LLM API pricing (October 2026): every model's cost per million tokens
TL;DR: below is a reference table of current LLM API list prices — input, output, and (where the provider publishes it) cached-input — per 1 million tokens, in USD. Every row was checked against the provider's own pricing page in early October 2026. Prices move fast and often downward, so treat this as a dated snapshot and confirm against the provider before you budget. The practical takeaway is at the bottom: the gap between the frontier tier and the cheap tier is 10x to 100x, and most tasks do not need the frontier.
How to read this
- Input / Output are the list price per 1M tokens sent / generated.
- Cached input is what a cache hit costs (Anthropic cache read, OpenAI cached input, Gemini cached content). It bills far below the normal input rate — usually about a tenth — which is why caching re-sent context is the single biggest lever on an agent's bill. We cover the mechanics, and a trap it creates, in The caching trap.
- A dash (—) means the provider does not publish a separate cached rate for that model.
OpenAI
| Model | Input | Output | Cached input |
|---|
| GPT-5 | $1.25 | $10.00 | $0.125 |
| GPT-5 nano | $0.05 | $0.40 | $0.005 |
| GPT-4.1 | $2.00 | $8.00 | — |
| GPT-4.1 mini | $0.40 | $1.60 | — |
| GPT-4.1 nano | $0.10 | $0.40 | — |
| GPT-4o | $2.50 | $10.00 | $1.25 |
| GPT-4o mini | $0.15 | $0.60 | $0.075 |
| o1 | $15.00 | $60.00 | — |
| o3-mini | $1.10 | $4.40 | — |
Anthropic (Claude)
| Model | Input | Output | Cached input |
|---|
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 |
| Claude Opus 5 | $5.00 | $25.00 | $0.50 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.20 |
| Claude Sonnet 4.5 | $3.00 | $15.00 | $0.30 |
| Claude Haiku 4.5 | $1.00 | $5.00 | — |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 |
Google Gemini
| Model | Input | Output | Cached input |
|---|
| Gemini 2.5 Pro | $1.25 | $10.00 | $0.125 |
| Gemini 2.5 Flash | $0.30 | $2.50 | $0.03 |
| Gemini 2.5 Flash-Lite | $0.10 | $0.40 | $0.01 |
| Gemini 2.0 Flash | $0.10 | $0.40 | — |
xAI Grok
| Model | Input | Output | Cached input |
|---|
| Grok 4 | $2.00 | $6.00 | $0.50 |
| Grok 4.3 | $1.25 | $2.50 | $0.20 |
DeepSeek
| Model | Input | Output | Cached input |
|---|
| DeepSeek Flash | $0.30 | $1.20 | $0.006 |
| DeepSeek V4 Pro | $1.32 | $3.96 | $0.044 |
DeepSeek lists a ~50% off-peak discount in a daily UTC window; the figures above are peak rates.
Mistral, Cohere, Perplexity
| Model | Input | Output |
|---|
| Mistral Large 3 | $0.50 | $1.50 |
| Mistral Medium 3.5 | $1.50 | $7.50 |
| Mistral Small 4 | $0.15 | $0.60 |
| Cohere Command R+ | $2.50 | $10.00 |
| Cohere Command R | $0.15 | $0.60 |
| Perplexity Sonar | $1.00 | $1.00 |
| Perplexity Sonar Pro | $3.00 | $15.00 |
Open-weight (hosted)
| Model | Input | Output |
|---|
| gpt-oss-120b | $0.15 | $0.60 |
| gpt-oss-20b | $0.075 | $0.30 |
Open-weight models are priced by whoever hosts them, so rates vary by inference provider (Groq, Together, Fireworks, your own GPUs). The figures above are representative hosted rates, not a single canonical price.
Embeddings
| Model | Price (input) |
|---|
| text-embedding-3-small | $0.02 |
| text-embedding-3-large | $0.13 |
| voyage-3-lite | $0.02 |
Legacy models — and what to use instead
If your code still points at one of these, you are almost certainly overpaying for a model a newer one beats on both price and quality. This is the most common finding in our teardowns — the model string gets written once and never revisited (see MetaGPT's stale config default and the Swarm name-extractor).
| Legacy model | Its price (in/out) | Modern replacement | Replacement price |
|---|
| gpt-4-turbo | $10 / $30 | GPT-4o | $2.50 / $10 |
| gpt-4 | $30 / $60 | GPT-4o or GPT-5 | $2.50 / $10 · $1.25 / $10 |
| gpt-3.5-turbo | $0.50 / $1.50 | GPT-4o mini | $0.15 / $0.60 |
| claude-3-opus | $15 / $75 | Claude Opus 5.5 | $4 / $20 |
| claude-3.5-sonnet | $3 / $15 | Claude Sonnet 5.5 | $2 / $10 |
| text-embedding-ada-002 | $0.10 | text-embedding-3-small | $0.02 |
In every row the replacement is both newer and cheaper. Upgrading is usually a one-line change — though, as always, a model swap is a candidate until you verify it holds quality on your task (here is how to prove that).
The one thing to take away
The spread between tiers is enormous: a frontier model runs $2–$20 per 1M output tokens, while the cheap tier (GPT-5 nano, Gemini Flash-Lite, Claude Haiku, the open-weight models) runs cents. A task like classification, routing, extraction, or a short label almost never needs the frontier, and the price gap is routinely 10x to 100x. The skill is matching each call to the cheapest model that still holds quality — which is exactly what the complete guide to LLM cost optimization walks through.
Methodology: list prices per 1M tokens, verified against each provider's official pricing page in early October 2026. Prices change frequently — this is a dated snapshot, not a live feed; confirm with the provider before budgeting. erabot keeps an internally-verified pricing table current (199 tests guard it) and uses it to estimate the dollar cost of the LLM calls in your code. Scan your repo free at erabot.ai.