How to compare fairly
List prices per million tokens hide the shape of real workloads: input-heavy tasks (document analysis, RAG) reward cheap input rates and caching, while output-heavy tasks (drafting, code generation) are dominated by the output price. Set the token mix to match your actual requests before comparing. Prices are standard-tier list prices from each provider's official pricing page — batch discounts (typically 50%) and provider-specific caveats are noted on each model's own page.
Not sure how many tokens your prompt is? Paste it into the token counter first.
How API pricing works
Every major provider bills the same way: a rate per million tokens you send (input) and a separate, higher rate per million tokens the model generates (output). Output costs more because generating a token requires a full forward pass through the model, while a prompt is processed in one parallel pass — GPT-5, for example, prices output at roughly 8× its input rate. The shape of your workload therefore matters as much as the model you pick: document analysis and RAG are input-heavy and reward cheap input rates, while drafting and code generation are dominated by the output price. That spread is wide — input rates across the registry currently run from $0.05 (GPT-5 nano) to $10 (Claude Fable 5.1) per million tokens.
Prompt caching, explained
If your requests share a prompt prefix — a system prompt, few-shot examples, a long document you keep asking questions about — the provider can reuse the computed state from last time instead of reprocessing it, and charges a discounted rate for those tokens. On GPT-5 the cached rate is $0.125 per million against $1.25 standard — 90% off. For chat assistants and agents, where the same instructions ride along with every request, caching is routinely the single biggest cost lever. Tick "input is cached" above to see it applied across every model; note that some providers also charge a small storage fee for keeping a cache warm, flagged on the relevant model pages.
Estimating a monthly bill
Per-request numbers are too small to build intuition — $0.0031 for a chat message on GPT-5 sounds like nothing. Volume is what turns token prices into a budget line: the same request at a thousand requests a day is about $93.75 a month. Enter your own requests-per-day above to get a monthly column for every model at once, and price the workload you actually expect — an agent session at a few hundred requests a day is a very different bill from a chat widget at a few thousand.
API pricing vs running open weights
Several models in the table are open-weight: you can pay the provider's API rate, use a third-party host, or run them on your own hardware for the cost of electricity. Self-hosting wins when you have steady volume, privacy constraints, or hardware already on hand — and loses when a workload is spiky or small enough that list prices round to pocket change. If you're weighing that trade-off, the VRAM calculator tells you whether a given model fits your GPU at all, and each open model's cost page links to its local-requirements twin.
How we source and verify prices
Every figure on this page comes from one shared model registry, populated only from provider's official pricing pages — never from blogs, news coverage or other calculators — with each entry's source URL and verification date recorded. The registry is re-verified monthly (last pass: 2026-09-03) and an automated test fails our build if any entry goes stale. The methodology page documents the process and every formula this site uses.
Frequently asked questions
Which LLM API is cheapest in 2026?
Of the 29 models we track, GPT-5 nano has the lowest list price at $0.05 per million input tokens and $0.4 per million output. But cheapest-per-token is not cheapest-per-task: a stronger model that succeeds in one attempt often beats a cheap one that needs three. Sort the table above by your own workload before deciding.
Why are output tokens more expensive than input tokens?
Generating a token requires a full forward pass through the model, while input tokens are processed in parallel much more efficiently. Most providers price output 4–6× above input, so long responses dominate cost in chat-style workloads.
What is cached input and when does it apply?
If your requests repeat the same prompt prefix (a system prompt, few-shot examples, a document), providers can reuse the computed state and charge a fraction of the normal input rate — 90% off at OpenAI, Anthropic and Mistral, and up to 98% off at DeepSeek. The caching toggle above prices your input at each provider's cached rate; note Google also charges a small per-hour storage fee for its cache.
How many tokens is a typical request?
A short chat message with a modest system prompt is roughly 500 input and 250 output tokens. A 10-page document is about 7,000 tokens. Agent-style coding sessions routinely reach tens of thousands of input tokens per request because of accumulated context. Use the presets above, or count your own text with the token counter.
What is a batch API and when is the discount worth it?
Most providers run an asynchronous batch tier: you submit a file of requests, results come back within a target window (typically 24 hours), and both input and output are billed at a steep discount — commonly half the standard rate. It suits evaluations, backfills, classification runs and anything else that is not interactive. The prices in the table above are standard-tier; halve them as a rough batch estimate and confirm on the provider's pricing page.
Do these prices include free tiers or rate limits?
No — the table shows standard pay-as-you-go list prices only. Free tiers, promotional credits, rate limits and enterprise discounts vary by provider and change frequently, so we deliberately leave them out rather than show numbers that may be stale. Treat the table as the like-for-like baseline and check the provider's own page for whatever concession applies to your account.
How often are these prices updated?
Every price on this page comes from a single model registry that is re-verified against each provider's official pricing page on a monthly cycle — last verified 2026-09-03. An automated staleness test fails our build if any entry goes more than sixty days without re-verification, so an outdated price cannot silently stay on the site.
Per-model cost guides
OpenAI: GPT-5.6 Sol · GPT-5.6 Terra · GPT-5.6 Luna · GPT-5.4 · GPT-5 · GPT-5 mini · GPT-5 nano · GPT-4o · GPT-4o mini · o3 · o4-mini
Anthropic: Claude Fable 5.1 · Claude Fable 5 · Claude Opus 5 · Claude Sonnet 5 · Claude Haiku 4.5 · Claude Opus 4.8
Google: Gemini 3.8 Flash · Gemini 3.7 Flash · Gemini 3.6 Flash · Gemini 3.1 Pro Preview · Gemini 2.5 Pro · Gemini 2.5 Flash · Gemini 2.5 Flash-Lite
DeepSeek: DeepSeek V4 Flash · DeepSeek V4 Pro
Mistral: Mistral Medium 3.5 · Mistral Large 3 · Mistral Small 4
Head-to-head comparisons
Claude Fable 5 vs GPT-5.6 Sol · GPT-5 vs Claude Sonnet 5 · Gemini 2.5 Pro vs GPT-4o · DeepSeek V4 Flash vs GPT-4o mini · Gemini 2.5 Flash vs GPT-5 mini
Last updated 2026-09-03. Prices verified against official provider pricing pages; see the methodology.