Compare 239 AI models: LLM API pricing, context and capabilities

Prices from official vendor price lists · October 2026

Compare 239 AI models and LLM APIs from 41 vendors — GPT, Claude, Gemini, Grok, DeepSeek, Qwen and more — by API pricing per 1M tokens, context window and capabilities. Output prices run from $0.06 to $180 per 1M tokens, a spread of roughly 3,000×; the median across the 199 models with a paid per-token price is $2.

That gap is why an AI model comparison pays off before you commit: most workloads do not need a frontier LLM, and 80 of the 239 models have a free tier to test on first. Filter the table by task or vendor, then open a model to see its pricing, limits and what your own token volume would cost.

#	AI model	API $/1M (in / out)	Context window	Price / mo	Output $/1M	Overview
61	DeepSeek V3.1 Terminus DeepSeek	$0.27 / $1	163.8K tokens	API only	1	View
62	Mistral Medium 3.5 Mistral AI	$1.5 / $7.5	256K tokens	API only	7.5	View
63	Nemotron 3.5 Lightning NVIDIA	$0.08 / $0.2	262.1K tokens	API only	0.2	View
64	Nemotron 3 Super NVIDIA	$0.085 / $0.4	262.1K tokens	API only	0.4	View
65	Mistral Small 4 Mistral AI	$0.15 / $0.6	256K tokens	API only	0.6	View
66	Mercury 2 Inception	$0.25 / $0.75	128K tokens	API only	0.75	View
67	Trinity Large Thinking Arcee Ai	$0.25 / $0.8	262.1K tokens	API only	0.8	View
68	Qwen3 Coder Next Qwen	$0.12 / $0.8	262.1K tokens	API only	0.8	View
69	Mistral Large 3 2512 Mistral AI	$0.5 / $1.5	262.1K tokens	API only	1.5	View
70	Devstral 2 2512 Mistral AI	$0.4 / $2	262.1K tokens	API only	2	View
71	Nemotron 3 Nano 30B A3B NVIDIA	$0.05 / $0.2	262.1K tokens	API only	0.2	View
72	Solar Pro 3 Upstage	$0.15 / $0.6	131.1K tokens	API only	0.6	View
73	Ministral 3 14B 2512 Mistral AI	$0.2 / $0.2	262.1K tokens	API only	0.2	View
74	Ministral 3 8B 2512 Mistral AI	$0.15 / $0.15	262.1K tokens	API only	0.15	View
75	Ministral 3 3B 2512 Mistral AI	$0.1 / $0.1	131.1K tokens	API only	0.1	View
76	GPT-5.4 OpenAI	$2.5 / $15	1.05M tokens	API only	15	View
77	GPT-5.3-Codex OpenAI	$1.75 / $14	400K tokens	API only	14	View
78	GPT-Realtime-2.1 OpenAI	$4 / $24	128K tokens	API only	24	View
79	GPT-Realtime-2.1 Mini OpenAI	$0.6 / $2.4	128K tokens	API only	2.4	View
80	Gemini 3.1 Flash-Lite Google	$0.25 / $1.5	1.05M tokens	API only	1.5	View

Browse by need

Reasoning models169 models

Coding assistants17 models

Vision and images156 models

Long context 1M+82 models

Free tier available80 models

Open weights88 models

How to choose an AI model in 2026

The table above compares 239 AI models from 41 vendors on the numbers that decide the bill: API price per million tokens, context window, subscription cost and what each model can actually do. This guide explains how to read those numbers and which model fits which job. Prices come from the vendors' official price lists in USD; for dirhams multiply by 3.6725.

Two prices, not oneEvery API model bills input and output separately. Output is typically 3-5x dearer, and answers are what you pay for most.

Subscription is not APIA $20/mo chat plan and per-token API access are different products. Heavy automation almost always belongs on the API.

Caching changes the mathsRepeated context can cost a tenth of the base rate. For long system prompts this matters more than the headline price.

AI model pricing: what a million tokens costs

Output price per 1M tokens, standard tier, October 2026. A million output tokens is roughly 700,000 words - about seven full-length novels.

GPT-6 Astra$50

Claude Fable 5.1$50

Claude Opus 5$25

GPT-5.6 Sol$20

GPT-5.6 Terra$12

Claude Sonnet 5$10

Mistral Medium 3.5$7.50

Grok 4.6$6

Qwen3.8-Max$6

Claude Haiku 4.5$5

Gemini 3.8 Flash$3.75

DeepSeek V4 Pro$3.15

Mistral Large 3$1.50

GPT-5.6 Luna$1.20

Qwen3.8-Flash$0.38

DeepSeek V4 Flash$0.18

Green - frontier tier. Amber - mid tier. Grey - budget tier. Gemini 3.8 Flash is promotional through 31 December 2026 and doubles from 1 January 2027. DeepSeek prices are the Alibaba Model Studio USD listing; DeepSeek's own price list is published in yuan only.

Best AI model for your task

Chat and writingCodingLong documentsVoice and audioHigh volume, low cost

Claude Sonnet 5~~$2 / $10 per 1M~~The balanced default: frontier-level writing at mid-tier prices, 1M context, adjustable thinking.

GPT-5.6 Terra~~$2 / $12 per 1M~~Strong general reasoning with a 1.05M window and the widest tooling ecosystem.

Gemini 3.8 Flash~~$0.75 / $3.75 per 1M~~Cheapest of the current flagships while promotional pricing lasts, with a free tier for testing.

Claude Opus 5~~$5 / $25 per 1M~~Built for agentic coding and long refactors; 1M context, 128K output.

GPT-5.3-Codex~~$1.75 / $14 per 1M~~OpenAI's dedicated coding model, 400K context, priced well below the flagships.

Grok 4.6~~$2 / $6 per 1M~~500K context and configurable reasoning depth at the lowest output price of any frontier model.

Claude Fable 5.1~~$10 / $50 per 1M~~Long-horizon agentic work where accuracy outweighs cost; cache reads at 2.5% of base.

Claude Opus 5.5~~$4 / $20 per 1M~~1M context with vision, prompt caching and batch pricing for large document sets.

DeepSeek V4.1 Flash~~$0.3 / $1.2 per 1M~~1M context and open weights at a budget price; off-peak billing halves it.

GPT-Realtime-2.1~~$4 / $24 per 1M~~Speech-to-speech with 128K context for live agents and phone-grade latency.

GPT-Realtime-2.1 Mini~~$0.6 / $2.4 per 1M~~Live voice with the same 128K context at a fraction of the full Realtime price.

Gemini 3.5 Transcribe~~$2 / $12 per 1M~~Batch transcription with diarisation, word timestamps and custom vocabulary.

DeepSeek V4 Flash~~$0.065 / $0.18 per 1M~~1M context at a rounding-error price; off-peak billing halves it again.

Qwen3.8-Flash~~$0.11 / $0.38 per 1M~~Open weights, 1M context, and a free quota on Model Studio for trials.

GPT-5.6 Luna~~$0.20 / $1.20 per 1M~~The cheapest way to stay inside the OpenAI ecosystem for classification and routing.

AI releases and price changes in 2026

Every major vendor shipped a new generation this year. If a comparison you are reading still names GPT-5.4, Claude 4.8 or Llama 4, it predates the current line-up.

September

Meta retires Llama, launches MuseLlama disappears from Meta's developer docs; the paid API now sells Muse Spark 1.3 at $1.25 / $4.25 with a 1M window, plus the Apache-licensed Muse Glimmer 30B for self-hosting.

September

Qwen3.8 tops Alibaba's line-upQwen3.8-Max at $2 / $6 with a 1M context.

August

Sonnet 5 keeps introductory pricingAnthropic cancels the planned rise to $3 / $15 - $2 / $10 becomes the standard rate.

Summer

GPT-6 Astra and the 5.6 familyOpenAI splits the mid tier into Sol, Terra and Luna, with a 1.05M window across the board and Fast mode at double rate.

Spring

DeepSeek V4 replaces V3 and R1Flash and Pro variants, 1M context, 384K max output, and off-peak pricing at half rate.

Monthly cost: API usage and subscriptions

Worked examples at current rates, assuming a 1:3 input-to-output ratio. Dirham figures use 3.6725 per dollar.

Workload	Budget tier	Mid tier	Frontier tier
Support bot, 5M tokens/mo	$1 (AED 3)	$48 (AED 174)	$100 (AED 367)
Coding assistant, 30M tokens/mo	$5 (AED 17)	$285 (AED 1,047)	$600 (AED 2,204)
Document pipeline, 100M tokens/mo	$15 (AED 56)	$950 (AED 3,489)	$2,000 (AED 7,345)
One person, chat only	Free tier	$20 (AED 73)	$100-200 (AED 367-734)

Budget tier calculated on DeepSeek V4 Flash, mid tier on GPT-5.6 Terra, frontier tier on Claude Opus 5. Batch processing cuts these figures roughly in half; prompt caching cuts the input share by up to 90%.

Consumer subscriptions worth knowing

Plan	Price	What you get
ChatGPT Go	$8/mo (AED 29)	Luna and Thinking Mini, 54K context; no Sol
ChatGPT Plus	$20/mo (AED 73)	GPT-6 Astra, Sol, Terra, 256K reasoning context
Claude Pro	$17-20/mo (AED 62-73)	Opus, Sonnet, Haiku, Claude Code, 200K context
Google AI Pro	$19.99/mo (AED 73)	Gemini 3 Pro, Deep Search, 5 TB storage
SuperGrok	$30/mo (AED 110)	Grok 4.6, image and video generation
Mistral Vibe Pro	$14.99/mo (AED 55)	6x messages, all-day coding, $10 API credits

Annual billing is not offered on ChatGPT Go, Plus or Pro. Claude Pro is $17/mo when paid yearly. Free tiers exist on Gemini, DeepSeek, Qwen and Mistral; ChatGPT Free is limited to Luna.

Who each option suits

Students

DeepSeek V4

Qwen3.8-Flash

Free tiers and sub-cent pricing, strong in Arabic and English. Enough for coursework, notes and translation.

Businesses / startups

GPT-5.6 Terra

Claude Opus 5

Frontier quality for code and documents, 1M context, batch and caching discounts that make the bill predictable.

Content creators

GPT-5.6 Terra

Text, images, voice and live web search in one window, with a free tier to start on.

Open weights, self-hostable

DeepSeek V4

Muse Glimmer 30B

Mistral Small 4

Weights you can download and run on your own infrastructure - no vendor lock-in, no per-token bill.

Frequently asked questions about AI models

Which AI model is the best in 2026?

There is no single winner. For reasoning and long agentic tasks the frontier tier is GPT-6 Astra and Claude Fable 5.1, both at $50 per 1M output tokens. For everyday chat and writing, Claude Sonnet 5 ($10) and GPT-5.6 Terra ($12) deliver most of that quality at a fifth of the price. For volume work DeepSeek V4 Flash and Qwen3.8-Flash cost under $0.40. Filter the table by capability and sort by output price to see the trade-off for your own case.

What is the cheapest AI API?

DeepSeek V4 Flash at roughly $0.065 input and $0.18 output per 1M tokens, with off-peak rates halving that again. Qwen3.8-Flash is close at $0.11 and $0.38 and adds open weights. Both offer a 1M-token context window, so cheap no longer means short.

How much does an AI model cost per month in the UAE?

API billing is in US dollars everywhere; convert at roughly 3.6725 dirhams per dollar. A single-person chat subscription runs AED 29 to 73 a month. A production support bot handling 5M tokens costs about AED 3 on a budget model and AED 367 on a frontier one. Cards issued in the UAE are accepted by every major vendor.

What is a context window and how much do I need?

The context window is how much text the model holds in one conversation, counted in tokens of roughly 0.75 words each. Most current flagships offer 1M tokens, about 750,000 words. Ordinary chat rarely exceeds 30K. Long windows matter for whole codebases, legal bundles and multi-hour transcripts, and some vendors charge more above 200K.

Is the API cheaper than a subscription?

For one person doing ordinary chat, no - a $20 plan is hard to beat. The API wins as soon as usage is automated or shared: you pay only for tokens consumed, batch processing cuts the bill in half, and prompt caching cuts repeated input by up to 90%. The crossover typically arrives around 5M tokens a month.

What is the difference between input and output tokens?

Input is everything you send - the prompt, system instructions and documents. Output is what the model writes back, including the hidden thinking tokens that reasoning models generate. Output costs three to five times more, so answer length drives the bill far more than prompt length.

Which models have open weights?

Around 90 models in this rating ship downloadable weights: the Qwen and GLM families, DeepSeek, Mistral Apache-licensed models including Large 3 and Small 4, Gemma 4, Meta Muse Glimmer 30B, Granite, EXAONE, Solar, Nemotron and Phi. Use the Capabilities filter to isolate them.

How is this rating compiled?

Prices and context windows are taken from each vendor official price list and model documentation, in USD, and rechecked as vendors update them. Capabilities come from the same documentation. An empty cell means the vendor does not publish that figure rather than that the feature is missing.

How this rating is built

Every price in this table comes from the vendor’s own published price list, not from a reseller or an aggregator. All 29 sources were opened and re-read in October 2026; the full list sits in the Sources section below.

What the numbers mean

PriceUS dollars per 1M tokens, input then output, standard tier. Batch, caching and off-peak discounts are not folded into the headline number — the calculator on each model page applies them.

Context windowThe maximum the vendor advertises. Usable context is normally shorter: quality drops well before the stated ceiling on every model that has been benchmarked.

DirhamsConverted at AED 3.6725 to the dollar. Vendors bill in USD, so the dirham figure is indicative, not a quoted rate.

An empty cellMeans the price could not be confirmed against a public vendor price list. It never means the model is free.

Where a vendor bills differently

Cohere moved the Command A family to per-instance-hour billing, so a per-million-token figure does not exist for those models.
Thinking Machines charges prefill, sample and train separately; prefill maps to input and sample to output in this table.
DeepSeek publishes peak and off-peak rates that differ by roughly half, depending on the hour the request lands.
Baidu, ByteDance, Tencent, StepFun and InclusionAI publish in yuan; the dollar figure is converted and the source page is linked.

Five vendors publish nothing

Liquid, Poolside, ByteDance Seed, Nex AGI and TheDrummer have no public rate card. Liquid offers a free tier below $10M revenue and quotes enterprise deals on request; Poolside sells self-managed deployments; the other three ship weights rather than an API. Their rows say so instead of showing a blank.

Alibaba, MiniMax and NVIDIA publish a model catalogue but keep prices behind a console login, so those figures come from the catalogue page and are marked in the table.

How often this is re-checked

Prices in this category move fast — they fell roughly 80% between early 2025 and early 2026, and a model shipped three months ago can already be superseded. The table is reviewed monthly, and out of cycle whenever a vendor ships a new version or changes a published rate.

Any figure here can be checked against its source: every vendor price page used is linked below, with the line that was read from it.

Sources

Twenty-nine vendors publish a price page that can be opened without an account. Each link below goes to the page the figures in this table were read from, not to a vendor homepage. All checked in October 2026.

OpenAI — API pricingGPT-5.6 Sol $4 / $20 per 1M

Anthropic — pricingOpus 5 $5 / $25 per MTok

Google — Gemini API pricingGemini 3.8 Flash input $0.75

xAI — models and pricingGrok 4.6 $2.00 / $6.00 per 1M

DeepSeek — API pricingV4 Pro $1.05 / $3.15 per 1M

Z.ai — pricing docsGLM-5.3 $1.40 / $4.40 per 1M

Mistral AI — API pricingMedium 3.5 $1.50 / $7.50 per 1M

Moonshot AI — Kimi pricingkimi-k3 $3.00 / $15.00 per 1M

Meta — model APImuse-spark-1.3 $1.25 / $4.25 per Mtok

Microsoft — Foundry model pricingPhi-4 128K $0.125 / $0.50 per 1M

Amazon — Bedrock pricingPer-provider, per-region token rates

IBM — watsonx.ai pricinggranite-4h-small $0.0636 / $0.265 per 1M

Cohere — pricing docsCommand A $40.00 per instance-hour

AI21 Labs — pricingJamba Large $2 / $8 per 1M

Perplexity — pricingSonar Pro $3 / $15 per 1M

Upstage — API pricingSolar Pro 4 $0.30 / $1.20 per 1M

Inception — MercuryMercury 2.5 $0.20 / $0.75 per 1M

Aion Labs — model docsAion 3.0 $3.00 / $6.00 per 1M

Arcee AI — pricingTrinity Large Thinking $0.25 / $0.80

Perceptron — Mk1 model page$0.15 / $1.50 per 1M, cached $0.0375

Sakana — console pricingfugu-ultra-v2.0 $5 / $30 per 1M

Thinking Machines — Tinker modelsInkling prefill $1.87, sample $4.68

Xiaomi — MiMo pricingmimo-v2.5-pro $0.435 / $0.87 overseas

Meituan — LongCat API pricingLongCat-2.0 $0.75 / $2.95 per 1M

ByteDance — Volcano Engine Arkdoubao-seed-2.1-pro ¥6 / ¥30 per 1M

Baidu — Qianfan pricingERNIE 5.1 ¥0.004 / ¥0.018 per 1K

Tencent — Hunyuan billingHunyuan-a13b ¥0.5 / ¥2 per 1M

StepFun — pricing detailsstep-3.7-flash ¥1.35 / ¥8.10 per 1M

InclusionAI — Ling pricingLing-2.6-1T ¥4.50 / ¥18.00 per 1M

Catalogue only, prices behind a login

Alibaba — Model Studio catalogueRates shown inside the console

MiniMax — model docsContext and capability, no rate card

NVIDIA — NIM catalogueFree and partner endpoints, no rates

Liquid, Poolside, ByteDance Seed, Nex AGI and TheDrummer publish no public rate card at all — their rows say so rather than showing a figure from a third party.

Ratings 5

Avg. score 5

Rate this page

Help us improve