What does it really cost to run an AI model?
Live API prices per million tokens across every major lab — verified against provider docs, re-checked on a schedule, and honest about what the numbers don't tell you.
Pricing table
USD per 1 million tokens. Click any column header to sort. "Blended" = 1M input + 1M output tokens, a simple like-for-like scale.
| Provider | Model | Input / 1M | Output / 1M | Blended | Status |
|---|---|---|---|---|---|
| Loading data.json… | |||||
List prices only. Your invoice also depends on caching, batch discounts, peak windows, thinking tokens, and tokenizer differences — see cost per task. Prices move fast; check the provider doc linked in Method before signing anything.
Model picker
Answer three questions and get a cheapest-capable recommendation with an estimated monthly bill — computed from the live table above.
How the picker thinks
- It only recommends verified prices — models flagged unverified in the table never appear as picks.
- Coding favors strong reasoning-per-dollar: DeepSeek V4.1 Flash for tight budgets, Grok 4.7 in the middle, Claude Opus 5.5 when quality wins.
- Support chat is a volume game: Gemini 2.5 Flash-Lite handles thousands of short turns for cents, and Claude Sonnet 5 takes the quality-sensitive tier.
- Writing rewards cheap output tokens: GPT-6 Luna at the low end, Gemini 3.8 Flash in the middle.
- Research needs long context: Gemini 3.1 Pro Preview (1M context) balanced, Claude Fable 5.1 when it must be the best.
- Agentic automation burns output tokens on every tool call: Grok 4.20 or 4.7 for cost control, Claude Opus 5.5 for reliability.
Estimate assumes output tokens ≈ 25% of input tokens. Real agent loops can be far more output-heavy — measure on your own workload.
Cost per task: the honest metric
Token price is the sticker. Task price is the bill. A model that charges half per token can still cost you more if it needs twice as many tokens to finish the job.
Worked example. A 100K-token document plus a 10K-token answer on Grok 4.7: 0.10 × $2.00 + 0.01 × $6.00 = $0.26. Run that a thousand times a month and you're at $260 — the table lets you re-run the same math for any model in seconds.
Why the cheapest token isn't the cheapest answer
- Verbose models tax you on output. One published benchmark comparison found Gemini 3.8 Flash generates roughly 30% more output tokens per task than its predecessor — raising cost-per-task by about 40% at the same list price. Output is where the money goes.
- Reasoning burns tokens too. Thinking models charge their chain-of-thought at output rates. A "cheap" reasoner that thinks long can outspend a pricier model that answers directly.
- Tokenizers differ. The same English paragraph can tokenize into very different token counts on different models — Anthropic notes Claude 4.7+ produces about 30% more tokens for equivalent text. Compare measured tokens, not vibes.
- Context cliffs. xAI doubles the whole request's rate past a 200K-token prompt. Gemini 3.1 Pro doubles above 200K prompt tokens. One long document can flip your cheapest pick.
Five levers that actually cut the bill
- Prompt caching. Repeated system prompts and documents billed as cache hits: Grok cached input is $0.50 vs $2.00 fresh; DeepSeek cache hits are ~2% of input rate. Cache-hit rates above 50% are common in production.
- Batch APIs. OpenAI, Google, and Anthropic discount async batch requests ~50% when you don't need an instant answer.
- Off-peak scheduling. DeepSeek halves every rate outside weekday UTC peak hours (01:00–04:00, 06:00–10:00) — scheduled jobs cost half just by moving the clock.
- Routing. Serve routine traffic on a cheap model (Gemini 2.5 Flash-Lite, Llama 3.1 8B) and escalate only failures to a frontier model. The newest expensive model is not automatically the cheapest system.
- Output budgets. Cap max tokens and ask for concise formats. Every unnecessary sentence is billed at output rates.
Price moves
The rate card never sits still. Every notable change we confirm, logged with its source.
Loading…
Method & verification policy
How prices are verified
Every row carries a last verified date and a source URL. A row is marked verified only when confirmed against a provider pricing document or a reputable pricing index checked in September 2026. Conflicting or stale sources get the row marked unverified — the figure stays for orientation, but don't treat it as authoritative.
How the table stays fresh (the anti-rot loop)
Pricing tables rot within weeks — staleness is this site's biggest risk. This lab runs refresh-data.py, an open Python script that re-checks provider pricing pages, diffs what it finds against the table, and updates verification timestamps. It's designed to run on a schedule, so the "last verified" badge above is a live promise, not a launch-day stamp. The raw data file (data.json) is published alongside the site for anyone to audit.
What list prices don't include
Caching behavior, batch discounts, peak/off-peak windows, thinking-token billing, long-context price cliffs, tokenizer differences, platform markups (Bedrock, Vertex, Azure), and taxes. Two models with identical table rows can still produce very different invoices. Always re-check the provider doc before committing spend.
Affiliate & sponsorship disclosure
This page currently carries no affiliate links and no sponsored placements. If that changes, placements will be clearly labeled and will never change model ordering. Model ordering is by the numbers, not by payment.
Provider pricing documents we check
OpenAI: developers.openai.com/api/docs/pricing · Anthropic: platform.claude.com/docs/en/about-claude/pricing · Google: ai.google.dev/gemini-api/docs/pricing · xAI: xAI developer docs · Groq: groq.com/pricing · DeepSeek: deepseek.ai/pricing · Mistral: docs.mistral.ai/platform/pricing. All trademarks belong to their owners and appear here only for identification.