LLM API Pricing Comparison 2026: Cost per Million Tokens for Every Major Model

LLM API prices in October 2026 range from $0.10 per million input tokens (Claude Haiku 5.5 and GPT-6 Luna) to $10 per million (GPT-6 Astra and Claude Fable 5.1), a 100-fold spread. Output tokens cost more, from $0.50 to $50 per million. For most apps, a mid-tier model such as Claude Sonnet 5.5 or GPT-6.1 Sol at $2 input and $10 output gives near-flagship quality, while budget models handle high-volume tasks for a few dollars a month.

This LLM API pricing comparison lists the standard price per million tokens for every major model from OpenAI, Anthropic, Google, xAI, DeepSeek and Meta, shows what a typical workload costs on each, and explains the discounts and hidden charges that change your real bill. Prices are list rates in US dollars from official pricing pages unless marked otherwise.

LLM API pricing comparison table (October 2026)

The last column shows a typical monthly workload of 10 million input tokens and 2 million output tokens, calculated as input price x 10 plus output price x 2. For example, Claude Sonnet 5.5 costs 10 x $2 + 2 x $10 = $40. If tokens are new to you, start with our explainer on what tokens are in AI.

OpenAI API pricing

OpenAI’s current family has three tiers. GPT-6 Astra is the flagship at $10 input, $1 cached input and $50 output per million tokens. GPT-6.1 Sol, launched at DevDay 2026, costs $2, $0.10 cached and $10 and nearly matches Astra on coding, according to OpenAI. GPT-6 Luna is the budget tier at $0.10, $0.01 cached and $0.50. Requests with more than 272K input tokens are billed at 2x input and 1.5x output for the whole request. The older GPT-5.6 Sol is on a promotional $4 / $20 rate at least through November 21, 2026. Full details are in our OpenAI API pricing guide and on the official pricing page.

Anthropic Claude API pricing

Anthropic’s 5.5 family covers every tier. Claude Opus 5.5 costs $4 / $20, Claude Sonnet 5.5 costs $2 / $10, and Claude Haiku 5.5 costs $0.10 / $0.50 for prompts up to 100K tokens ($0.50 / $2.50 above that). Claude Fable 5.1 is the premium tier at $10 / $50. Cache reads are much cheaper: $0.20 per million on Opus 5.5, $0.10 on Sonnet 5.5 after the October 7 cut, and $0.01 on Haiku 5.5. The Batch API takes 50% off. See our Claude API pricing breakdown, plus our reviews of Sonnet 5.5 and Haiku 5.5, and the official Claude pricing page.

Google Gemini API pricing

Gemini 3.8 Flash costs $0.75 input and $3.75 output until December 31, 2026, then rises to $1.50 / $7.50 from January 1, 2027. Gemini 3.5 Flash costs $1.50 / $9.00 and Gemini 3.5 Flash-Lite $0.30 / $2.50. Batch pricing halves these. Gemini 4 Argon was reported at an introductory $2 / $10 with a standard rate of $4 / $20; check Google’s pricing page for current terms, since Argon access was still expanding in October 2026. Our Gemini API pricing guide has the full list.

xAI, DeepSeek and Meta API pricing

• xAI Grok 4.7: $2 input, $0.50 cached, $6 output per million tokens, with a 500K context window. Low output pricing makes it attractive for long generations.
• DeepSeek V4.1 Flash: $0.30 input and $1.20 output at peak times, and half that off-peak ($0.15 / $0.60), with cache hits from $0.003. It has a 1M context window. Our DeepSeek API pricing guide covers peak hours and the V4 Pro changes.
• Meta Muse Wonder: $4 input and $20 output on the Meta Model API, priced head to head with Claude Opus 5.5. See our Muse Wonder guide.

Discounts that change your real cost

Prompt caching

If your app resends the same instructions or documents, caching makes repeat input very cheap. On GPT-6.1 Sol, cached input is $0.10 instead of $2, a 95% saving. On Claude Haiku 5.5, a cache read is $0.01 instead of $0.10. Our guide to prompt caching explains how to structure prompts to hit the cache.

Batch APIs

OpenAI, Anthropic and Google all offer about 50% off for jobs processed asynchronously, usually within 24 hours. Sonnet 5.5 drops to $1 / $5 on Batch. See our Batch API guide.

Off-peak pricing

DeepSeek halves prices outside peak hours, which suits scheduled jobs.

Hidden costs to watch

• Long-context surcharges: OpenAI’s GPT-6 family bills the whole request at higher rates above 272K input tokens; Haiku 5.5’s rate jumps five times above 100K.
• Thinking tokens: reasoning models bill their internal thinking as output tokens, so higher effort levels raise cost even if the visible answer is short.
• Tokenizer differences: the same text can produce different token counts across providers. Anthropic says its 5.5 tokenizer produces about 30% more tokens than earlier Haiku models.
• Premium speed tiers: OpenAI’s Ultrafast costs six times the standard Astra rate, and Claude’s Fast mode doubles Opus 5.5 prices.

Our guides on estimating AI API costs and reducing token usage help you control these.

How to estimate your own LLM API cost

1. Measure one real request. Run a typical prompt and note the input and output token counts in the API response or dashboard.
2. Multiply by monthly volume. If one support reply uses 2,000 input and 400 output tokens, 50,000 replies a month is 100M input and 20M output tokens.
3. Apply the price. On GPT-6.1 Sol: 100 x $2 = $200 input, plus 20 x $10 = $200 output, for $400 a month. On Claude Haiku 5.5 the same volume costs 100 x $0.10 + 20 x $0.50 = $20.
4. Subtract caching. If 1,500 of the 2,000 input tokens are a fixed system prompt, 75M tokens can be cached. On GPT-6.1 Sol that portion drops from $150 to $7.50, cutting the bill to about $257.50.
5. Add a safety margin. Budget 20% to 30% extra for retries, longer conversations and reasoning tokens.

Repeat the calculation for two or three candidate models and test quality on your own examples before deciding. A model that costs ten times less but needs twice as many retries may not be cheaper in practice.

Which LLM API should you choose?

The best setups route each request to the cheapest model that can handle it; our model routing guide and list of the cheapest AI model for each task show how.

Frequently asked questions

The bottom line

For most new projects, start with a mid-tier model like Claude Sonnet 5.5 or GPT-6.1 Sol, add caching from day one, and route simple tasks to Haiku 5.5, GPT-6 Luna or DeepSeek. Use flagships only where they clearly earn their price. Prices in this table were checked in October 2026, and they change often, so confirm current rates on each provider’s pricing page before you build your budget.