinking, DeepSeek R-line and Gemini thinking charge). The reasoning ratio is your workload input — check your provider's usage report for the real figure. All math runs in your browser.
How the math works
With o3, the answer you read is only part of what you pay for. Before it replies, the model burns hidden "thinking" tokens you never see — and those bill at the output rate, $8/M, same as the visible answer. At a 5× reasoning ratio, a 500-token answer hides 2,500 thinking tokens behind it.
Same baseline, one identity: the cost you'd estimate from the visible answer is $0.006000/request (input + the 500 tokens you see). Add the hidden reasoning ($0.02) and the real cost is $0.03 — 4.3× the visible estimate. The thinking is 77% of your bill, invisible on any per-token quote of the answer.
The reasoning ratio is a workload number, not a fixed constant — it depends on the task and the effort level you request. Simple lookups think very little; hard multi-step problems can burn 10× the visible answer or more. On the o-series and extended-thinking models the effort/budget setting is the main dial: "high" reasoning multiplies the hidden tokens. Tune the slider to what you actually observe in your logs.
The lever is matching effort to task. Dropping from high to low reasoning effort, or routing simple requests to a non-reasoning model, cuts the hidden tokens directly — usually more than any input-side optimization, because reasoning is billed at the pricier output rate. At 500,000 requests/month the hidden thinking alone is $10,000.
Providers bill reasoning tokens as output tokens (OpenAI o-series, Anthropic extended thinking, DeepSeek R-line, Gemini thinking), so this uses o3's real output rate from the daily-synced catalog — the only estimate is the reasoning ratio, which is yours to set. Inference prices sync daily from OpenRouter (updated 2026-07-13). All math runs client-side with tested code.
Frequently asked questions
How much do hidden reasoning tokens cost on o3?
At this page's defaults — a 500-token answer with a 5× reasoning ratio — the hidden thinking is $0.02/request (77% of the bill), or $10,000/month at 500,000 requests. Your real per-request cost is 4.3× what the visible answer suggests.
Why are reasoning tokens billed if I never see them?
Because the model still generates them — the "thinking" is real computed output, just hidden from the response. Every major reasoning model (o-series, extended-thinking Claude, DeepSeek R-line, Gemini thinking) bills these at the output token rate. You pay for the reasoning whether or not it's shown.
What reasoning ratio should I use?
Check your API usage logs — providers report reasoning/thinking tokens separately. As a rough guide, low-effort or simple tasks run ~1–2× the visible output, medium ~3–6×, and high-effort multi-step reasoning can hit 10–20× or more. It's workload-dependent, so tune the slider to your own numbers rather than trusting a single figure.
What's the cheapest way to cut reasoning cost?
Match the effort to the task: use a lower reasoning-effort setting, or send simple requests to a non-reasoning model entirely. Because reasoning bills at the output rate, trimming hidden tokens saves more per token than trimming input. Capping the reasoning/thinking budget where the provider allows it also helps.
Are these prices current?
o3's input and output rates sync daily from OpenRouter (updated 2026-07-13). Reasoning tokens bill at the output rate, so no separate price is needed — the only input is your reasoning ratio. All math runs client-side with tested code.
Should o3 worry about this?
Yes — at a 5× ratio, hidden reasoning is 77% of your bill, so the thinking costs more than the answer. Effort tuning and model routing move real money here.