Expert Guide

The Best LLM APIs in 2026, Priced per Million Tokens

Seven LLM APIs compared on what one shared month of traffic costs, with batch, cache and peak-hour discounts, checked in October 2026.

Louis CorneloupFounder, Dupple · 930,000+ readers · Updated Oct 2026

Product links may be affiliate links. How we rate 7 tools compared

TL;DR

Short answer: for 10 million input and 2 million output tokens a month, Gemini 2.5 Flash-Lite is the cheapest model priced on this page at $1.80, and GPT-6 Luna and Claude Haiku 5.5 come in at the same total as the cheapest models in the GPT-6 and Claude 5.5 families. Prices checked on vendor pricing pages in October 2026.

The same traffic costs $40.00 on GPT-6.1 Sol or Claude Sonnet 5.5. Keep small models on classification and routing, send offline work to batch at half price, and reserve the flagship for the steps that need it. Coding-tool subscriptions are a different bill, covered in our Claude Code pricing guide.

Key facts7 tools compared: Google Gemini API, OpenAI API, Anthropic API, DeepSeek API, Mistral AI API…

Updated: October 11, 2026
Top pick: Google Gemini API (best for: High-volume classification, extraction and tagging, where the lowest bill matters most)
Top pick price as of October 11, 2026: Google Gemini API: Gemini 2.5 Flash-Lite $0.10 in, $0.40 out per 1M; Gemini 3.8 Flash $0.75 in, $3.75 out until Dec 31, 2026

7 tools compared: Google Gemini API, OpenAI API, Anthropic API, DeepSeek API, Mistral AI API, Together AI, xAI API
OpenAI API (best for: Teams already on the OpenAI API that want a current small model with batch and caching built in): GPT-6 Luna $0.10 in, $0.50 out per 1M; GPT-6.1 Sol $2.00 in, $10.00 out (short prompts)
Anthropic API (best for: Teams that route with a small model and call a larger one only for the hard requests): Claude Haiku 5.5 $0.10 in, $0.50 out per 1M (prompts to 100k tokens); Sonnet 5.5 $2.00 in, $10.00 out
DeepSeek API (best for: Cost-sensitive apps that can move batch jobs and bulk work into off-peak UTC hours): DeepSeek V4.1 Flash $0.15 in, $0.60 out per 1M off-peak; double at peak (UTC weekday windows)

Most model comparisons list prices and stop there. A developer needs the monthly bill. Every vendor here bills input and output tokens at separate rates, and output costs three to six times input, so one shared usage level makes the rates comparable.

We price 10 million input and 2 million output tokens a month at standard rates, unless a row says otherwise.

Toolradar data: we evaluated 479 API tools in our API category in October 2026, and 131 of the 200 we rank (66%) list a free plan.

That category covers API development, testing and documentation tools, not model APIs, so this ranking uses the per-token rate instead. For the wider API picks, see best API testing tools and best API gateways.

Top Picks

Based on features, real-world fit, and value for money.

Best LLM APIs in 2026: Price per Million Tokens Compared: 7 tools compared, updated Oct 2026
Tool	Pricing	Best for
Google Gemini API	Gemini 2.5 Flash-Lite $0.10 in, $0.40 out per 1M; Gemini 3.8 Flash $0.75 in, $3.75 out until Dec 31, 2026	High-volume classification, extraction and tagging, where the lowest bill matters most
OpenAI API	GPT-6 Luna $0.10 in, $0.50 out per 1M; GPT-6.1 Sol $2.00 in, $10.00 out (short prompts)	Teams already on the OpenAI API that want a current small model with batch and caching built in
Anthropic API	Claude Haiku 5.5 $0.10 in, $0.50 out per 1M (prompts to 100k tokens); Sonnet 5.5 $2.00 in, $10.00 out	Teams that route with a small model and call a larger one only for the hard requests
DeepSeek API	DeepSeek V4.1 Flash $0.15 in, $0.60 out per 1M off-peak; double at peak (UTC weekday windows)	Cost-sensitive apps that can move batch jobs and bulk work into off-peak UTC hours
Mistral AI API	Mistral Small 2603 $0.15 in, $0.60 out per 1M; Mistral Medium 3.5 $1.50 in, $7.50 out per 1M	Teams that want low-cost small models, a 50% batch discount and cheap repeated prompts
Together AI	gpt-oss-120B $0.15 in, $0.60 out per 1M on serverless inference, open-weight models from several labs	Teams that want open-weight models such as gpt-oss-120B on one bill and one API
xAI API	Grok 4.7 $2.00 in, $6.00 out per 1M tokens; 500k-token context window	Agent builders who want a 500k-token window and the Web Search and X Search tools in one API

Pricing read from each vendor's own published pricing page, checked Oct 2026. Every vendor here publishes a price.

Google Gemini API

Top Pick

Best for: High-volume classification, extraction and tagging, where the lowest bill matters most

PricingGemini 2.5 Flash-Lite $0.10 in, $0.40 out per 1M; Gemini 3.8 Flash $0.75 in, $3.75 out until Dec 31, 2026

+Gemini 2.5 Flash-Lite costs $1.80 at this usage level, the lowest total on this page

+Batch is half the standard rate across the paid tiers

+Gemini 3.1 Pro Preview's output rate on prompts up to 200k tokens is below Claude Opus 5.5's

−Gemini 2.5 Flash-Lite is a previous-generation model, so the cheapest line is not the newest one

−Gemini 3.8 Flash input and output double on January 1, 2027, which breaks any budget set in 2026

Visit Google Gemini API →

OpenAI API

Best for: Teams already on the OpenAI API that want a current small model with batch and caching built in

PricingGPT-6 Luna $0.10 in, $0.50 out per 1M; GPT-6.1 Sol $2.00 in, $10.00 out (short prompts)

+GPT-6 Luna is the cheapest GPT-6 model and ties Claude Haiku 5.5 on total

+Batch halves both models, so the GPT-6.1 Sol scenario drops by half

+Cached input has its own price line, so the cache discount shows up separately on the bill

−Long-context prompts cost more, and GPT-6.1 Sol's long-context output rate is one and a half times its short-context rate

−GPT-6-astra is the most expensive OpenAI model on the page, at five times Sol's output rate

Visit OpenAI API →

Anthropic API

Best for: Teams that route with a small model and call a larger one only for the hard requests

PricingClaude Haiku 5.5 $0.10 in, $0.50 out per 1M (prompts to 100k tokens); Sonnet 5.5 $2.00 in, $10.00 out

+Batch halves input and output for Sonnet 5.5 and the other listed models

+A cache hit costs 5% of the standard input price, which is $0.10 per 1M on Sonnet 5.5

+The price page lists cache writes separately, so you can cost a long-lived cache before you build it

−Haiku 5.5 changes price above 100,000 prompt tokens, to $0.50 in and $2.50 out

−Claude Fable 5.1 costs $10.00 in and $50.00 out, one of the most expensive Claude rates on the page

Visit Anthropic API →

DeepSeek API

Best for: Cost-sensitive apps that can move batch jobs and bulk work into off-peak UTC hours

PricingDeepSeek V4.1 Flash $0.15 in, $0.60 out per 1M off-peak; double at peak (UTC weekday windows)

+Off-peak V4.1 Flash ties the Mistral Small and gpt-oss-120B entries for the scenario total

+A cache hit on the flash model is $0.003 per 1M off-peak, the lowest cached rate here

+One pricing table lists off-peak and peak columns, so the budget is easy to model by hour

−Peak pricing doubles every rate, so an unscheduled job that runs in a weekday peak costs twice as much

−V4 Pro costs about four times the Flash bill for the same month, even off-peak

Visit DeepSeek API →

Mistral AI API

Best for: Teams that want low-cost small models, a 50% batch discount and cheap repeated prompts

PricingMistral Small 2603 $0.15 in, $0.60 out per 1M; Mistral Medium 3.5 $1.50 in, $7.50 out per 1M

+Mistral Small 2603 ties the DeepSeek and gpt-oss-120B entries for the scenario total

+The pricing FAQ states the batch and cache discounts in plain terms, which makes budgeting simple

+Mistral Medium 3.5 is priced below GPT-6.1 Sol and Claude Sonnet 5.5 at the scenario total

−Mistral Medium 3.5's output rate is about twelve times Small's, so the bill moves fast when you step up

−The pricing FAQ's worked example uses rates that differ from Mistral's model cards, so budget from the model card rates

Visit Mistral AI API →

Together AI

Best for: Teams that want open-weight models such as gpt-oss-120B on one bill and one API

Pricinggpt-oss-120B $0.15 in, $0.60 out per 1M on serverless inference, open-weight models from several labs

+gpt-oss-120B ties the cheapest DeepSeek and Mistral Small entries for the scenario total

+The serverless table lists dozens of open-weight models with input and output prices in one place

+A Batch API price column sits beside the serverless prices on the same page

−Prices change by model and by host, so the same open-weight model can cost more on another provider

−Some listed promotions have end dates, so check the date on any promo line before you budget against it

Visit Together AI →

xAI API

Best for: Agent builders who want a 500k-token window and the Web Search and X Search tools in one API

PricingGrok 4.7 $2.00 in, $6.00 out per 1M tokens; 500k-token context window

+Its output rate is below the output rate of GPT-6.1 Sol and Claude Sonnet 5.5

+Web Search and X Search are listed as tools in the same API docs, so real-time retrieval sits on one API

−Input is priced the same as GPT-6.1 Sol and Claude Sonnet 5.5, so the saving is on output only

−Batch is not available on every model, so check the model page before you plan bulk work

Visit xAI API →

What it is

An LLM API bills by the token: a token can be a word, a number or a punctuation mark. Each request is charged for the input tokens you send and the output tokens the model returns, at separate rates.

Why it matters

Output is the expensive side. On these price lists, output costs three to six times input. Claude Sonnet 5.5 and GPT-6.1 Sol both price output at five times input, and Google's Gemini 3.1 Pro Preview at six times. A support bot that writes long answers pays more for its replies than for its questions, so measure both sides before you pick a model.

The same two-rate shape applies to voice pricing, covered in best text-to-speech APIs.

Price the scenario, not the sticker. A cheap input rate and a cheap output rate look harmless on their own. At 2 million output tokens, the output line is half of the GPT-6 Luna bill.

Key features to look for

Input and output ratesEssential

Every model here bills the two separately, and output costs three to six times input. A model that writes long answers can cost more than one that reads long prompts, so measure both.

Batch processing

OpenAI, Anthropic, Google and Mistral all list batch at half the standard rate. Use it for overnight jobs, evaluations and backfills, where a delay of hours costs nothing.

Prompt caching

Repeated prompt prefixes bill at a fraction of the input rate. Anthropic charges 5% of the input price for a cache hit, OpenAI charges 5% on GPT-6.1 Sol, and Mistral says cached input is up to 90% cheaper.

Time-of-day pricing

DeepSeek charges half its peak rate off-peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, so a scheduled job outside those windows costs half as much per token.

Context length

Context windows differ: Mistral Small and Medium list 256k tokens, Grok 4.7 lists 500k and DeepSeek lists 1M. Long prompts often move a request to a higher price tier, so check the tier before you rely on the window.

Pricing

Methodology: we read each vendor's own pricing page in October 2026 and took the entry-level current text model on each list, plus a flagship for context. For Together AI we priced gpt-oss-120B as the reference open-weight model, because its serverless list also carries cheaper small models.

Items are ordered by the entry model's cost at 10 million input and 2 million output tokens a month.

The scenario cost is the input volume times the input rate, plus the output volume times the output rate, with no caching or batch unless the row says so.

Google's Gemini API has the cheapest entry rate, from Gemini 2.5 Flash-Lite on Google's price list.

Its newer Gemini 3.8 Flash is priced at one rate through December 31, 2026 and rises on January 1, 2027, so the same month of traffic costs $15.00 this year and $30.00 next year.

OpenAI and Anthropic price their current small models at the same scenario total, which is why they tie.

Claude Haiku 5.5 moves to a higher rate above 100,000 prompt tokens.

Gateways add their own markup on top of these rates; our OpenRouter alternatives page covers those fees.

DeepSeek's price list sets V4.1 Flash at a lower off-peak rate and doubles it in the UTC peak windows.

The Mistral Small model card gives its rates, and the Mistral pricing FAQ confirms the 50% batch discount.

For coding-tool bills rather than API bills, see our Cursor pricing guide.

Plan	Price	Best for
Gemini 2.5 Flash-Lite	$1.80 a month	Previous-generation model and the lowest total on the list
GPT-6 Luna	$2.00 a month	Current small model; batch halves the rate
Claude Haiku 5.5	$2.00 a month	Current small model; prompts over 100k tokens cost more
DeepSeek V4.1 Flash	$2.70 off-peak, $5.40 at peak	Peak hours double the rate
gpt-oss-120B on Together AI	$2.70 a month	Open-weight model on serverless hosting
Mistral Small 2603	$2.70 a month	256k context window; batch at half the rate
Gemini 3.8 Flash	$15.00 to 2026, $30.00 from 2027	Input and output rates double on January 1, 2027
GPT-6.1 Sol	$40.00 a month, $20.00 on batch	Flagship-class OpenAI model
Claude Sonnet 5.5	$40.00 a month, $20.00 on batch	Mid-tier Claude model; cache hits cost 5% of input
Grok 4.7	$32.00 a month	500k-token context window
Mistral Medium 3.5	$30.00 a month	256k context window
DeepSeek V4 Pro	$10.56 off-peak, $21.12 at peak	The Pro tier of DeepSeek V4
Claude Opus 5.5	$80.00 a month	For long-running agentic coding and knowledge work
Claude Fable 5.1	$200.00 a month	For demanding reasoning and long-horizon agentic work

Mistakes to avoid

×Comparing per-token rates with monthly plan prices. Convert every line to the same unit before you put two numbers side by side.

×Budgeting DeepSeek at the off-peak rate. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, and it doubles the bill for scheduled jobs that run in those windows.

×Missing the January 1, 2027 step on Gemini 3.8 Flash, where input and output both double. A budget signed off in 2026 misses it.

×Sending long documents without checking the tier. Claude Haiku 5.5 and Gemini 3.1 Pro both change price above a prompt-size threshold, and GPT-6.1 Sol lists a separate long-context rate.

×Paying for cache writes without repeat traffic. Anthropic bills a 1-hour cache write at twice the input rate, so caching pays only when the same prefix is read again and again.

Expert tips

→Price the scenario with your own token counts. Log input and output tokens for a week, then run them through each rate card, because the output share often decides the ranking.

→Send offline work to batch. OpenAI, Anthropic, Google and Mistral all list batch at half the standard rate, so evaluations, backfills and nightly enrichment cost half as much.

→Route by task. Use a small model for classification and extraction and escalate only the hard cases. At this usage level, a flagship on every call costs twenty times more than a small model.

→Re-check the price list before you commit. Gemini's January 2027 step-up and DeepSeek's peak windows show that the same model can cost a different amount on another date.

The bottom line

Default to Gemini 2.5 Flash-Lite or GPT-6 Luna for high-volume jobs with short outputs, and send anything that can wait to batch. Step up to GPT-6.1 Sol or Claude Sonnet 5.5 only for the calls that need it, because the same traffic costs twenty times more.

If the calls run on a serverless platform, the hosting bill is separate; see best serverless platforms.

Watch the calendar. Gemini 3.8 Flash doubles on January 1, 2027, and DeepSeek charges double in its UTC peak windows.

For another per-unit cost comparison in the same format, read object storage pricing compared. Cite this: Devshot, "Best LLM APIs in 2026: Price per Million Tokens Compared", October 2026.

Frequently asked questions

What is the cheapest LLM API in 2026?

For 10 million input and 2 million output tokens a month, Gemini 2.5 Flash-Lite is the cheapest model priced on this page at $1.80, checked on Google's pricing page in October 2026. GPT-6 Luna and Claude Haiku 5.5 come out at the same total and are the cheapest models in the GPT-6 and Claude 5.5 families.

How much does GPT-6.1 Sol cost per million tokens?

GPT-6.1 Sol lists $10.00 per 1M output tokens for short prompts, and half that on the batch tier, per OpenAI's pricing page in October 2026. Long prompts pay a higher input and output rate, so check the tier before you budget a long document.

Is the batch API really half price?

Yes, for the vendors that publish it. OpenAI, Anthropic and Google price batch at 50% of the standard rate on input and output, and Mistral states a 50% reduction in its pricing FAQ. DeepSeek's discount comes from off-peak hours instead.

Does prompt caching cut the bill?

On repeated prompt prefixes, yes. Anthropic charges 5% of the input rate for a cache hit, OpenAI charges 5% on GPT-6.1 Sol, and Mistral says cached input is up to 90% cheaper. Cache writes cost extra on Anthropic, so measure your repeat rate before you enable caching.

Is DeepSeek cheaper than OpenAI?

Not at this usage level. DeepSeek V4.1 Flash costs $2.70 off-peak and $5.40 at peak, which is more than GPT-6 Luna and Claude Haiku 5.5 at the same usage level. The answer depends on which models you compare and when your jobs run, so cost the scenario with your own token counts.

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

Get the Devshot brief

Free daily newsletter, read in 5 minutes.

Subscribe free