LLM Price Comparison
Compare the cost of one request across GPT, Claude, Gemini, Grok, DeepSeek and Mistral models, ranked cheapest to priciest. Free, runs in your browser.
LLM Price Comparison
| Model | Cost / request | Relative |
|---|
How to Use the LLM Price Comparison
Enter your token counts
Put in the input and output tokens for a typical request. Not sure? Measure the input with our Token Counter first.
Read the ranked table
Every model is priced for your exact request and sorted cheapest to priciest, so the best-value option is always at the top. Where a vendor charges more for long prompts, the higher rate is used once your input passes its threshold, and the row says so; a model whose context window is too small for the request is flagged and listed after every model that can take it.
Check the "relative" column
It shows how many times more expensive each model is than the cheapest — a fast way to see whether a premium model costs 2× or 20× more for your workload.
Adjust the input/output mix
Output tokens are priced higher than input. Raise the output figure to model a chatty, generation-heavy task and watch the ranking shift.
Comparing LLM Prices the Right Way
A headline price tells you almost nothing
Provider pricing pages quote a per-million-token input rate and a separate, higher output rate — but the number that matters is what your request costs, and that depends entirely on your input/output mix. A model with a cheap input rate but an expensive output rate can be the bargain for a summarisation task and the worst choice for a creative one. This tool removes the guesswork: you give it one realistic request, and it prices that exact request across every model, then ranks them. The cheapest option for a long-context, short-answer job is often a completely different model than the cheapest for a short-prompt, long-answer job.
The relative column is where the insight usually lands. Seeing that a flagship model is "14× the cheapest" for your workload reframes the decision: is the quality difference worth fourteen times the cost at your volume? Sometimes emphatically yes; often a mid-tier model does the job for a fraction of the price. Comparing on real numbers, rather than vibes or headline rates, is how teams avoid quietly overpaying.
"Don't compare price tags — compare what your actual request costs on each. The ranking can flip the moment your output-to-input ratio changes."
What the comparison leaves out (on purpose)
These figures are standard list rates, and they deliberately exclude the discounts that vary by usage: cached-input pricing (cheap when you reuse a long shared prompt), batch processing (often around half price for non-urgent jobs), and any negotiated or volume tier. They also can't capture quality — a cheaper model that needs longer prompts or more retries may not be cheaper in practice. Use the ranking to shortlist, confirm the live rates on the provider's page, and weigh cost against the quality your task actually needs. For a single model's full monthly bill, pair this with our LLM Cost Calculator.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, whose models are compared on this page. Readers should weigh that interest.
The cheapest model depends on your input-output mix, not the headline rate
Models charge a separate input and output rate — output is almost always the pricier of the two.
The cheapest model for your job depends on your input/output mix, not the headline rate.
For the same request, the dearest models on this card cost over a hundred times what the cheapest do.
Prices are quoted per million tokens, which hides how a high-volume workload adds up.
Gemini 2.5 Pro charges a higher rate once a prompt exceeds 200,000 input tokens.
Cached input is often billed at a tenth of the input price or less — 2.5% on Claude Fable 5.1, per Anthropic — though some models discount far less.
Batch APIs typically offer around 50% off for jobs you don't need answered immediately.
A cheaper but weaker model isn't always cheaper — retries and longer prompts add up.
Model prices drift over time, so any comparison needs an "as at" date and periodic refresh.
This comparison runs entirely in your browser — your numbers are never uploaded.
Frequently Asked Questions
- There's no single answer — it depends on your request. Enter your input and output tokens and the tool ranks every model for that exact workload. The cheapest for a long-input, short-output job is often different from the cheapest for a short-input, long-output job.
- It shows how many times more expensive each model is than the cheapest option for your request. A value of 5.00× means that model costs five times what the top-ranked model costs for the same input and output tokens.
- Generating tokens costs more than reading them — each output token needs a full pass through the model, while input is processed more efficiently. That's why a generation-heavy task can be much more expensive than its input size suggests.
- They're list rates gathered on the date shown under the table and refreshed periodically. Providers change pricing without notice, so confirm the live rate on the provider's page before committing to a budget.
- No — it uses standard list rates so the comparison is apples-to-apples. Batch processing and cached-input pricing can cut real costs significantly, but they depend on your usage pattern, so they're left out of the headline ranking.
- This tool ranks the per-request cost across all models so you can choose one. The LLM Cost Calculator focuses on a single model and projects a full monthly bill from your request volume. Use this to pick, that to budget.
- Not necessarily. The cheapest model may need longer prompts, produce weaker results, or require more retries — which can erase the saving. Use the ranking to shortlist, then weigh cost against the quality your task actually needs.
- Several, each with its own threshold on the input of a single request: OpenAI's GPT-6, GPT-5.6, GPT-5.5 and GPT-5.4 (not the mini and nano) above 272K input tokens, Google's Gemini 3.1 Pro and 2.5 Pro above 200K, xAI's Grok models at 200K or more, and Claude Haiku 5.5 above 100K. Past the threshold the whole request is billed at the higher rate, so the tool switches that row to it and labels the tier.
- No. All calculation happens in your browser. Your token counts are never sent to any server or third party, and nothing is stored.
- Completely free, with no account or sign-up, and no limit on use. It runs in your browser and collects no data.
Method & sources
How it computes
Cost of one request = input tokens ÷ 1e6 × input rate + output tokens ÷ 1e6 × output rate, at each vendor's list rate from the shared rate card; past a vendor's long-prompt threshold the whole request bills at the higher rate. Ranked cheapest first among models whose context window holds input + output, then the rest, flagged; relative = cost ÷ the cheapest fitting model's cost (LlmCostKernel).
What this tool implements
- Rates are each vendor's own list price, read from its pricing or model page and recorded per row with the page and the date; the date under the table is the OLDEST row's check, with the newest beside it, so it is a floor rather than a boast.
- Long-prompt tiers are applied as the vendors bill them, on the input tokens of a single request and to the whole request: OpenAI above 272K (GPT-6, GPT-5.4–5.6), Google above 200K (Gemini 3.1 Pro, 2.5 Pro), xAI at 200K or more, Anthropic's Claude Haiku 5.5 above 100,000. The row is labelled with the tier.
- A model whose context window cannot hold input + output is never ranked as the cheapest: it is listed after every model that fits and flagged.
- DeepSeek rows are its off-peak rate, the one charged 17 of 24 hours on weekdays and all weekend, with the peak rate in the row's note. Introductory or promotional prices carry their end date in the note.
- Excluded on purpose and said so: batch and cached-input discounts, regional or priority-processing surcharges, taxes.
- Corrected 2026-10-26: xAI's tier read ">200K" for a rate that starts at 200K; an overflowing request was ranked, even first, with no warning; "flagships cost 10–50× budget models" (over 100× on the current card); "cached input cuts up to ~90%" (97.5% on Claude Fable 5.1); the tier FAQ named only Gemini 2.5 Pro. The card itself gained the current Anthropic, OpenAI, Google and xAI models and GPT-5.5 / GPT-5.4 long-prompt tiers.
Sources
- Anthropic. Claude API pricing (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 with its 100,000-token prompt-length pricing; cache hits at 2.5% of input on Fable 5.1), read 2026-10-26: https://platform.claude.com/docs/en/about-claude/pricing
- Anthropic. Models overview (current lineup, 1M-token context windows), read 2026-10-26: https://platform.claude.com/docs/en/models/overview
- OpenAI. API pricing (GPT-6 Astra / 6.1 Sol / 6 Luna, GPT-5.6, GPT-5.5, GPT-5.4, GPT-4o), read 2026-10-26: https://developers.openai.com/api/docs/pricing
- OpenAI. GPT-6.1 Sol model page (1,050,000-token window; above 272K input tokens 2x input and 1.5x output for the full request), read 2026-10-26: https://developers.openai.com/api/docs/models/gpt-6.1-sol
- OpenAI. GPT-5.5 model page (above 272K input tokens 2x input and 1.5x output), read 2026-10-26: https://developers.openai.com/api/docs/models/gpt-5.5
- Google AI for Developers. Gemini Developer API pricing (Gemini 4 Argon and 3.8 Flash introductory rates; Gemini 3.1 Pro and 2.5 Pro ≤200k / >200k), read 2026-10-26: https://ai.google.dev/gemini-api/docs/pricing
- Google AI for Developers. Gemini deprecations (2.5 models limited to existing users; gemini-3.5-flash routed to gemini-3.6-flash), read 2026-10-26: https://ai.google.dev/gemini-api/docs/deprecations
- xAI. Grok models and pricing (grok-4.7, 4.5, 4.3; long context ≥ 200k tokens), read 2026-10-26: https://docs.x.ai/developers/pricing
Abridged — the full review record for this tool runs longer than the list above.
Pick up where you left off
Stored only in this browser — never sent to our servers.