LLM API Prices 2026: Cost Per Million Tokens Across 20 Models

The cheapest LLM APIs in 2026 are OpenAI's gpt-6-luna and Anthropic's Claude Haiku 5.5, both $0.10 per million input tokens and $0.50 per million output tokens. Mistral Small 4 ($0.15 and $0.60) and DeepSeek's deepseek-flash follow. The most expensive are gpt-6-astra and Claude Fable 5.1 at $10 and $50, a hundred times the cheapest rate.

We read every price on the six vendors' own API pricing pages on 19 October 2026, so the sheet is as of October 2026, in US dollars at the standard (non-batch) tier and short-context rates. Our view: picking the model is the decision that moves an API bill by 100x. Batch discounts, caching and off-peak hours move it by 2x, so get the routing right first.

AIpresso data: across the 20 models on this sheet, the median price is $1.29 per million input tokens and $3.86 per million output tokens, and half of them (10 of 20) charge exactly five times more for output than for input. Only 3 models cost under $1 per million output tokens.

The price sheet: 20 models, October 2026

Price per million tokens, standard tier, short prompts. Long-context and priority rates are higher; batch rates are lower.

Sources, all read on 19 October 2026: OpenAI API pricing, Claude API pricing, Gemini API pricing, xAI API pricing, DeepSeek API pricing, and Mistral's model cards for Medium 3.5, Large 3 and Small 4.

Five findings from the sheet

1. The flagship tier has converged on one price. OpenAI's gpt-6-astra and Anthropic's Fable 5.1 charge identical rates, while Google's top model, Gemini 4 Argon, sits at a fifth of that until its promotion ends.
2. Mid-tier pricing clusters at one rate too. gpt-6.1-sol, Claude Sonnet 5.5 and Gemini 4 Argon all charge the same today, which makes quality per task, not price, the deciding factor in that band.
3. Output is where the money goes. Ten of the 20 models charge five times more per output token than per input token, so a chatty model that writes long answers costs far more than its input price suggests.
4. Two Google rates are temporary. Gemini 4 Argon rises to $4 in and $20 out on 1 December 2026, and Gemini 3.8 Flash doubles to $1.50 and $7.50 on 1 January 2027, so a budget built on today's Google prices is half the real one.
5. xAI and Mistral price output low. Grok 4.3, grok-build-0.1 and Mistral Large 3 charge two to three times their input price for output, against five times at most rivals, which favors them for long generated answers.

Which LLM API is cheapest per million tokens?

At list price, gpt-6-luna and Claude Haiku 5.5 tie as the cheapest, with Mistral Small 4 next. Among models that sit closer to the mid-tier, Mistral Large 3 and Gemini 3.1 Flash-Lite are the low-cost options from vendors whose flagships compete at the top.

DeepSeek changes the ranking by the clock. Its off-peak rates are half the peak ones, so deepseek-flash costs $0.15 in and $0.60 out outside peak hours, matching Mistral Small 4. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which overlaps the European working morning, so a US team running jobs during its own working day pays the off-peak rate.

Google publishes the clearest free tier for API use: its Flash and Flash-Lite models are free of charge on it, with rate limits, while Gemini 4 Argon and Gemini 3.1 Pro Preview are paid only. For prototypes that never leave a developer's laptop, that free tier is the cheapest API of all.

What one workload costs on each model

Per-token prices are hard to compare in the abstract, so we priced one workload: a support assistant that reads 10 million input tokens and writes 2 million output tokens a month, at standard rates.

The same job costs $2 or $200 depending on the model, while models in the same tier from different vendors land within a few dollars of each other. After Google's promotions end, the Gemini 4 Argon line doubles to $80 and Gemini 3.8 Flash to $30.

At these volumes the absolute amounts are small for any business. The ratio is what matters once traffic grows: a workload 100 times larger turns the $40 mid-tier bill into $4,000 a month and the flagship bill into $20,000.

AIpresso is the short daily brief on which AI prices actually moved. Get it here if you would rather not recheck six API price lists every time a rate changes.

The discounts that halve a bill

Every major vendor sells a cheaper way to run the same model, and most cut the price by half.

• Batch processing. OpenAI, Anthropic, Google and Mistral all price batch requests at half the standard rate, in exchange for results that arrive within hours instead of seconds. Any job that is not user-facing (summaries, classification, evals) belongs there.
• Off-peak hours. DeepSeek halves both input and output prices outside its weekday peak windows, with no change to the model.
• Caching. Repeated prompt prefixes cost far less when cached: gpt-6.1-sol charges $0.10 per million cached input tokens, a twentieth of its standard input rate, and Mistral says cached input cuts input cost by up to 90%.
• OpenAI Flex. OpenAI's Flex tier matches its batch price, half the standard rate, for workloads that tolerate slower responses.

The surcharges run the other way. OpenAI's Fast mode doubles the standard rate, Anthropic adds a 1.1x multiplier for US-only inference on recent models, and long prompts move several models to a higher band: Claude Haiku 5.5 charges $0.50 in and $2.50 out above 100,000 tokens, and Gemini 3.1 Pro Preview $4 and $18 above 200,000.

How to pick an API on price

Start with the cheapest model that passes your own quality test on real prompts, since the model choice sets the bill's order of magnitude. Luna, Haiku 5.5 and Mistral Small 4 cover classification, extraction and routing; the mid-tier band covers most assistant and coding work; the flagships earn their price only on the hardest reasoning tasks.

Then apply the 2x levers: batch everything that can wait, cache long system prompts, and on DeepSeek schedule bulk jobs off-peak. Budget Google at its post-promotion rates. For subscription prices rather than API rates, see our AI subscription prices 2026 sheet, and for each vendor's consumer and API plans, our Claude pricing, Gemini pricing, Grok pricing and ChatGPT pricing pages. DeepSeek vs ChatGPT, Claude vs ChatGPT and Gemini vs ChatGPT compare the assistants on more than price, and our AI model release tracker logs each new model as it ships. Developers choosing an API on features as well as cost can read DevShot's best LLM APIs ranking.

How we built this sheet

Louis Corneloup, founder of Dupple and Toolradar, compiled this price sheet. We read each model's standard input and output price per million tokens on the vendor's own pricing page or model card on 19 October 2026, picked each vendor's current flagship, mid-tier and low-cost models for a total of 20, and used short-context, non-batch rates throughout. Medians, ratios and the workload costs are our arithmetic from those list prices. Prices exclude tax, tool calls such as web search, and negotiated discounts. We plan to re-read the sheet each quarter and after each major model launch.

FAQ

Which LLM API is cheapest per million tokens in 2026?

OpenAI's gpt-6-luna and Anthropic's Claude Haiku 5.5 tie at $0.10 per million input tokens and $0.50 per million output tokens, read on vendor pages on 19 October 2026. Mistral Small 4 is next at $0.15 and $0.60, and DeepSeek's deepseek-flash matches it during off-peak hours.

What is the most expensive LLM API?

OpenAI's gpt-6-astra and Anthropic's Claude Fable 5.1, both $10 per million input tokens and $50 per million output tokens at standard rates. That is 100 times the cheapest models on the sheet, and OpenAI's priority tiers cost more again.

How much does the GPT-6 API cost?

gpt-6-astra costs $10 in and $50 out per million tokens, gpt-6.1-sol a fifth of that, and gpt-6-luna $0.10 and $0.50, at OpenAI's standard short-context rates. Batch and Flex processing halve those prices, and Fast mode doubles them.

Is the Claude API cheaper than OpenAI's?

At matching tiers they are priced the same: Claude Sonnet 5.5 and gpt-6.1-sol charge the same per million tokens, and Haiku 5.5 matches gpt-6-luna. Claude Opus 5.5, at $4 and $20, has no OpenAI equivalent at that price.

Are Gemini API prices going up?

Yes. Google lists Gemini 4 Argon at $2 in and $10 out until 30 November 2026, rising to $4 and $20 on 1 December. Gemini 3.8 Flash rises from $0.75 and $3.75 to $1.50 and $7.50 on 1 January 2027.

Is there a free LLM API?

Google's Gemini API has a free tier for its Flash and Flash-Lite models, with rate limits and no charge per token, while its top models are paid only. Mistral's Free plan also covers getting started in its Studio console, with lower limits than its paid plan.

Cite this: AIpresso, "LLM API Prices 2026: Cost Per Million Tokens Across 20 Models", November 2026.

If you want price moves like these without rechecking six vendor pages, get the AIpresso brief.

Some offers on this page may be paid placements or contain affiliate links.