What is the maximum thinking budget for Gemini 2.5 Pro?

Gemini 2.5 Pro accepts a thinkingBudget from 128 to 32,768 tokens. Its default is dynamic (-1), where the model decides how much to think. Unlike Flash and Flash-Lite, Pro cannot fully disable thinking — a value of 0 is rejected.

Are Claude thinking tokens billed as output tokens?

Yes. Anthropic bills extended and adaptive thinking tokens at the model's output rate — for Claude Opus 4.8 that is $25 per million. To find the thinking portion of a response, subtract your visible answer tokens from the reported output_tokens.

How do I turn off thinking in Gemini 2.5 Flash?

Set thinkingBudget to 0. That disables thinking on Gemini 2.5 Flash and Flash-Lite, so you pay only for the prompt and the visible answer. Gemini 2.5 Pro does not allow 0 — the smallest budget it accepts is 128 tokens.

Does reasoning_effort=high cost more on GPT-5?

Yes. reasoning_effort is qualitative (minimal, low, medium, high) with no user-set token cap — a higher effort makes the model spend more hidden reasoning tokens, and those are billed at the output rate. This calculator prices each level with a representative token estimate so you can compare them.

What is the minimum budget_tokens value for Claude extended thinking?

On older Claude extended-thinking models (such as Sonnet 4.5), budget_tokens must be at least 1,024 and strictly less than your max_tokens. On current Claude models (Opus 4.8, Haiku 4.5) budget_tokens is removed — thinking is adaptive and you set depth with the effort control instead.

What does thinkingBudget = -1 mean?

It sets dynamic thinking: the model decides how many thinking tokens to spend based on the request. It is convenient but unpredictable for budgeting, because a single hard turn can quietly generate thousands of output-priced tokens. This tool prices -1 with a representative estimate, clearly marked.

Why is my reasoning-model bill so much higher than the prompt suggests?

Because the thinking tokens are invisible in the response but fully billed at the output rate. A thinkingBudget of 8,000 silently adds 8,000 output-priced tokens to every request. On Gemini 2.5 Pro at $10 per million output, that is $0.08 of thinking per call — often 90%+ of the whole bill.

Does this calculator make live API calls or use live prices?

No. Everything runs in your browser against a static, dated price and budget-range table — nothing you type leaves the page. The Anthropic rows are cross-checked against the Anthropic pricing page; Google and OpenAI rows are published list prices, so re-check them against the linked source before relying on a figure.

AI · Developer tools

AI Thinking Budget Calculator

Every reasoning model controls thinking a different way — thinkingBudget, budget_tokens, or reasoning_effort. This tool shows the valid range for your model and turns your setting into one honest number: what it costs per call and per month. No signup, sources cited below.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 18, 2026

Thinking budget & cost

List price · verify

Provider & model

Currency

thinkingBudget

Range: 128 – 32,768
Default: Dynamic (-1)
Turn off (0): Not allowed
Dynamic (-1): Allowed

Range 128–32,768. Default is dynamic (-1). Pro cannot fully disable thinking — 0 is rejected.

Thinking budget

Input (prompt) tokens

Your prompt, system message, tools, and context.

Visible output tokens

The answer you actually see in the response.

Requests per month

How many of these calls you make in a month.

Example setups

Thinking tokens

4,000

Dynamic (-1) · estimate

Cost / request

$0.0443

thinking = $0.04

Thinking share

90.4%

of the per-request bill

Monthly cost

$442.50

latency: Moderate

What you pay for per request

Input: $0.00133%
Thinking: $0.0490%
Visible output: $0.0037%

Per-request cost breakdown

Component	Tokens	Rate	Cost
Input tokens	1,000	$1.25/1M	$0.0013
Thinking tokens (billed as output)	4,000	$10/1M	$0.04
Visible output tokens	300	$10/1M	$0.003
Cost per request			$0.0443

Thinking adds $400.00 per month at 10,000 requests — the invisible line a prompt+answer estimate never shows (thinking tokens are an estimate for this control). Excludes prompt caching, batch discounts, and per-tenant contract pricing.

Cost across settings

Setting	Thinking tokens	Cost / request	Monthly
Minimum128	128	$0.0055	$55.30
Default (dynamic)-1selected	4,000≈	$0.0443	$442.50
High (max)32,768	32,768	$0.3319	$3,319.30

Same prompt, visible output, and request volume; only the thinking setting changes. Values marked ≈ are planning estimates (dynamic or effort settings) — read your real thinking-token usage for exact figures.

Thinking / reasoning tokens are billed at the model's output rate by Google, Anthropic, and OpenAI. Anthropic rates are cross-checked against the Anthropic pricing page; Google and OpenAI rows are published list prices that change without notice. Budget ranges and rates last verified 2026-07-18. Full sources are listed below the calculator.

How it works

Reasoning models think before they answer. That thinking is a stream of hidden tokens — Google calls the control thinkingBudget, Anthropic's older models use budget_tokens, and OpenAI exposes a qualitative reasoning_effort. You almost never see the thinking, but every provider counts and bills it, and crucially bills it at the model's output rate — the higher of the two prices.

The first job of this tool is to state the valid range for the model you pick, because it differs sharply and getting it wrong is an API error:

Gemini 2.5 Pro: 128 – 32,768. Default dynamic (-1). Cannot set 0 (no off).
Gemini 2.5 Flash: 0 – 24,576. 0 = off, -1 = dynamic (default).
Gemini 2.5 Flash-Lite: 512 – 24,576. Off by default; 0 = off, -1 = dynamic.
Claude (legacy budget_tokens): ≥ 1,024 and < max_tokens.
Claude current / OpenAI: qualitative effort levels, no token cap.

The second job is cost. Let Pin and Pout be the per-token input and output prices (the per-million rate ÷ 1,000,000), with I input tokens, V visible output tokens, and B the effective thinking budget:

inputCost = I × Pin
thinkingCost = B × Pout — the hidden line item, billed at the output rate.
outputCost = V × Pout
costPerRequest = inputCost + thinkingCost + outputCost, and monthly is costPerRequest × requests.

Because the thinking and visible tokens are both billed at Pout, the per-request cost can also be written inputCost + (B + V) × Pout. The tool computes it both ways and they agree to the cent — a built-in cross-check on the math. For dynamic (-1) and effort settings there is no fixed token count, so those are priced with a clearly labelled representative estimate; read your real thinking-token usage from the API response for an exact figure.

Worked examples

Gemini 2.5 Pro — thinkingBudget 8,000

Input $1.25/1M, output $10/1M · I=1,000 V=300 B=8,000 · 10,000 req/mo

inputCost = 1,000 × $0.00000125 = $0.00125
thinkingCost = 8,000 × $0.00001 = $0.08
outputCost = 300 × $0.00001 = $0.003
costPerRequest = $0.08425
thinking share = 0.08 ÷ 0.08425 = 95.0%
monthly = $0.08425 × 10,000 = $842.50 — the budget IS the bill

Same request — budget cut to the 128 minimum

Gemini 2.5 Pro · I=1,000 V=300 B=128 · 10,000 req/mo

thinkingCost = 128 × $0.00001 = $0.00128
costPerRequest = $0.00125 + $0.00128 + $0.003 = $0.00553
monthly = $0.00553 × 10,000 = $55.30
8,000 → 128 cuts the month from $842.50 to $55.30 (≈93% saving)

Edge — thinking OFF on Gemini 2.5 Flash

Input $0.30/1M, output $2.50/1M · I=1,000 V=300 B=0 · 10,000 req/mo

inputCost = 1,000 × $0.0000003 = $0.0003
thinkingCost = 0 (budget 0 disables thinking on Flash)
outputCost = 300 × $0.0000025 = $0.00075
costPerRequest = $0.00105 · thinking share = 0%
monthly = $10.50 — on Pro this exact setting is rejected (no off)

Frequently asked questions

Sources & references

Anthropic rates were cross-checked against the Anthropic pricing page on 2026-07-18. Google and OpenAI budget ranges and rows are published values that change without notice — re-check them against the linked source before relying on a figure. No live API calls are made; the price and range table is static and dated.

Related tools

LiveAI

Reasoning Token Cost Calc

Estimate the true cost of reasoning-model API calls by accounting for the hidden reasoning/thinking tokens that o-series, Claude, and Gemini bill at output rates. See per-call and monthly cost in USD and LKR, plus how much more it is than a naive estimate.

Open tool

LiveAI

AI Chatbot Cost Calculator

Estimate the monthly API cost of a multi-turn AI chatbot across Claude, GPT, and Gemini. Models the quadratic context re-sending that single-call calculators miss, with and without prompt caching, in USD and LKR.

Open tool

LiveAI

AI Code Execution Cost

Compare the server-side code execution (code interpreter / sandbox container) fee that Anthropic, OpenAI, and Google Gemini bill separately from model tokens. Enter your monthly runs and container uptime to see each provider's execution fee after free tiers, the effective cost per run, and which is cheapest at your volume.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want to suggest an improvement?

Email me at [email protected] — most fixes ship within 24 hours.