LLM API pricing in 2026: what a million tokens costs at 12 APIs

We read the pricing pages of twelve LLM APIs, from OpenAI and Anthropic to DeepSeek and the open-model hosts: the price per million tokens in and out, the condition attached, and the lines that inflate the bill.

Short answer: a million output tokens costs from $0.10 (Ministral 3 3B) to $40 (GPT-6.2 Astra and Claude Fable 5.2) across the reference models on the cards below. The mid-range has converged: Claude Sonnet 5.5, GPT-6.2 Sol and Kimi K3.1 Pro all charge $2 input and $10 output, and Gemini 4 Carbon matched them at its launch rate, which rises to $4 / $20 in December. Grok 4.7 costs $2 / $6, Mistral Large 4 $0.70 / $2.10, and DeepSeek V4.1 Flash $0.30 / $1.20 at peak hours. Below, each price with its condition, a month priced on sixteen models, and ten traps.

To pick an API on quality, not price, see our best LLM API ranking; for the cheapest options only, our cheap LLM API guide.

Romain Cochard
Romain CochardCEO
Updated 202612LLM APIs priced$0.10 to $40per 1M output tokens2026pricing pages reread

The index is editorial: a sponsored placement is always labeled on its card, and no affiliate link changes a price or an order.

Prices in 2026

The price of a million tokens at 12 LLM APIs

An LLM API lets your code send text to a language model hosted by someone else and get text back, without running the model yourself. You pay per token: the small unit of text a model reads and writes, a word, part of a word, a number or a punctuation mark. Prices are quoted per million tokens (often written 1M or MTok), with two rates: input for what you send, output for what the model writes, usually three to five times higher, and eight times on Gemini 3.5 Flash-Lite.

Three discounts sit on top of the list price. Cached input is a repeated prompt prefix the provider keeps in memory, billed at a fraction of the input rate. Batch processing returns results within hours at a discount: half price at most providers, 40% off at Moonshot, 20% off on xAI's Grok models. Off-peak pricing, at DeepSeek, halves the rate outside set hours. Each card shows the flagship or reference model of the provider, its entry-level model, and the conditions that change the bill.

What to require from an LLM API before you compare: a published per-token price for input, cached input and output; the prompt length at which the rate changes; the price of a cache write; the batch discount and which models it covers; and where your data is processed.

USD before tax. Standard pay-as-you-go rates, input / output per 1M tokens, short prompts. Read on each provider's pricing page in 2026.

OpenAI

GPT-6.2 Astra

$8.00 / $40.00input / output per 1M tokens

GPT-6 Luna $0.10 / $0.50

  • Cached input $0.40
  • Over 272K input tokens: 2x input
  • Batch at 50%, Fast mode 2x
See OpenAI pricing →

Anthropic

Claude Fable 5.2

$8 / $40input / output per 1M tokens

Claude Haiku 5.5 $0.10 / $0.50

  • Opus 5.5 $4 / $20, Sonnet 5.5 $2 / $10
  • Batch API 50% off
  • US-only inference 1.1x
See Claude pricing →

Google Gemini

Gemini 4 Carbon

$2.00 / $10.00input / output per 1M tokens, launch rate

Then $4.00 / $20.00

  • Not on the free tier
  • Gemini 3.5 Flash-Lite $0.30 / $2.50
  • Grounding $14 per 1,000 after 5,000 free
See Gemini API pricing →

xAI

Grok 4.7

$2.00 / $6.00input / output per 1M tokens

Grok 4.3 $1.25 / $2.50

  • Cached input $0.50
  • From 200k tokens: $4.00 / $12.00
  • 500k-token context
See xAI pricing →

Moonshot AI

Kimi K3.1 Pro

$2.00 / $10.00input / output per 1M tokens

Kimi K3 $3.00 / $15.00

  • Cached input $0.20
  • 5-minute cache write $2.00
  • Highspeed version $4.00 / $20.00
See Kimi API pricing →

DeepSeek

DeepSeek V4 Pro, peak hours

$1.32 / $3.96input / output per 1M tokens

Flash $0.30 / $1.20 at peak

  • Off-peak rates are half
  • Flash cache hit $0.006 at peak
  • 1M-token context
See DeepSeek pricing →

Mistral AI

Mistral Large 4

$0.70 / $2.10input / output per 1M tokens

Ministral 3 3B $0.10 / $0.10

  • Mistral Small 4 $0.15 / $0.60
  • Batch 50% off
  • Prices in USD or EUR
See Mistral pricing →

Together AI

DeepSeek V4.1 Flash

$0.30 / $1.20input / output per 1M tokens

US geo-pinned $0.375 / $1.50

  • GLM-5.3 $1.40 / $4.40
  • Kimi K3.1 Pro $2.00 / $10.00
  • gpt-oss-120B $0.15 / $0.60
See Together AI pricing →

Fireworks AI

DeepSeek V4.1 Flash

$0.30 / $1.20input / output per 1M tokens

Priority $0.375 / $1.50

  • Batch 50% off
  • Spot 30% off on select models
  • No cache write fee
See Fireworks pricing →
Lowest V4.1 Flash rate

DeepInfra

DeepSeek V4.1 Flash 1017

$0.098 / $0.294input / output per 1M tokens, 51% off

List $0.20 / $0.60

  • DeepSeek V4 Pro $1.30 / $2.60
  • Card or prepayment required
  • 1024k-token context
See DeepInfra pricing →

Groq

GPT OSS 120B

$0.15 / $0.60input / output per 1M tokens

GPT OSS 20B $0.075 / $0.30

  • 500 tokens per second (120B)
  • 131,072-token context
  • Batch and Flex on the Developer plan
See Groq pricing →

OpenRouter

Gateway to many providers

List priceplus 5.5% on credit purchases ($0.80 minimum)

BYOK: $25,000 a month fee-free

  • 25+ free models
  • Free models: 50 requests a day
  • Failed attempts not billed
See OpenRouter pricing →

Twelve APIs. Model makers (OpenAI, Anthropic, Google, xAI, Moonshot, DeepSeek, Mistral) show their flagship or reference model; hosts (Together AI, Fireworks, DeepInfra, Groq) show a shared open model; OpenRouter bills each provider's list price plus a fee on credits. Gemini 4 Carbon's price is a launch rate; DeepSeek's is the peak-hour rate; DeepInfra's includes a 51% discount shown on its page.

API by API

The prices that really matter, API by API

For each API: what the reference price buys, the entry-level model, and the condition that changes the bill.

01

OpenAI: four GPT-6 price points, and cache writes on top

OpenAI's current line runs from GPT-6 Luna at $0.10 input and $0.50 output per million tokens, through GPT-6.1 Terra ($1.00 / $5.00) and GPT-6.2 Sol ($2.00 / $10.00), up to GPT-6.2 Astra at $8.00 / $40.00. Cached input costs a fraction of the input rate: $0.40 on Astra, $0.01 on Luna.

Two lines matter for the bill. Cache writes are priced: $10.00 per million on Astra for 30 minutes of retention, $20.00 for 12 hours. And a prompt over 272K input tokens doubles the input and cache rates and adds 50% to output, for the whole request. Batch costs half the Standard rate (Flex too, but only on the older models that still offer it, not on GPT-6.2); Fast mode costs twice. Regional processing for data residency adds 10%. If you are leaving OpenAI, our guide to switching from OpenAI lists the APIs teams move to.

GPT-6.2 Astra
$8.00 / $40.00
GPT-6 Luna
$0.10 / $0.50
See OpenAI pricing →
02

Anthropic: Claude from $0.10 to $8 per million input tokens

Claude comes in four current models: Fable 5.2 at $8 input and $40 output per million tokens, Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10, and Haiku 5.5 at $0.10 / $0.50 for prompts up to 100,000 tokens. Above that length Haiku 5.5 moves to $0.50 / $2.50, while Claude 4.6 and later models (Haiku 5.5 aside) keep their standard rate across the full 1M-token context.

Prompt caching has two durations: a 5-minute write costs 1.25x the input price and a 1-hour write 2x, and a cache hit costs $0.20 on Fable 5.2 and Opus 5.5, $0.10 on Sonnet 5.5. The Batch API takes 50% off input and output. Pinning inference to the US costs 1.1x on every token category. Our Anthropic review and Claude review cover the models beyond price; Anthropic alternatives lists the other APIs.

Claude Sonnet 5.5
$2 / $10
Claude Haiku 5.5
$0.10 / $0.50
See Claude pricing →
03

Google Gemini: a free tier, and a launch rate that ends

Gemini 4 Carbon, Google's most capable reasoning model, costs $2.00 input and $10.00 output per million tokens at its launch rate, and Google's own pricing table already shows the next one: $4.00 / $20.00 from the start of December. Gemini 4 Argon has the same two rates; Gemini 4 Neon costs $3.00 / $15.00. None of the three is on the free tier.

The free tier covers models such as Gemini 4 Hydrogen ($1.50 / $7.50 on the paid tier), Gemini 3.8 Flash and Gemini 3.5 Flash-Lite ($0.30 / $2.50 paid). On the free tier, your content is used to improve Google's products; on the paid tier it is not. Output prices include thinking tokens. Grounding with Google Search gives 5,000 free requests a month across Gemini 3 and newer models, then $14 per 1,000. See our Gemini review and Gemini alternatives.

Gemini 4 Carbon now
$2.00 / $10.00
Next rate
$4.00 / $20.00
See Gemini API pricing →
04

xAI: Grok 4.7 at $2 / $6, double past 200k tokens

Grok 4.7, xAI's flagship for code and general use, costs $2.00 input and $6.00 output per million tokens, with cached input at $0.50 and a 500k-token context. Its output rate is low for a frontier model: $6.00 is 40% below the $10.00 of Claude Sonnet 5.5 and GPT-6.2 Sol.

The condition is the prompt length. Once a request reaches 200k tokens, every token in it bills at the long-context rate: $4.00 input and $12.00 output. Grok 4.3 costs $1.25 / $2.50 with a 1M-token context, and web search costs $5 per 1,000 calls. Our Grok review and Grok alternatives go beyond the API rate.

Grok 4.7
$2.00 / $6.00
From 200k tokens
$4.00 / $12.00
See xAI pricing →
05

Moonshot AI: Kimi K3.1 Pro and a priced cache write

Moonshot's Kimi K3.1 Pro costs $2.00 input and $10.00 output per million tokens, with cached input at $0.20. Its caching works like Claude's: a 5-minute write is billed at $2.00 and a 1-hour write at $4.00. The highspeed version doubles the base rates to $4.00 / $20.00, and Kimi K3 costs $3.00 / $15.00.

The same Kimi K3.1 Pro is also served by Together AI and Fireworks AI at $2.00 / $10.00: the host changes the rate limits and the region, not the price. The cheaper K2 models (kimi-k2.6 and kimi-k2.7-code, both $0.95 / $4.00) are no longer open to new accounts, and Moonshot plans to retire them by the end of 2026: budget a new project on the K3 line. Batch inference on K3.1 Pro and K3 is billed at 60% of the standard price.

Kimi K3.1 Pro
$2.00 / $10.00
1-hour cache write
$4.00
See Kimi API pricing →
06

DeepSeek: two models, two clocks

DeepSeek sells two models through one API. At peak hours, deepseek-flash (V4.1 Flash) costs $0.30 input and $1.20 output per million tokens and deepseek-v4-pro $1.32 / $3.96; off-peak rates are half, so $0.15 / $0.60 and $0.66 / $1.98. A cache hit on Flash costs $0.006 at peak. Both have a 1M-token context.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; weekends are off-peak in full. A batch job scheduled outside those windows pays half. Concurrency is capped at 2,500 requests for Flash and 500 for V4 Pro. If its data location rules DeepSeek out, our DeepSeek alternatives page lists other options.

Flash at peak
$0.30 / $1.20
Flash off-peak
$0.15 / $0.60
See DeepSeek pricing →
07

Mistral AI: Mistral Large 4 under $1 of input

Mistral Large 4 costs $0.70 input and $2.10 output per million tokens, Mistral Small 4 $0.15 / $0.60, and Ministral 3 3B $0.10 for input and output alike. Mistral also hosts third-party models on the same API, such as DeepSeek V4.1 Flash 1017 at $0.30 / $1.20. Prices are shown in USD or EUR, which helps a European team budget without currency drift.

One catch: the FAQ on Mistral's commercial pricing page still uses $0.5 / $1.5 as its example price for Mistral Large, below the API table's $0.7 / $2.1. Budget on the API table. Batch processing takes 50% off. See Mistral AI alternatives for comparable APIs.

Mistral Large 4
$0.70 / $2.10
Ministral 3 3B
$0.10 / $0.10
See Mistral pricing →
08

Together AI: open models, with a US surcharge

Together AI serves open models on serverless endpoints: DeepSeek V4.1 Flash 1017 at $0.30 input and $1.20 output per million tokens, GLM-5.3 at $1.40 / $4.40, Kimi K3.1 Pro at $2.00 / $10.00 and gpt-oss-120B at $0.15 / $0.60.

Several models have a US geo-pinned version, at a higher price: DeepSeek V4.1 Flash goes to $0.375 / $1.50, 25% more, and GLM-5.3 to $1.75 / $5.50. Pick the pinned version only if your contract requires US processing.

DeepSeek V4.1 Flash
$0.30 / $1.20
US geo-pinned
$0.375 / $1.50
See Together AI pricing →
09

Fireworks AI: five modes on one endpoint

Fireworks prices the same model in several modes. DeepSeek V4.1 Flash costs $0.30 input and $1.20 output per million tokens in Standard and $0.375 / $1.50 in Priority; GPT OSS 120B costs $0.15 / $0.60 and Kimi K3.1 Pro $2.00 / $10.00. Batch takes 50% off input and output, and Spot takes 30% off on best-effort capacity for select models, with throughput that can drop to zero.

Cached input is billed at half price or less, with no write fee: on a workload with a long repeated prompt, that removes the write premium you pay at Anthropic or Moonshot.

DeepSeek V4.1 Flash
$0.30 / $1.20
Batch
50% off
See Fireworks pricing →
10

DeepInfra: the discounted host

DeepInfra lists DeepSeek-V4.1-Flash-1017 at $0.098 input and $0.294 output per million tokens, 51% off its $0.20 / $0.60 list price: the lowest rate for that model in this index. DeepSeek V4 Pro costs $1.30 / $2.60 there, against $1.32 / $3.96 at DeepSeek's own peak rate.

Discounts on a host are not contracts: a percentage shown next to a model can change. You have to add a card or prepay before the first call, and there is no long-term contract.

V4.1 Flash 1017
$0.098 / $0.294
V4 Pro
$1.30 / $2.60
See DeepInfra pricing →
11

Groq: open models priced by speed

Groq runs open models on its own LPU hardware. GPT OSS 120B costs $0.15 input and $0.60 output per million tokens at 500 tokens per second, GPT OSS 20B $0.075 / $0.30 at 1,000 tokens per second, and Qwen3.8-27B $0.80 / $4.00. Both GPT OSS models have a 131,072-token context, far from the 1M of the big APIs.

The token price matches Fireworks and Together on GPT OSS 120B: what Groq sells is speed. Batch and Flex processing, and the higher rate limits, come with the Developer plan.

GPT OSS 120B
$0.15 / $0.60
GPT OSS 20B
$0.075 / $0.30
See Groq pricing →
12

OpenRouter: list prices, plus a fee on credits

OpenRouter routes one API key to many providers and bills inference at each provider's list price. Its margin is a 5.5% fee when you buy credits, with a $0.80 minimum, or 8% on the Business plan. Bring your own provider keys (BYOK) and the first $25,000 of list-price inference a month goes through with no fee, then 5%.

25+ free models are capped at 50 requests a day, or 1,000 a day once your account has bought at least $10 of credits. An attempt that errors is not charged, so a fallback chain bills only the run that answers. OpenRouter also takes 1% off usage costs if you opt in to logging your prompts and completions, a trade to keep for non-sensitive traffic. Prices exclude taxes. Our best AI gateway ranking and OpenRouter alternatives compare the routing layers.

Credit fee
5.5%, $0.80 minimum
BYOK
$25,000 a month fee-free
See OpenRouter pricing →
Pay less

Ten traps and one tip to pay less per token

Comparison tables list two numbers per model. These are the lines further down the pricing pages that change what you really pay, each with its number.

01

Trap 1: a launch price that doubles

Gemini 4 Carbon's $2.00 / $10.00 is a launch rate: Google's table lists $4.00 / $20.00 from the start of December. Gemini 3.8 Flash goes from $0.75 / $3.75 to $1.50 / $7.50 in 2027, and its cache storage from $0.50 to $1.00 per million tokens per hour. Budget a new model at its second price.

Gemini 4 Carbon
$2.00 then $4.00 input
02

Trap 2: the older model costs more

New generations are often cheaper. GPT-6 Astra costs $10.00 / $50.00 against $8.00 / $40.00 for GPT-6.2 Astra; Claude Fable 5.1 costs $10 / $50 against $8 / $40 for Fable 5.2. A code base that pins the older version pays 25% more on every token: on 50 million output tokens a month, $2,500 instead of $2,000.

GPT-6 Astra
$10.00 / $50.00
03

Trap 3: a long prompt reprices the whole request

Each provider has its own threshold. At OpenAI, a prompt over 272K input tokens doubles input and cache rates and multiplies output by 1.5, for the full request. At xAI, a request that reaches 200k tokens bills every token at the long-context rate ($4.00 / $12.00 on Grok 4.7). Claude Haiku 5.5 moves from $0.10 / $0.50 to $0.50 / $2.50 past 100,000 tokens, five times the rate, cache reads included in the count.

Haiku 5.5 past 100K
$0.50 / $2.50
04

Trap 4: writing to the cache is billed

A cache hit is cheap; storing the prefix is not free. At Anthropic, a 5-minute write costs 1.25x the input price and pays off after one cache read; a 1-hour write costs 2x and pays off after two. OpenAI bills cache writes too: $10.00 per million on GPT-6.2 Astra for 30 minutes, $20.00 for 12 hours. Moonshot bills a 5-minute write at the input price ($2.00 on Kimi K3.1 Pro). Fireworks charges no write fee. If your users come back less than once per cache window, caching costs more than it saves.

Anthropic 1-hour write
2x input
05

Trap 5: DeepSeek's peak hours

DeepSeek's peak rate applies 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday: the European working morning and part of the American evening. Off-peak, deepseek-flash costs $0.15 / $0.60 instead of $0.30 / $1.20. A nightly job moved out of those windows halves its bill without changing model.

Flash off-peak
$0.15 / $0.60
06

Trap 6: keeping data in the US costs extra

Pinning inference to the US multiplies every Claude 4.6 and later token by 1.1. OpenAI adds 10% for regional processing on models released since early 2026. Together AI's US geo-pinned DeepSeek V4.1 Flash costs $0.375 / $1.50 instead of $0.30 / $1.20, 25% more. xAI's US regional endpoint bills 1.1x the global rates, and Fireworks charges a 50% premium for US-only serverless on select models. Ask for residency only where a contract needs it.

Claude, US-only
1.1x
07

Trap 7: fast modes double the rate

Speed has a price list of its own. Claude Fable 5.2 in fast mode costs $16 / $80, twice its standard rate, and fast mode does not work with the Batch API. OpenAI's Fast mode costs twice the Standard rate, and Ultrafast three times, then six times in 2027. Turn them on for the user-facing call that needs them, not for the whole application.

Fable 5.2 fast
$16 / $80
08

Trap 8: tools bill per call, on top of tokens

Web search is priced per call: $10 per 1,000 searches at Anthropic, $10 per 1,000 calls at OpenAI, $5 per 1,000 at xAI, plus the retrieved content billed as input tokens. Gemini grounding with Google Search is free for 5,000 requests a month, then $14 per 1,000, and one request can trigger several searches. An agent that searches on every turn spends more on search than on tokens.

Gemini grounding
$14 per 1,000
09

Trap 9: the gateway fee on small top-ups

OpenRouter's 5.5% credit fee has a $0.80 minimum. On a $10 top-up, that is $0.80, 8% of the purchase; on $100 it is $5.50. Top up in larger amounts, or bring your own keys: BYOK is fee-free up to $25,000 of list-price inference a month.

$10 top-up
$0.80 fee
10

Trap 10: the same text, more tokens

A price per million tokens compares badly across providers, because each one cuts text into tokens its own way. Anthropic states that Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text than Claude Sonnet 4.6 and earlier; Moonshot counts 1 token as roughly 3 to 4 characters. Before you compare two APIs on price, run the same sample of your own prompts through each provider's token counter, then multiply.

Claude 4.7 and later
about 30% more tokens
11

The tip: one open model, several prices

An open-weight model is sold by its maker and by hosts. DeepSeek V4.1 Flash costs $0.30 / $1.20 at DeepSeek's peak rate, at Together AI, at Fireworks and at Mistral; $0.15 / $0.60 off-peak at DeepSeek; and $0.098 / $0.294 at DeepInfra with its current discount. GPT OSS 120B costs $0.15 / $0.60 at Groq, Fireworks and Together alike, so compare those hosts on speed and limits instead.

DeepInfra V4.1 Flash
$0.098 / $0.294
How pricing works

What an LLM API really costs per month

How tokens add up, one workload priced on sixteen models, and what caching and reasoning change.

01

How the meter runs

Every call is billed on two counters: the tokens you send (your prompt, the conversation history, the documents you attach) and the tokens the model writes back. Output usually costs three to five times input: $8 against $40 on GPT-6.2 Astra and Claude Fable 5.2, $2 against $6 on Grok 4.7. Gemini 3.5 Flash-Lite goes to eight times, $0.30 against $2.50. Reasoning models also bill their thinking as output, even when you never see it: Google states its output price includes thinking tokens, and xAI lists reasoning tokens as a billed token type.

Output vs input
3x to 8x
02

One workload, priced on the frontier models

Take a support assistant answering 100,000 conversations a month, each with 2,000 input tokens and 500 output tokens: 200 million input tokens and 50 million output tokens. At standard rates, GPT-6.2 Astra and Claude Fable 5.2 cost $3,600 a month each, Claude Opus 5.5 $1,800, and Claude Sonnet 5.5, GPT-6.2 Sol, Kimi K3.1 Pro and Gemini 4 Carbon $900 (Carbon $1,800 at its next rate). Grok 4.7 comes to $700.

GPT-6.2 Astra
$3,600 a month
Grok 4.7
$700 a month
03

The same workload on budget models

On the same 200 million input and 50 million output tokens, DeepSeek V4 Pro costs $462 at peak rates, Mistral Large 4 $245, DeepSeek V4.1 Flash $120 at peak or $60 off-peak, GPT OSS 120B on Groq $60, and GPT-6 Luna or Claude Haiku 5.5 $45 (Haiku as long as each prompt stays under 100,000 tokens). Ministral 3 3B comes to $25, the lowest bill on the cards for this workload (GPT OSS 20B on Groq $30, DeepSeek V4.1 Flash on DeepInfra about $34): the same month costs 144 times more on GPT-6.2 Astra.

Mistral Large 4
$245 a month
GPT-6 Luna
$45 a month
04

What caching changes

If 1,500 of the 2,000 input tokens are a fixed system prompt read from cache, Claude Sonnet 5.5 drops from $900 to about $615 a month: $15 for 150 million cached tokens at $0.10, $100 for the 50 million uncached, $500 for output, before cache writes. GPT-6.2 Astra drops from $3,600 to $2,460. Output does not shrink with caching: on Astra it is $2,000 of the $2,460.

Sonnet 5.5, cached
about $615
05

What thinking tokens add

If a reasoning model writes 1,000 hidden thinking tokens per conversation, the same month adds 100 million output tokens: $1,000 more on Sonnet 5.5, more than the $900 base bill, and $4,000 more on GPT-6.2 Astra. Set a reasoning effort per call, and keep the deepest setting for the questions that need it. To see where tokens go call by call, our best LLM observability tools ranking lists the tracing layers.

Sonnet 5.5
+$1,000 a month
06

Through AWS or Azure

OpenAI models on Amazon Bedrock and Microsoft Azure are billed through those services. Claude Platform on AWS bills through AWS Marketplace in Claude Consumption Units at $0.01 each, after rating your tokens at Anthropic's standard per-model rates. Buying through your cloud puts the spend on an existing commitment, not necessarily at a lower rate.

Claude on AWS
$0.01 per CCU

Do you make a tool in this category? Email us: hello@hackceleration.com

Our method

How we collected the prices

For each API, we opened the provider's official pricing page or model page in 2026 and recorded the standard pay-as-you-go price per million tokens for input, cached input and output, on short prompts, with the conditions written next to it: prompt-length thresholds, cache-write prices, batch and fast-mode multipliers, peak hours, residency surcharges. Model makers are shown with their flagship or reference model and their cheapest text model; hosts with an open model they share, so the same model can be compared across hosts. Where a price is temporary (a launch rate, a discount, a peak hour), the card says so. The monthly budgets multiply those rates by one fixed workload, 200 million input and 50 million output tokens, without tax. We compare prices, not answer quality: to choose a model, read our best LLM API ranking.

  1. SourceOnly the provider's pricing or model page
    Official page
  2. UnitInput and output, in USD
    1M tokens
  3. TierPay-as-you-go, short prompts
    Standard
  4. ConditionsCache, batch, long prompts, hours
    Written
  5. HostsDeepSeek V4.1 Flash, GPT OSS 120B
    Same model
  6. Budget200M input, 50M output tokens
    One workload
12APIs
250Mtokens per budget

The index is editorial: a sponsored placement is always labeled on its card.

FAQ

LLM API pricing: your questions

  • Which LLM API has the largest context window for the price?

    Several models priced here pair a window of about one million tokens with an entry-level price, among them GPT-6 Luna, Claude Haiku 5.5 and DeepSeek V4.1 Flash. GPT-6 Luna has a 1,050,000-token context window at $0.10 input and $0.50 output, but a prompt over 272K input tokens is billed at 2x input and 1.5x output for the whole request, so $0.20 / $0.75. Claude Haiku 5.5 has a 1M-token window at $0.10 / $0.50, and $0.50 / $2.50 once a prompt passes 100,000 tokens. DeepSeek's deepseek-flash has a 1M-token context at $0.30 / $1.20 at peak, and DeepInfra lists DeepSeek-V4.1-Flash-1017 with a 1024k context at $0.098 / $0.294 with its current discount. GPT OSS 120B on Groq stops at 131,072 tokens.

  • How long does a cached prompt stay in the cache?

    From minutes to hours, and the longer option costs more to write. Anthropic offers 5 minutes (write at 1.25x input) or 1 hour (2x); OpenAI prices 30-minute and 12-hour writes ($2.50 and $5.00 per million on GPT-6.2 Sol). Gemini 4 Carbon supports implicit caching only, so you cannot set its duration at all.

  • Does the Batch API discount apply to every model?

    No. xAI says not every model accepts Batch API requests, and its batch discount is 20%, not 50%, on the Grok models it lists; Moonshot bills batch at 60% of the standard price, on Kimi K3.1 Pro and K3. Anthropic's fast mode is not available with the Batch API, and Claude Managed Agents sessions have no batch mode. At Groq, Batch and Flex processing come with the Developer plan. Check the model page before you plan a nightly job around a discount.

  • Do you pay for a request that fails?

    Not for its model tokens on OpenRouter: an attempt that errors is not charged, so a fallback chain bills only the run that answers. The bigger risk is retries you control. In 2026, one developer traced a bill at 40 times normal to a retry policy that resent a long prompt all night; cap retries and log their cost.

  • How often do LLM API prices change?

    Often enough to recheck every quarter. Google publishes dated price changes in its table, DeepSeek reserves the right to adjust its prices, Together AI runs promotions with an end date (in 2026, Qwen4 Max at $1.05 / $3.15 instead of $1.50 / $4.50), and Moonshot plans to retire its K2 models by the end of 2026. OpenAI's Ultrafast multiplier is already scheduled to rise from 3x to 6x in 2027.

  • Can you negotiate a volume discount on an LLM API?

    Yes, at enterprise level. Google's Enterprise tier lists volume-based discounts and provisioned throughput, OpenRouter offers fee discounts on Enterprise contracts with prepaid credits and annual commitments, and Fireworks sells reserved throughput by the minute with a 99.9% uptime SLA. Below that, the levers are batch, caching and off-peak hours.

  • Do LLM API prices include VAT?

    No. Prices are listed in USD before tax: OpenRouter states its prices exclude applicable taxes and may add VAT or GST on invoices, and Moonshot's prices also exclude applicable taxes. A European team should add its VAT rate to each per-token price, or pick a provider that shows its prices in EUR, as Mistral does, to avoid currency drift on top.

Hack'celeration Lab

Get the next pricing index in your inbox

LLM API prices move every few weeks. We'll write when one changes.

No spam. Unsubscribe anytime.