LLM API Pricing, October 2026: Cost per Million Tokens

This page lists October 2026 API prices for LLMs, per million tokens, read straight from vendor and host pages. The spread is wide. The most expensive flagship in the table charges $50 per million output tokens; the cheapest small models charge $0.10 to $0.15. About fifty models sit between those ends, from OpenAI, Anthropic, Google, DeepSeek, xAI, Mistral and the main open-model hosts. Each row links to the page its number came from, so you can re-check a figure before it goes into a budget.

Prices checked on 30 October 2026 against the vendor pages linked in each row.

The price picture

The chart puts input and output prices for twelve widely used models on one linear scale, sorted by output price. The output bar drives the cost of chat and generation. The input bar drives summarisation, extraction and any prompt that carries long documents.

Three things stand out. First, GPT-6 Astra tops the chart at $10 in and $50 out, the same pair as Claude Fable 5.1 in the table below. Anthropic's main flagship, Claude Opus 5.5, costs $4 and $20. Second, the middle of the market has settled on $2 input and $10 output. Gemini 4 Argon (until 30 November), Claude Sonnet 5.5 and GPT-6.1 Sol all list exactly that pair. Gemini 3.1 Pro Preview is $2 and $12 for prompts up to 200k tokens. Third, DeepSeek V4 Pro, Gemini 3.8 Flash and Mistral Large 4 all price output between $2 and $4.

Output tokens cost 3 to 6 times more than input tokens at every vendor in the chart. Ministral 3, in the budget chart below, is the exception, with input and output at the same price.

The budget tier

Below $1 per million output tokens, the gap between vendors shrinks to cents. Open-weight models on third-party hosts compete directly with the vendors' own small models.

GPT-6 Luna and Claude Haiku 5.5 list the same pair, $0.10 in and $0.50 out. Haiku moves to $0.50 and $2.50 once a prompt exceeds 100k tokens. The Groq-hosted gpt-oss models and the Llama 3.3 70B route on OpenRouter land in the same band. OpenRouter's headline price for gpt-oss-120B is lower still, at $0.03 in and $0.15 out from its cheapest provider. That figure depends on which provider the router picks and is not a vendor list price. It therefore appears in the table with the host named, and stays out of the chart.

Full price table

One row per model as listed on the vendor page. Prices are US dollars per million tokens, standard tier, synchronous requests, as of 30 October 2026. "Not shown" means the vendor page fetched on that date did not display the figure. We left those gaps open rather than fill them from memory. Where a cached-input or batch figure comes from OpenRouter rather than the vendor, the cell and the Source column say so.

Notes on reading the table:

• OpenAI's GPT-6 Astra, GPT-5.6 Luna and GPT-5.5 pages add a surcharge when a request's input exceeds 272K tokens. It is 2x on input and cached input and 1.5x on output, applied to the whole request. Astra's Fast mode doubles the price; Flex mode is billed at the batch rate. GPT-5.6 cache writes are billed at 1.25x the uncached input rate.
• Anthropic bills cache writes separately from cache reads; the two write figures are for the 5-minute and 1-hour cache. Fast mode for Opus 5.5 is $8 and $40 and is not available with batch. Web search is $10 per 1,000 searches.
• Google's caching column is the price of a cached token plus a storage charge per million tokens per hour. Gemini 4 Argon has implicit caching only and no storage charge on the page. Priority tier is 1.8x the standard price; Flex tier is billed at the batch price.
• DeepSeek defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese holidays. Off-peak is half of peak. Thinking mode does not change the per-token price.
• Mistral's two pages disagree on Mistral Large 4: docs.mistral.ai lists $0.70 and $2.10, while mistral.ai/en/pricing shows $0.50 and $1.50. Both were fetched on 30 October. The table uses the docs figure; check the one you are billed against.
• xAI applies its long-context rate from 200k prompt tokens.
• Context windows for the other Gemini models are not in this table because the pricing page does not state them. The Argon figure is from Google's models page.

Announced changes

Several of the figures above carry an expiry date on the vendor's own page.

The Gemini increases are the ones to put in a calendar. A workload on Gemini 3.8 Flash costs twice as much on 1 January 2027 as it does today unless something else changes before then.

How to read a price

A model's price has four parts: input, output, cached input and batch. The lowest headline pair often does not mean the lowest bill in production.

Input is what you pay for every token you send: system prompt, conversation history, documents, tool definitions, tool results.

Output is what you pay for every token the model writes, including hidden reasoning tokens at vendors that bill them.

Cached input is the reduced rate for the part of a prompt the vendor has seen recently and kept in memory. Anthropic and Google also charge for writing to or storing that cache. OpenAI's pages mention a write charge only for GPT-5.6, at 1.25x the input rate.

Batch is the rate for asynchronous requests that the vendor may take up to 24 hours to answer. OpenAI, Anthropic, Google and Mistral give 50% off; xAI gives 20%.

A worked example shows how the four parts interact. Take a product that sends a 2,000-token prompt and gets a 500-token answer. Of the prompt, 1,500 tokens are a fixed system prompt that can be cached, and 500 tokens are the user's fresh input. The product makes 1,000,000 calls a month. That adds up to 1.5 billion cacheable input tokens, 0.5 billion fresh input tokens and 0.5 billion output tokens. The table assumes every cacheable token is a cache hit. It leaves out Anthropic's cache-write charge and Google's storage charge, which depend on how often the cache is refreshed.

Two observations follow. First, output is about 80% of the bill at every price level (79% to 81%), even with three quarters of the prompt cached. Output tokens cost 5x input, and the cache only touches input. A shorter answer or a model that stops sooner saves more than a better cache. Second, for this shape of traffic the batch discount saves more dollars than the cache does. A workload that can wait a few hours is better batched first and cached second. Anthropic states that the two discounts stack. Google bills cache hits inside a batch at the standard caching rate. OpenAI's pages do not state the cached-input rate inside a batch.

Our prompt caching guide explains how each cache works, including how long entries live and what counts as a hit. Batch APIs Compared sets out batch limits, turnaround times and the queue sizes each vendor allows.

Embedding prices

Vendors bill embedding models on input tokens only. The prices below could be read on vendor pages on 30 October 2026.

The range is narrow, $0.02 to $0.15. For most retrieval pipelines, the embedding bill is a rounding error next to the generation bill. Embeddings get expensive when you re-index a large corpus every time the model version changes.

Where prices stood in mid-2025

A reference point helps when reading the current table. Sixteen months before this snapshot, OpenAI cut the price of o3 by 80%, and Simon Willison summarised the market in one tweet.

The mid-tier pair has stayed close to $2 and $10 since then. The flagship tier moved in both directions. Claude Opus went from $15 and $75 to $4 and $20 for Opus 5.5. OpenAI's current flagship, GPT-6 Astra, lists $10 and $50, against the $2 and $8 of o3 in that tweet. Gemini 2.5 Pro still lists the same $1.25 and $10 it did in June 2025. LLM Prices 2023–2026 has the year-by-year picture.

What's missing from this snapshot

Some vendors are absent because their pages did not render a usable price on 30 October 2026. This snapshot prints only figures read from a page. Amazon Nova prices on the Bedrock pricing page appeared as headings with no values. The only verifiable statement there is that Amazon models are available for batch inference at 50% below on-demand. Cohere's pricing page showed Model Vault hourly rates and legacy Command prices, but no per-token Embed or Rerank price. Google's Gemini embedding price did not appear in four reads of the pricing page. Groq lists Llama 3.3 70B and 3.1 8B as "contact sales". Together's serverless table did not include Llama 4 Maverick or Scout. Cerebras per-token prices were not reachable. OpenAI's public pricing page now shows seat-based plans only, so OpenAI token prices here come from the developer model pages. Some cells come from OpenRouter's OpenAI provider listings instead of OpenAI: the cached-input figure for GPT-6.1 Terra and the batch figures for GPT-6.1 Sol, Terra and GPT-6 Luna. Treat those cells as provisional.

What to do next

Copy the three or four rows for the models you call into your own sheet, with the source links. Re-check them on the first of each month; the Gemini rows change on 1 December and 1 January. Then run the worked example above with your own prompt shape: average prompt length, how much of it is fixed, average output length, calls per month. If output is more than 60% of the bill, look at answer length and stop sequences before model choice. If the traffic can wait, price the batch rate before the cache rate. The Cheapest LLM APIs in 2026, by Task lists the cheapest viable option per task type. Free LLM APIs in 2026 covers free tiers and their limits. We re-run this snapshot when a vendor changes a list price.

Sources

• OpenAI: GPT-6 Astra model page
• OpenAI: models overview
• Anthropic: Claude pricing
• Google: Gemini API pricing
• DeepSeek: models and pricing
• Mistral: inference pricing
• xAI: pricing
• Groq: gpt-oss-120B model page
• Voyage AI: pricing
• OpenRouter: OpenAI provider listing