AI & data ranking
10 Best LLM API Providers of 2026
By the FruitPunch research deskUpdated November 20, 2026Next review January 4, 2027
The short answer
The cheapest LLM API per million tokens is Fireworks AI, from $0.05 per 1M input tokens, but Google Gemini API is the best LLM API provider for developers and small teams in 2026 because it has a free tier with free input and output tokens and paid models from $0.10 per 1M input tokens (Gemini 2.5 Flash-Lite). OpenAI API suits teams that want one vendor for text, images and voice, with GPT-6 Luna at $0.10 per 1M input tokens, and the Anthropic API suits long-document and coding work, with a 1M token context window at standard pricing on most current Claude models. Teams that must buy through AWS or Azure contracts should look at Amazon Bedrock or Azure OpenAI instead; all prices were checked on vendor pages in November 2026.
The podium
Our verdict
Start with the Google Gemini API: its free tier lets you build and test without a card, and moving to paid unlocks context caching, the 50% Batch API discount and keeps your prompts out of Google's product training. If you would rather standardize on one vendor that covers text, image generation, realtime voice and transcription, OpenAI API is the closest runner-up, and the Anthropic API is the pick when your prompts are long or your product is a coding agent. For pure cost on open-weight models, Fireworks AI, Groq and Together AI all list models at or below $0.15 per 1M input tokens, and Groq adds a free plan and speeds of up to 1,000 tokens per second on GPT OSS 20B. Two mistakes to avoid: comparing input prices only, when output tokens usually cost 4 to 5 times more, and assuming a launch price lasts, since Google lists Gemini 4 Carbon at $2.00 per 1M input tokens through November 30, 2026 and $4.00 from December 1.
About this ranking
This page is for developers, indie builders and small product teams choosing which language model API to put behind an app, a chatbot, an internal tool or a coding agent. We compared 10 providers on three things: the quality of the models they serve, how quickly a small team can integrate them (self-serve keys, OpenAI-compatible endpoints, spend limits, docs), and the per-token price of the cheapest useful model. On November 20, 2026 we reopened each vendor's own pricing page, model list or docs and compared documented features and published USD prices; we did not run our own benchmarks. Scores weigh quality at 40%, ease of use at 25% and value at 35%. Amazon Bedrock, Azure OpenAI and Cohere load their per-token prices with scripts that did not render for us, so those three prices are marked pending.
Figures worth knowing
- 54.6%In the St. Louis Fed's Real-Time Population Survey, 54.6% of US adults aged 18 to 64 used generative AI in August 2025, up 10 percentage points in a year.Source: Federal Reserve Bank of St. Louis
- 19.8%In the Census Bureau's Business Trends and Outlook Survey, 19.8% of U.S. businesses reported using AI in the collection period ending May 3, 2026.Source: U.S. Census Bureau
- 20%Between December 2025 and May 2026, between 17% and 20% of U.S. businesses told the Census Bureau they were using AI in their business functions, and 20% to 23% expected to within six months.Source: U.S. Census Bureau
Every number, side by side
The taste test
Tap a column to sort. Data from each maker's own pages.
| Free plan or trial | ||||||
|---|---|---|---|---|---|---|
| 1Google Gemini API | 9.3 | 9.2 | 9.0 | 9.6 | $0.10 | Free plan |
| 2OpenAI API | 9.1 | 9.4 | 9.2 | 8.8 | $0.10 | No |
| 3Anthropic API | 9.0 | 9.5 | 8.9 | 8.4 | $0.10 | No |
| 4Groq | 8.7 | 8.0 | 9.0 | 9.2 | $0.07 | Free plan |
| 5Fireworks AI | 8.6 | 8.1 | 8.4 | 9.4 | $0.05 | No |
| 6Mistral AI | 8.6 | 8.2 | 8.5 | 9.0 | $0.10 | No |
| 7Together AI | 8.5 | 8.0 | 8.4 | 9.0 | $0.15 | No |
| 8Azure OpenAI | 8.1 | 9.0 | 7.0 | 7.8 | Check price | No |
| 9Amazon Bedrock | 8.0 | 9.0 | 7.0 | 7.6 | Check price | No |
| 10Cohere | 7.5 | 7.8 | 7.6 | 7.2 | Check price | No |
10 products compared
The full rankings
1 
Top pick
Google Gemini API
9.3out of 10
Google's developer API for Gemini models, with a free tier in Google AI Studio and pay-as-you-go paid tiers.
Best for: small teams that want a free tier to build on
The Gemini API comes first because it is the only provider of the three big frontier labs that lets a small team build, test and even run a light workload at no cost. Google's pricing page lists a free tier with free input and output tokens on models such as Gemini 3.5 Flash-Lite, Gemini 3.6 Flash and Gemini 2.5 Flash, plus Google AI Studio access. Moving to the paid tier adds higher rate limits, context caching and the Batch API at a 50% cost reduction, and Google says paid content is not used to improve its products. On price, Gemini 2.5 Flash-Lite is $0.10 per 1M text input tokens and Gemini 3.5 Flash-Lite is $0.30 in and $2.50 out. The top model, Gemini 4 Carbon, is paid only and its input price rises from $2.00 to $4.00 per 1M tokens on December 1, 2026, so budget at the later rate. The free tier trade-off is data: free tier content is used to improve Google's products.
Google Gemini API is the best choice for small teams and startups because its free tier includes free input and output tokens and paid models start at $0.10 per 1M input tokens. Move to the paid tier before you send customer data, since free tier content is used to improve Google's products.
What we like
- Free tier covers both input and output tokens on several Flash and Flash-Lite models
- Paid content is not used to improve Google's products
- Wide price ladder, from $0.10 per 1M input tokens up to Gemini 4 Carbon
Watch out for
- Free tier content is used to improve Google's products
- Gemini 4 Carbon and Gemini 4 Argon are not available on the free tier
- Gemini 4 Carbon input doubles from $2.00 to $4.00 per 1M tokens on December 1, 2026
2 
Ranked 2
OpenAI API
9.1out of 10
OpenAI's developer platform for GPT-6 text models, GPT-Image, realtime voice and transcription, billed per token.
Best for: teams that want text, image and voice from one vendor
OpenAI API is the runner-up because it covers the most ground from one account. The pricing page lists four GPT-6 text models: GPT-6 Luna at $0.10 in and $0.50 out per 1M tokens for high-volume tasks, GPT-6.1 Terra at $1.00 in, GPT-6.2 Sol at $2.00 in for everyday work and agents, and GPT-6.2 Astra at $8.00 in for the hardest projects. The same account also serves GPT-Image models, realtime voice and two transcription models, which saves a small team from stitching vendors together. Cost controls are practical: the Batch API cuts input and output prices by 50%, cached input on GPT-6 Luna is $0.01 per 1M tokens, and you can set a monthly budget after which requests stop. The gaps are a missing free API tier, since Playground use is billed like regular API use, and a 10% premium for data residency. API usage is also billed separately from any ChatGPT subscription.
OpenAI API is the best choice for teams that want text, image and voice models on one bill because one account serves GPT-6 text, GPT-Image and realtime voice, with GPT-6 Luna at $0.10 per 1M input tokens. Pick Gemini instead if you need a free tier to start.
What we like
- GPT-6 Luna matches the lowest frontier-lab input price at $0.10 per 1M tokens
- One account covers text, image, voice and speech-to-text models
- Monthly budgets and email alerts help small teams cap spend
Watch out for
- No free API tier listed on the pricing page; Playground use is billed
- API is billed separately from ChatGPT Plus, Business and Enterprise plans
- Data residency adds 10% to the standard rate
3 
Ranked 3
Anthropic API
9.0out of 10
Anthropic's Claude Developer Platform, serving Claude Fable, Opus, Sonnet and Haiku models per token.
Best for: coding agents and long-document apps
The Anthropic API ranks third overall but first on model quality for long inputs. Anthropic's pricing docs say Claude 4.6 and later models, except Claude Haiku 5.5, include the full 1M token context window at standard pricing, so a 900k-token request costs the same per token as a 9k-token one. That matters for apps that read whole codebases, contracts or research papers. Claude Sonnet 5.5 costs $2 in and $10 out per 1M tokens, Claude Opus 5.5 $4 in and $20 out, and Claude Fable 5.2 $8 in and $40 out. For volume work, Claude Haiku 5.5 is $0.10 in for prompts up to 100,000 tokens and $0.50 above that. The Batch API halves token prices and prompt caching hits cost a fraction of base input. Two cost notes: no free tier is listed on the pricing page, and Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text, so compare costs per task, not per token.
Anthropic API is the best choice for coding agents and long-document apps because most current Claude models include a 1M token context window at standard pricing. Budget-first teams should start on Claude Haiku 5.5 at $0.10 per 1M input tokens.
What we like
- A 900k-token request is billed at the same per-token rate as a 9k-token request on most current models
- Haiku 5.5 is $0.10 per 1M input tokens for prompts up to 100,000 tokens
- Code execution tool includes 1,550 free hours per organization per month
Watch out for
- Haiku 5.5 rises to $0.50 per 1M input tokens for prompts over 100,000 tokens
- Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text
- Top model Claude Fable 5.2 costs $8 in and $40 out per 1M tokens
4 
Ranked 4
Groq
8.7out of 10
GroqCloud inference API for open-weight models such as GPT OSS, with an OpenAI-compatible endpoint.
Best for: apps that need very fast responses
Groq is the pick when response speed matters more than having the biggest model. Its model docs list GPT OSS 20B at 1,000 tokens per second for $0.075 in and $0.30 out per 1M tokens, and GPT OSS 120B at about 500 tokens per second for $0.15 in and $0.60 out, both with a 131,072-token context window. Integration is quick because Groq documents an OpenAI-compatible base URL, so existing OpenAI SDK code needs only a new key and URL. There is a free plan with lower rate limits, and the Developer plan adds higher limits plus Batch and Flex processing. The trade-off is range: the self-serve production list is short, and models such as Llama 3.3 70B are marked Enterprise with contact-sales pricing. For voice bots, live autocomplete or agents that make many short calls, the speed and price make it a strong second provider next to a frontier lab.
Groq is the best choice for apps that need very fast responses because it lists GPT OSS 20B at 1,000 tokens per second for $0.075 per 1M input tokens. Pair it with a frontier model for harder tasks.
What we like
- Fastest listed throughput in this comparison, up to 1,000 tokens per second
- Works with the OpenAI SDK by changing the base URL
- Free plan to test before paying
Watch out for
- Small production catalog: two GPT OSS text models are self-serve priced
- Llama 3.3 70B and Llama 3.1 8B are listed as Enterprise, contact sales
- 131,072-token context window on the production text models
5 
Ranked 5
Fireworks AI
8.6out of 10
Serverless and reserved inference for open-weight models such as DeepSeek, Kimi, GLM and GPT OSS.
Best for: the lowest per-token price on open models
Fireworks AI answers the cheapest-per-token question: its pricing page says serverless inference starts from $0.05 per 1M input tokens, the rate for NVIDIA Nemotron 3.5 Lightning 30B A3B, with output at $0.20. It hosts many popular open-weight models, including DeepSeek V4.1 Flash at $0.30 in and $1.20 out, GLM 5.3 Flash at $0.15 in and OpenAI GPT OSS 120B at $0.15 in. Each request can pick a mode: Standard, Priority, Fast for 100+ tokens per second, Batch at 50% off, or Spot at 30% off on best-effort capacity. Cached input is billed at half price or less with no write fee. Billing uses prepaid credits with monthly spend limits and Auto Reload, which suits a small team watching cash. Fireworks does not list a free plan, Spot capacity can drop to zero, and US-only serving costs 50% more on select models. Quality depends on the open model you choose, so test a few before committing.
Fireworks AI is the best choice for the lowest per-token price because serverless models start at $0.05 per 1M input tokens. Choose a frontier lab API instead if you need the strongest closed models.
What we like
- Lowest starting price in this comparison at $0.05 per 1M input tokens
- Switch between Standard, Batch and Spot per request without changing endpoints
- Reserved Throughput adds a 99.9% uptime SLA for steady traffic
Watch out for
- No free plan; usage draws down prepaid credits
- Spot throughput can drop to zero, so you need retry or fallback logic
- US-only serverless costs a 50% premium on select models
6 
Ranked 6
Mistral AI
8.6out of 10
Mistral's API for its own Mistral and Ministral models plus selected partner models, with EU-based hosting.
Best for: teams that want a European provider
Mistral AI is the value pick among model makers that train their own models. Its API pricing page lists Ministral 3 3B at $0.10 per 1M tokens for both input and output, Ministral 3 8B at $0.15 for both, Mistral Small 4 at $0.15 in and $0.60 out, and the flagship Mistral Large 4 at $0.70 in and $2.10 out. Codestral serves code completion at $0.30 in and $0.90 out, and the same account covers OCR, Voxtral transcription and speech, and embeddings. Mistral says batch processing cuts prices by 50% and cached input tokens reduce input cost by up to 90%. The model range is narrower at the top end than the big US labs, and the API pricing page does not state a free API tier. For European teams that want a home-region vendor, or anyone who wants output tokens priced as cheaply as input, it is worth a test.
Mistral AI is the best choice for teams that want a European provider because Ministral 3 3B costs $0.10 per 1M tokens for both input and output. Use Mistral Large 4 at $0.70 in for harder tasks.
What we like
- Ministral 3 3B has the same $0.10 price for input and output tokens
- Flagship Mistral Large 4 is $0.70 per 1M input tokens
- Free moderation model (Mistral Moderation 2)
Watch out for
- No free API tier stated on the API pricing page
- Fewer frontier-class models than OpenAI, Google or Anthropic
- Some models (Shieldstral, Timestral) are self-hosted only
7 
Ranked 7
Together AI
8.5out of 10
Serverless, batch and dedicated inference plus fine-tuning for open-weight models.
Best for: teams that may fine-tune open models later
Together AI is a broad open-model host that suits teams expecting to customize later. Its pricing page lists serverless chat models such as MiniMax M3.1 Flash and GLM-5.3-Flash at $0.15 per 1M input tokens, DeepSeek V4.1 Flash at $0.30 in and $1.20 out, and larger models such as Kimi K3 at $3.00 in and $15.00 out. Beyond chat, the same account covers image, video, audio, embeddings, rerank and moderation models. Together says most teams start with serverless inference and move to dedicated endpoints at scale, and it also offers Batch Inference, Provisioned Throughput with SLAs, and managed fine-tuning with SFT and DPO jobs. Prices are a little higher than Fireworks on the cheapest models, no free plan is listed, and the page notes that displayed prices refer to the lowest settings. Pick it when you want one vendor for open-model inference and fine-tuning.
Together AI is the best choice for teams that may fine-tune open models later because inference, managed SFT and DPO fine-tuning and dedicated endpoints sit on one account. Its cheapest chat models cost $0.15 per 1M input tokens.
What we like
- Large catalog of open chat, code, vision and image models
- Path from serverless to dedicated endpoints as traffic grows
- Fine-tuning and inference with the same vendor
Watch out for
- No free plan listed on the pricing page
- Displayed prices refer to the lowest settings and actual prices might vary
- Some listed prices are short-term promotions
8 
Ranked 8
Azure OpenAI
8.1out of 10
OpenAI models served from Microsoft Azure as part of Microsoft Foundry, billed through an Azure subscription.
Best for: teams already on Microsoft Azure
Azure OpenAI is for teams whose company already runs on Microsoft Azure. Microsoft's pricing page describes three ways to buy: Standard on-demand billing for input and output tokens, Provisioned throughput units with monthly and annual reservations, and a Batch API that returns completions within 24 hours at a 50% discount on Global Standard pricing. Deployments can be Global, Data Zone or regional, which helps when data must stay inside a geography. The page lists current OpenAI models such as GPT-6 Sol and GPT-6 Luna in short and long context variants. The trade-off for a small team is setup: you need an Azure subscription and a resource before the first request, while the direct OpenAI API needs only a key. The per-token figures on Microsoft's page are filled in by a script that did not load from our location, so we mark the price pending and suggest checking the Azure pricing calculator for your region.
Azure OpenAI is the best choice for teams already on Microsoft Azure because it bills OpenAI models to an existing Azure subscription and offers Data Zone deployments. Small teams without Azure should use the OpenAI API directly.
What we like
- OpenAI models under Microsoft Azure contracts and billing
- Data Zone deployments for teams with data processing boundaries
- Reservations lower the cost of provisioned throughput
Watch out for
- Per-token prices did not load for us, so verify in the Azure pricing calculator
- Needs an Azure subscription and resource setup before the first call
- Mostly OpenAI models, not a multi-vendor marketplace like Bedrock
9 
Ranked 9
Amazon Bedrock
8.0out of 10
AWS's managed service for foundation models from Amazon and many third-party model makers, billed through AWS.
Best for: teams already on AWS that want many vendors
Amazon Bedrock is the broadest marketplace here. The AWS pricing page lists models from AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Google, Luma AI, Meta, MiniMax AI, Mistral AI, Moonshot AI, NVIDIA, OpenAI, Qwen, Stability AI, TwelveLabs, Writer, xAI and Z AI, plus Amazon's own Nova family. That lets an AWS-based team try several vendors without new contracts, and keep traffic inside existing AWS networking and IAM. Anthropic's docs note that on Bedrock, regional endpoints for Claude 4.5 and later models carry a 10% premium over global endpoints. For a small team not yet on AWS, setup is heavier than a direct API key, and the per-token prices on the AWS page are filled in by a script that did not render for us, so we mark the price pending. Use the AWS pricing page or calculator for the exact rate in your region.
Amazon Bedrock is the best choice for teams already on AWS that want many vendors because one AWS account reaches models from Anthropic, Meta, Mistral AI, OpenAI, xAI and Amazon's own Nova family. Teams without AWS will find direct APIs faster to set up.
What we like
- Widest vendor choice in this comparison under one API and bill
- Fits existing AWS security, networking and procurement
- Batch pricing for jobs that can wait
Watch out for
- Per-token prices did not load for us, so verify on the AWS pricing page
- Regional and multi-region endpoints for newer Claude models cost 10% more than global ones
- AWS account, IAM and model access setup take longer than a direct API key
10 
Ranked 10
Cohere
7.5out of 10
Enterprise AI company offering Command generative models, Embed and Rerank via API and private deployments.
Best for: search and retrieval apps
Cohere is aimed more at enterprises than at indie developers, but its retrieval models make it useful in a small team's stack. Cohere's docs say its generative models, including Command A+, Command A and Command R7B, are priced per token, Rerank is priced per search (one query with up to 100 documents), and Embed is priced per token embedded. Every account gets a free Trial API key, but trial keys are rate limited and not permitted for production or commercial use, and moving to a Production key requires filling out a Go to Production form. Bills are issued at the end of each month or when the outstanding balance reaches $250. The current Command token prices on Cohere's pricing page did not load for us, so we mark the price pending. For a chat-first startup, the direct APIs above are simpler; for search, Rerank is the reason to add Cohere.
Cohere is the best choice for search and retrieval apps because its Rerank model scores up to 100 documents per search unit. Use another provider as your main chat model unless you need Cohere's private deployments.
What we like
- Rerank and Embed models fit retrieval pipelines next to any chat model
- Trial API key is free for testing
- Private and on-premises deployment options
Watch out for
- Trial keys are not permitted for production or commercial use
- Production keys require an application through the Go to Production workflow
- Current Command token prices did not load for us
Before you choose
What to check first
- 1
Price your real token mix, not the headline
Every vendor here bills input and output tokens separately, and output costs more. GPT-6 Luna is $0.10 in and $0.50 out per 1M tokens; Claude Haiku 5.5 is $0.10 in and $0.50 out for prompts up to 100,000 tokens. Estimate a typical request (say 2,000 tokens in, 500 out), multiply by your daily volume, and compare that number across providers.
- 2
Use batch and caching discounts before switching vendors
OpenAI, Anthropic, Google, Mistral, Fireworks and Azure OpenAI all document a 50% discount for batch jobs that can wait up to 24 hours. Cached input is also far cheaper: OpenAI lists GPT-6 Luna cached input at $0.01 per 1M tokens against $0.10 uncached. For nightly classification or bulk summaries, these discounts often matter more than the base rate.
- 3
Check the free tier's data terms
A free tier is useful for prototypes, but read what it costs you in data. Google's pricing page says content on the Gemini free tier is used to improve its products, while paid tier content is not. Cohere trial keys are free but rate limited and not permitted for production or commercial use. Groq's free plan has lower rate limits than its Developer plan.
- 4
Favor OpenAI-compatible endpoints if you may switch
Groq documents an OpenAI-compatible base URL (https://api.groq.com/openai/v1), and Fireworks, Together AI and others serve many of the same open-weight models. Writing your app against the OpenAI SDK format keeps the cost of moving between these hosts low, so you can follow price cuts without a rewrite.
- 5
Buy through your cloud only when procurement requires it
Amazon Bedrock and Azure OpenAI bill through your existing AWS or Azure account, which helps with enterprise contracts, data zones and provisioned throughput. For a small team without those needs, the direct APIs from Google, OpenAI and Anthropic have simpler setup and publish their prices on plain pages.
FAQ
Questions readers ask
- Which LLM API is cheapest per million tokens?
- Fireworks AI has the lowest starting price we found, from $0.05 per 1M input tokens on its serverless open models. Groq follows at $0.075 for GPT OSS 20B, and Google, OpenAI, Anthropic and Mistral each list a model at $0.10 per 1M input tokens, as of November 2026.
- Which LLM API is best for a small startup?
- Google Gemini API is the best fit for most small startups. Its free tier includes free input and output tokens on several models, and paid models start at $0.10 per 1M input tokens with a 50% Batch API discount.
- Which LLM API providers offer a free tier?
- Google Gemini API and Groq offer free tiers you can call from code. Google's free tier covers input and output tokens on models such as Gemini 2.5 Flash, and Cohere gives a free Trial API key that is not allowed for production or commercial use.
- Is the OpenAI API included in ChatGPT Plus?
- No. OpenAI's pricing page says the API is billed separately from ChatGPT Plus, Business, Enterprise and Edu, and Playground usage is billed like regular API usage.
- Which LLM API is best for long documents?
- Anthropic API. Its pricing docs say Claude 4.6 and later models, except Claude Haiku 5.5, include the full 1M token context window at standard pricing, so long prompts carry no surcharge.
- How can I cut my LLM API bill?
- Use batch processing and prompt caching first. OpenAI, Anthropic, Google, Mistral and Fireworks all list a 50% batch discount, and OpenAI charges $0.01 instead of $0.10 per 1M tokens for cached GPT-6 Luna input.
- Should I use Amazon Bedrock or Azure OpenAI instead of a direct API?
- Only if your company already buys through AWS or Azure. Both bill through your cloud account and offer regional or data zone deployments, but setup takes longer than creating a direct API key.
How this ranking was scored
Each product gets a 0 to 10 rating on three criteria, and the order follows the weighted total. Ratings rest on each maker's documented features and published prices, not hands-on testing. Read the method.
- 40%Quality
- 25%Ease of use
- 35%Value
Data behind this ranking
Spotted an error or a new price? Write to editorial@fruitpunch.ai. How we work is in our editorial policy. Makers change prices often, so check the seller page before you buy.
Keep browsing