LLM API Pricing in 2026: What OpenAI, Claude, and Gemini Cost per Token

In October 2026, the major LLM APIs charge from $0.05 to $10 per million input tokens and from $0.40 to $50 per million output tokens. At standard context lengths, output costs five to nine times as much as input on every model below. Two levers cut the bill on any of them: prompt caching bills a repeated prompt prefix at 2.5% to 10% of the input price, and batch processing halves the price of work that can wait up to a day.

Below: each provider's prices, what a real workload costs per month, the charges the price list doesn't show, and how Amazon Bedrock and Gemini Enterprise Agent Platform, Google's new name since April 22, 2026 for Vertex AI, price the same models. Every price links to the provider's own page. By OpenAI's rule of thumb of about three quarters of an English word per token, a million tokens is roughly 750,000 words. OpenAI on tokens.

How much does each LLM API cost per million tokens?

OpenAI

OpenAI names GPT-6 Astra its flagship, GPT-6.1 Sol the balanced option, and GPT-6 Luna the one for cost-sensitive work. On models from GPT-5.6 on, writing a prompt to the cache costs 1.25 times the input price for the default 30 minutes, or 2.5 times for 12 hours, and a cache read costs a tenth of the input price or less. OpenAI prompt caching.

Anthropic Claude

Writing to Anthropic's cache costs 1.25 times the input price for five minutes or twice the input price for an hour, and each cache hit restarts the clock at no charge. On Opus 5.5 and Sonnet 5.5 a cache read costs 5% of the input price. Claude 4.6 and later models, except Haiku 5.5, read their full 1M-token context at the standard rate. Claude prompt caching.

Google Gemini

Gemini caches repeated prompt prefixes on its own, with no write fee, once the prefix reaches 4,096 tokens on Gemini 4 Argon and 3.8 Flash, and bills cached tokens at 5% of the input price on Gemini 4 Argon and 10% on 3.8 Flash. Gemini context caching.

What does a real workload cost per month?

Take a support assistant that answers 10,000 questions a day. Each request sends a 4,500-token system prompt with instructions and product facts, plus 500 tokens from the customer, and the model writes a 400-token answer. Over 30 days that's 1.5 billion input tokens, 1.35 billion of them the same prefix, and 120 million output tokens. The cache stays warm all month, so its few writes are left out.

Caching the prefix cuts the bill by 58% to 61%, because the system prompt is 90% of every request. A live assistant can't use batch, since customers won't wait hours for an answer, but nightly jobs such as tagging tickets or summarizing calls can, at half price. The choice of model moves the bill twenty- to fortyfold in this table, so route the easy questions to the small model and send the hard ones up.

Why is my LLM bill higher than the price list?

• Reasoning tokens bill as output. Gemini's output price includes thinking tokens, Anthropic bills thinking tokens as output, and OpenAI's max_output_tokens caps visible and hidden tokens together.
• A chat resends the whole conversation on every turn. Turn ten pays again for the nine before it, unless the earlier turns hit the cache.
• Long prompts cross a price line: 272,000 input tokens on OpenAI's GPT-6 models, 200,000 on Gemini 3.1 Pro, and 100,000 on Claude Haiku 5.5, where the rate goes up fivefold.
• Token counts differ between models. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier Claude models. Claude glossary.
• Regional processing costs more: OpenAI adds 10% for data residency on models released since March 5, 2026, and Anthropic's US-only inference costs 1.1 times the standard rate.
• Cache writes cost more than plain input: 1.25 to 2.5 times on OpenAI, 1.25 to 2 times on Anthropic. A cache that expires between requests pays the write again and again.

Do Amazon Bedrock and Google's Agent Platform (Vertex AI) charge the same as the direct APIs?

For global endpoints, yes. Claude on Bedrock's global cross-region inference costs the same as Anthropic's own API, and Gemini on Agent Platform's global endpoint costs the same as the Gemini API. Regional endpoints cost 10% more on both: Claude Opus 5.5 is $4.40 / $22 in a single region on Bedrock, and Gemini 4 Argon is $2.20 / $11 outside Agent Platform's global endpoint. Amazon Bedrock pricing · Agent Platform pricing.

• For Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1, the 50% batch discount is on Anthropic's own API; Bedrock lists batch for Claude Opus 5 at $2.50 / $12.50.
• Bedrock sells OpenAI's GPT-6 Astra at OpenAI's rates on its global endpoint and at a stated 10% premium in a single region. GPT-6 Astra on Bedrock.
• Agent Platform shows Gemini 4 Argon's $2 input price as a promotion paid through 50% credits back on net spend, against a $4 list price.
• Bedrock and Agent Platform bill through your AWS or Google Cloud account, so the spend lands in the same bill, budgets, and commitments as the rest of your cloud.

How do you spend fewer tokens?

• Put the fixed part of the prompt first and keep it identical between requests, so it hits the cache. OpenAI caches prefixes from 1,024 tokens on GPT-5.6 and later, Claude Fable, Opus 5.5, Sonnet 5.5, and Haiku 5.5 from 512, and Gemini 4 Argon and 3.8 Flash from 4,096.
• Send anything that can wait to the batch API at half price: OpenAI and Gemini return within 24 hours, and Anthropic says most batches finish within an hour. On Anthropic, batch and caching discounts stack. Claude batch processing.
• Route by difficulty. All three providers tell you to use a smaller model for simple tasks, and the small models in the tables cost a twentieth to a hundredth of the large ones.
• Cap the output with each API's maximum-output setting and ask for short answers. Output is the expensive side of every price list.
• Count before you send. OpenAI's POST /v1/responses/input_tokens, Anthropic's free /v1/messages/count_tokens, and Gemini's countTokens tell you what a prompt costs before you pay for it.
• Trim the context: drop old turns, summarize long threads, and send only the documents the question needs.

Common mistakes with LLM API costs

• Comparing models by input price alone, when output and reasoning tokens make up most of many bills.
• Putting the user's question before the system prompt, which breaks the cached prefix on every request.
• Running everything on the flagship model because it was the one used in the prototype.
• Using a regional endpoint by default when no data rule requires it, and paying 10% more for it.
• Tracking the bill per provider only. Without spend per project, key, or team, nobody knows which feature drives it. How to track AI costs by team sets that split up at each provider.

LevelFour reads your OpenAI, Anthropic, and cloud AI bills next to the rest of your cloud and SaaS spend, by model, key, and team. See what LevelFour can save you in 5 minutes, with read-only access.

Frequently asked questions