Are reasoning tokens billed?

Yes. OpenAI, Anthropic and Google all bill reasoning or thinking tokens at the model’s output price, even though you do not see them in the answer.

Where do your reasoning-token numbers come from?

They are our estimates for short classification calls, deliberately on the low side. Providers say reasoning is adaptive and only report the real count after each call. Measure yours from the usage data your API returns and use that number in the audit.

Will lower effort hurt accuracy?

It can, on hard inputs. On a narrow yes/no or extraction call it may make no difference, but the only way to know is to compare answers at both levels on your own inputs.

Can I lower effort and route to a smaller model?

Yes. They stack: the calls the small model is unsure about go to your current model, and those can run at a lower effort too.

Guide · prices checked 2026-11-03

What reasoning effort costs per call

Our reasoning-token estimates by level

For a short classification call. Field extraction uses 2.5× these. They are estimates, not provider figures:

Effort level	Est. reasoning tokens per call
none	0
off	0
minimal	10
low	40
medium	120
high	320
xhigh	800
max	1600

Not every model offers every level; each provider documents its own levels and default. Read the real count from your API’s usage data after a call (for example usage.output_tokens_details.reasoning_tokens on OpenAI).

Cost per 1,000 calls: default vs lowest effort

500 input and 5 visible output tokens per call, list prices checked 2026-11-03. The last two columns are a hybrid where Levanto Sage 2 keeps 70% of calls (a scenario) and the rest go to the current model.

Model	Default effort	Lowest effort	Saving	Hybrid, default	Hybrid, lowest
Claude Fable 5.1	$21.25 (high)	$7.25 (low)	66%	$6.40	$2.20
GPT-6 Astra	$11.25 (medium)	$5.75 (minimal)	49%	$3.40	$1.75
Claude Opus 5.5	$4.50 (medium)	$2.90 (low)	36%	$1.38	$0.897
Claude Sonnet 5.5	$4.25 (high)	$1.45 (low)	66%	$1.30	$0.462
GPT-6.1 Sol	$2.25 (medium)	$1.45 (low)	36%	$0.702	$0.462
Gemini 4 Argon	$2.25 (medium)	$1.45 (low)	36%	$0.702	$0.462
GPT-6.1 Terra	$1.13 (medium)	$0.525 (none)	53%	$0.365	$0.185
Gemini 3.8 Flash	$0.844 (medium)	$0.544 (low)	36%	$0.280	$0.190
GPT-5 mini	$0.375 (medium)	$0.155 (minimal)	59%	$0.140	$0.073

Estimates, not bills. Levanto Sage 2 on its own costs about $0.027 per 1,000 such calls. Batch and cached-input discounts excluded. Intro prices are noted on each model’s page.

When lowering effort beats switching

Try lower effort first when the call is already on a model you trust, the saving in the table is large, and you would rather not add a second vendor.
Switch or route when even the lowest effort is expensive at your volume, or the model’s lowest level still reasons (several stop at low instead of none).
Do both when volume is high: the hybrid’s fallback calls get cheaper too.

Either way, test on your own inputs: compare answers at the two effort levels, or between your model and a smaller one, before changing production.

Measure it on your calls

The free cutover audit lets you pick your model and effort level, or type your measured reasoning tokens, and projects the monthly cost with and without routing. Is my task too simple? · Cost by model · Routing calculator

Run the free audit