LLM Token & Cost Calculator
Compare what one request costs across 15 current models, then project it to a monthly bill. Runs in your browser.
Prices checked weekly against models.dev (last 20 July 2026). USD per 1M tokens, standard tier.
Advanced: measure your own prompt
1 token is about 4 characters of English.
Every model, one request
| Model | Input cost | Output cost | Per call | Per month |
|---|
The table above lies a little
It assumes every model emits the output tokens you set. They do not: reasoning models emit far more for the same task. This corrects your per-call cost using how many tokens each model actually burns, so a cheap per-token model that thinks a lot can cost more than a pricey one that does not.
| Model | Tokens vs answer | Real output | Real cost / call |
|---|
- Multipliers are measured on a hard reasoning benchmark at maximum effort, so this is the ceiling. On an easy task, or a lower effort setting, the gap is smaller.
- Effort settings alone swing token use about 10x on the same model: GPT-5.6 Terra runs 96M at max but 10M at medium.
- Reasoning vs non-reasoning models differ 60 to 80x in tokens on identical tasks. That gap, not the per-token price, usually decides the bill.
Token-per-task figures: Artificial Analysis Intelligence Index v4.1, retrieved 19 July 2026.
Claude token calculator
Four models, a 10x range: Haiku 4.5 ($1/$5), Sonnet 5 ($2/$10, intro rate to 31 Aug 2026), Opus 4.8 ($5/$25), Fable 5 ($10/$50). Opus 4.7 onward count ~30% more tokens for the same text, so a migration costs more than the rate delta suggests.
OpenAI GPT token calculator
The GPT-5.6 family (9 July 2026): Sol ($5/$30), Terra ($2.50/$15), Luna ($1/$6), all 1M-context. GPT-5.4 mini is the budget pick ($0.75/$4.50). Cached input bills at 10% of the input rate, which changes chat math a lot.
How the numbers are made
- Tokens estimated at 4 characters each; real tokenizers differ, so these are budgeting numbers.
- Standard pay-as-you-go prices, re-checked weekly against models.dev from a July 2026 provider-page baseline.
- Llama has no first-party API; those rows use DeepInfra and vary by host.
Prefactor measures this per run in production, not per estimate. Token usage tracking →
Frequently asked questions
How do I know my output token count?
You often cannot in advance, and output usually drives the bill. Pick the preset closest to your workload, or set a response length you expect. For open-ended jobs like "build an app", cost is many calls of accumulating context, not one, so it is highly variable.
Why do output tokens cost more?
Generating is more expensive than reading, so providers price output at roughly 5 to 6 times the input rate. That is why the split above matters.
Are these prices current?
Re-checked weekly against models.dev; last checked 20 July 2026, from a July 2026 provider-page baseline. Standard tier: batch, cached-input, and negotiated rates all land lower.
Does my text leave the browser?
No. Everything is computed client-side. Nothing you paste is sent anywhere.