Software advice
Best LLM API Providers in 2026
An LLM API provider is a company that lets software call a large language model over the internet and pays by usage, usually per million tokens of input and output. Some train their own models, some host other people's open-weight models, and some resell many vendors through a cloud account. It is not a chatbot subscription for people.
Quick answer
OpenAI's API ranks first for breadth of endpoints and a clear price ladder. Anthropic fits coding agents and long context, and Google's Gemini API fits prototypes thanks to its free tier. Mistral is the pick for buyers who want a vendor based in the EU.
A developer choosing an LLM API is not choosing a chatbot. The decision is which vendor your code will call thousands of times a day, what each million tokens will cost, and where your prompts go after the response comes back. This ranking covers ten providers across three types: model makers, open-model hosts and cloud marketplaces. It judges breadth of models, published prices, data handling terms from each vendor's own policy, how hard it is to leave, and which team each one suits. Prices were read from vendor pages in October 2026 and change often. We did not benchmark model quality, so quality is described as a design trade-off, not a score.
The 10 best LLM API providers in 2026, ranked
The order is the verdict, there is no score. No vendor paid for position. How we rank.
| # | Tool | Best for | Free plan | Entry price |
|---|---|---|---|---|
| 1 | OpenAI API | Teams that want the broadest set of text, voice, image and agent endpoints from one vendor. | No | $0.10 per million input tokens (gpt-6-luna) |
| 2 | Anthropic API | Developers building coding agents and long-running automations who want a small, clearly tiered model family. | No | $2 per million input tokens (Sonnet 5.5) |
| 3 | Google Gemini API | Builders who want a free tier to prototype and one key for text, images, audio and long-context work. | Yes | Free tier on most models, paid from $0.10 per million input tokens (Gemini 2.5 Flash-Lite) |
| 4 | Mistral AI | European teams that want a French vendor, EU-style contracts and strong open-weight options. | Yes | $0.50 per million input tokens (Mistral Large) |
| 5 | Cohere | Enterprises building retrieval, search and document workflows, where embedding and rerank models matter as much as chat. | Trial | Quoted |
| 6 | Groq | Latency-sensitive apps such as voice agents and live assistants that run open-weight models. | Yes | $0.075 per million input tokens (GPT OSS 20B) |
| 7 | Together AI | Teams that want many open-weight models and the option to fine-tune or reserve GPUs later. | No | $0.15 per million input tokens (gpt-oss-120B) |
| 8 | Fireworks AI | Production apps that need open models served fast, with fine-tuning and dedicated GPU options. | No | Quoted |
| 9 | Amazon Bedrock | Companies already on AWS that want many vendors' models behind existing IAM, billing and compliance. | No | Quoted |
| 10 | DeepSeek | Cost-driven teams that accept a China-based API or run the open-weight models themselves. | No | $0.15 per million input tokens (deepseek-flash, off-peak) |
- 1
OpenAI API
OpenAI's API takes first place because one account covers text models, realtime voice, transcription, image generation, hosted tools and an agents layer, all priced per token or per call. The three-tier line-up (luna, sol, astra) makes it easy to start cheap and move up. The cost is a pricing sheet with many modes, and the fine-tuning platform is closed to new users.
Strong on
- Documented endpoints for text, realtime voice, transcription and image generation under one billing account
- Batch and Flex processing both list half of the standard token price
- Data sent to the API is not used for training unless you opt in, per the vendor's data controls guide
Weak on
- Fast and Ultrafast modes multiply the standard price, so a careless setting inflates the bill
- Fine-tuning is closed to new users, so custom models mean prompting or another vendor
Skip it if: Buyers who need EU-only processing as the default, since regional endpoints add a 10% uplift on newer models. Smallest model, standard speed. Batch is half price. Flagship gpt-6-astra is $10 input, $50 output.
- 2
Anthropic API
Anthropic's API ranks second because its model family is short and easy to reason about: Haiku for volume, Sonnet for general work, Opus and Fable for hard problems. Prompt caching is priced explicitly, with cache hits at a tenth of the base input rate. The same models are sold through Bedrock and Google Cloud, which helps with procurement.
Strong on
- Prompt cache hits are billed at 0.1 times the base input price, which suits agents that resend long context
- Claude models are also sold on Amazon Bedrock and Google Cloud, so existing cloud commitments can pay for them
- Inputs and outputs are deleted from Anthropic's backend within 30 days by default, per the privacy center
Weak on
- The vendor warns newer tokenizers produce roughly 30% more tokens for the same text, which raises effective cost
- US inference geography adds a 1.1 times multiplier when you pin it
Skip it if: Teams that want image, speech or embedding models from the same vendor account. Sonnet 5.5 standard rate. Opus 5.5 is $4 input and $20 output per million.
- 3
Google Gemini API
Google's Gemini API sits third because it has the only generous free tier among the large labs and a wide price ladder, from 10 cents to 4 dollars per million input tokens. Batch halves the paid rate. The trade-off is that free-tier prompts may be read by human reviewers and used to improve Google products, so production traffic belongs on the paid tier.
Strong on
- Free tier on most text models, with Gemma 4 listed as free with no paid tier
- Paid price ladder runs from Gemini 2.5 Flash-Lite at $0.10 input to Gemini 4 Argon at $2.00 input
- Batch pricing is half of the standard paid rate where it is listed
Weak on
- Free-tier content may be used to improve Google products and read by reviewers
- Several list prices are scheduled to rise within months of the page date
Skip it if: Anyone who must keep prompts out of product improvement and therefore cannot use the free tier at all. Paid rate per million tokens. Several prices rise on December 1, 2026 or January 1, 2027.
- 4
Mistral AI
Mistral ranks fourth because it is the one large-scale model maker headquartered in the EU, with commercial terms that say customer data and outputs are not used to train its models unless you opt in. Its flagship is priced low, at $0.50 and $1.50 per million tokens. The pricing page publishes few model rates, so smaller models need a check in the console.
Strong on
- Registered in Paris, with commercial terms that exclude model training on customer data absent an opt-in
- Mistral Large lists $0.50 input and $1.50 output per million tokens
- Batch discount of 50% and up to 90% off cached input are stated on the pricing page
Weak on
- The public pricing page lists a single API model price, so other models need the console
- Studio free plan is real, but API free limits were not published on the page we read
Skip it if: Buyers who want one vendor for images, video, music and the widest consumer tooling around the model. Output is $1.50 per million. Batch is 50% off, and cached input saves up to 90%.
- 5
Cohere
Cohere ranks fifth because it sells the retrieval pieces (Embed, Rerank, Parse) alongside Command models, and offers private deployment in your own cloud or on-premises. That is rare in this list. The weak spots are a pricing page that hides current per-token rates behind tabs and a data policy that says customers can opt out of training, which implies a default in favor of it.
Strong on
- Embed, Rerank and Parse models sit beside the Command generation family in one account
- Model Vault runs dedicated instances from $3.00 per hour or $2,000 per month
- Private deployment in a VPC or on-premises is offered, and Cohere says it has no access to that environment
Weak on
- Current Command per-token prices were not shown in the pricing page text, only legacy rates
- Cohere's security page says customers can opt out of training, so the default needs confirming in the contract
Skip it if: Hobbyists and small apps that want a public per-token price list and a free tier for production. Trial keys are free but limited and barred from production. Current Command rates were not on the page.
- 6
Groq
Groq earns sixth place on speed and price rather than model choice. Its docs list GPT OSS 120B at about 500 tokens per second and 20B at about 1,000, at $0.15 and $0.075 per million input tokens. The catalog is small and mostly open-weight, and several models are enterprise-only. Retention defaults are good, but stored data sits in US cloud buckets.
Strong on
- Docs list GPT OSS 120B near 500 tokens per second at $0.15 input and $0.60 output
- By default Groq does not retain customer data for inference requests, and any customer can enable zero data retention
- Whisper Large V3 Turbo transcription is listed at $0.04 per hour of audio
Weak on
- Short model catalog, with Llama 3.3 70B and others listed as contact-sales only
- Retained data, such as batch jobs, is stored in US cloud buckets
Skip it if: Teams that need a closed frontier model or one vendor covering images and video. GPT OSS 120B is $0.15 input, $0.60 output. Free and developer tiers exist.
- 7
Together AI
Together AI lands seventh because its strength is breadth: dozens of open models across chat, vision, image, audio and video, with fine-tuning and dedicated GPUs on the same account. A buyer can start serverless and move to reserved capacity as volume grows, without changing vendors. The pricing page is long and changes often, so a quote should be re-read before each budget cycle.
Strong on
- Serverless table spans models from gpt-oss-120B at $0.15 input to Kimi K3 at $3.00 input
- Dedicated endpoints on H100 hardware list $5.49 per GPU hour, with reserved rates by contract
- Privacy notice says data is not used to train models without explicit opt-in
Weak on
- Dozens of text models make the choice slow without a test set
- No free tier was shown on the pricing page
Skip it if: Buyers who specifically need GPT, Claude or Gemini models under one contract. Serverless per token. Dedicated H100 endpoints start at $5.49 per GPU hour.
- 8
Fireworks AI
Fireworks AI takes eighth place as the privacy-friendliest of the open-model hosts: its docs say prompts and generations are not logged or stored for open models by default. It sells serverless tokens, managed fine-tuning and per-second GPU deployments. The public pricing page does not list per-token model rates, which pushes buyers to the docs, and its catalog is open models only.
Strong on
- Zero data retention is the default for open models, per the vendor's docs
- Per-second GPU billing on H100 and H200 at $8.00 per hour, with no extra charge for start-up
- Managed fine-tuning prices start at $0.50 per million training tokens for models up to 16B
Weak on
- The Response API keeps conversations for 30 days when store is true, which is the default
- Per-token serverless prices for popular models are on a separate documentation page
Skip it if: Teams that want a single bill that also includes closed models from OpenAI or Anthropic. Prepaid credits, billed per token. On-demand H100 and H200 GPUs are $8.00 per hour.
- 9
Amazon Bedrock
Amazon Bedrock ranks ninth for a reason that is not about the models: its prices vary by provider, region and service tier, so a quote needs a worksheet. What it offers is procurement simplicity for AWS customers, with models from Anthropic, OpenAI, Meta, Mistral and others, and the statement that inputs and outputs are not shared with model providers or used to train them.
Strong on
- Offers models from Anthropic, OpenAI, Mistral, Cohere, Meta, DeepSeek and others through one API
- FAQ states inputs and outputs are not used to train models and are not shared with model providers
- SOC 1, 2 and 3, ISO 27001 and HIPAA eligibility are listed in the vendor FAQ
Weak on
- No single price list, since rates differ by model, region and tier
- Cross-region inference needs a closer look if residency matters, since the FAQ does not spell out its effect
Skip it if: Startups outside AWS that would adopt a cloud account only to reach one model. On-demand per token, varying by model and region. Flex is 50% below Standard, Priority 75% above.
- 10
DeepSeek
DeepSeek is last on this list because its hosted API trades data comfort for price. Off-peak, deepseek-flash costs $0.15 input and $0.60 output per million tokens and offers a 1M context window, and the API accepts both OpenAI and Anthropic request formats. But the data controller is a Hangzhou company with its registered address in China, and its policy lists model training among the uses of personal data.
Strong on
- Off-peak input at $0.15 per million tokens for deepseek-flash, with 1M tokens of context
- Endpoints follow OpenAI and Anthropic request formats, which lowers migration effort
- Peak and off-peak pricing is published, with off-peak at half the peak rate
Weak on
- Data controller has a registered address in China and the policy lists training as a use of data
- Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays and double the price
Skip it if: Anyone sending personal, client or regulated data to a hosted API under EU or US data rules. Cache misses off-peak. Peak hours double the rate, and deepseek-v4-pro is $0.66 input and $1.98 output off-peak.
How these LLM API providers were judged
How do you judge model quality before you commit to an API?
Judge quality on your own prompts, because vendor leaderboards rarely match a specific workload. Each provider here publishes a model family that spans cheap and expensive tiers: OpenAI's luna, sol and astra, Anthropic's Haiku, Sonnet, Opus and Fable, and Google's Flash-Lite through Gemini 4 Argon. Open-model hosts such as Groq, Together AI and Fireworks AI serve models they did not train, so quality tracks the underlying weights. Build a test set of fifty real inputs, run each shortlisted API on it, and score the outputs yourself. Rerun it whenever a vendor retires a model name, which several did in 2026.
What does an LLM API actually cost per million tokens?
Entry prices span a wide range, as of October 2026. OpenAI's smallest listed model is $0.10 per million input tokens, Gemini 2.5 Flash-Lite is also $0.10, Groq's GPT OSS 20B is $0.075, and DeepSeek's flash model is $0.15 off-peak. Flagship tiers cost far more: OpenAI's gpt-6-astra is $10 input and $50 output, and Anthropic's Fable 5.1 is the same. Batch modes at OpenAI, Google and Mistral cut the rate by half. Output tokens cost several times input, so long answers dominate bills. Bedrock and Cohere need a quote worksheet, since rates vary by model and region.
Do LLM API providers train on your prompts and keep your data?
Terms differ and the defaults matter. OpenAI says API data is not used for training unless you opt in and keeps abuse logs for up to 30 days. Anthropic deletes API inputs and outputs within 30 days by default. Google's free Gemini tier may use content to improve products and let reviewers read it, while paid use does not. Mistral's commercial terms exclude training without opt-in, and Fireworks logs nothing for open models by default. Groq retains nothing by default. DeepSeek's policy lists training among the uses of personal data. Cohere's security page says customers can opt out of training. Read the policy for your tier.
How hard is it to switch LLM API providers later?
Switching costs sit in prompts and glue code more than in contracts. Most providers here accept OpenAI-style chat calls, and DeepSeek also accepts the Anthropic format, so a base URL swap gets you running quickly. What breaks is behavior: a prompt tuned for one model drifts on another, tool-calling formats differ, and cached prefixes and batch jobs do not move. Fine-tuned models are the hardest to take with you, and OpenAI has closed its fine-tuning platform to new users. Keep a thin wrapper around the call, store your evaluation set, and pay for a second vendor's tests before volume grows.
Which LLM API provider fits a startup, an enterprise or an EU company?
A startup prototyping on a budget should start with Google's free tier or OpenAI's small models, then compare against open-model hosts. A team building coding agents will look hard at Anthropic's tiering and caching. An enterprise on AWS may find Bedrock the easiest to approve, and one doing document search should look at Cohere. A European company that wants a vendor in its own jurisdiction should start with Mistral, which is registered in Paris, and ask every other vendor about regional endpoints and a data processing agreement.
How to choose a LLM API provider in three steps
- Write down the workload before the vendor list. Estimate monthly input and output tokens, how long a typical prompt is, whether answers must stream in under a second, and which data types will pass through. A voice agent needs low latency, which points to Groq or Fireworks AI. A nightly summarization job can wait for a batch endpoint at half price. A retrieval system needs embeddings and rerank as well as generation, which points to Cohere. Without this sheet, every vendor looks cheap on its own pricing page.
- Run the same fifty prompts through three finalists. Pick one first-party vendor, one open-model host and one cloud marketplace, then run the same fifty real prompts through each at the tier you would buy. Record cost per answer, time to first token and how often a human reviewer accepts the output. Free tiers make this cheap: Gemini has one, Cohere gives free trial keys with rate limits, and Groq lists a free tier. Do not use a free tier for sensitive text until you have read its data terms.
- Read the data terms and plan the exit. Open each finalist's privacy page and check three things: training on your data, retention length, and where data is stored. Ask for a data processing agreement if you handle personal data. Then plan the exit: keep prompts in version control, avoid vendor-specific features in core paths, and keep a second provider's API key warm. If a vendor retires a model, as DeepSeek's retired v4-flash name and Google's deprecated image model show, a tested fallback turns a crisis into a config change, and a stored evaluation set shows within an hour whether the replacement is good enough.
Frequently asked questions
What is the best LLM API provider in 2026?
OpenAI's API ranks first here because one account covers text, voice, transcription, image generation and agent tools, with a three-tier model line-up and batch pricing at half the standard rate. Anthropic ranks second for coding agents, and Google's Gemini API third for its free tier. The best choice still depends on your workload, so test three finalists on your own prompts.
Which LLM API is cheapest per million tokens?
Among listed standard prices as of October 2026, Groq's GPT OSS 20B is lowest at $0.075 input, then OpenAI's gpt-6-luna and Gemini 2.5 Flash-Lite at $0.10, and DeepSeek's flash model at $0.15 off-peak. Cheapest on paper is not cheapest in practice, since output tokens, retries and quality differences change the cost per useful answer.
Do OpenAI and Anthropic train on API data?
OpenAI says data sent to its API has not been used to train its models since March 1, 2023, unless you opt in, and keeps abuse monitoring logs for up to 30 days. Anthropic says it deletes API inputs and outputs within 30 days unless you agree otherwise, a usage policy review applies, or the law requires retention. Zero data retention is offered by agreement.
Is the Gemini API free for production use?
The free tier exists, but it is a poor fit for production. Google's terms say free-tier content may be used to improve its products, and human reviewers may read it after it is detached from your account. Paid services state that prompts and responses are not used for product improvement, though Google logs them for a limited time to detect abuse.
Which LLM API provider is best for EU data requirements?
Mistral AI is the only vendor here headquartered in the EU, with registered offices in Paris and commercial terms that exclude training on customer data without opt-in. OpenAI offers regional processing at a 10% uplift on newer models, and Bedrock stores content in the region you choose. Check cross-region routing and request a data processing agreement from any vendor.
Is DeepSeek's API safe to use for business data?
That depends on your data. DeepSeek's privacy policy names Hangzhou DeepSeek Artificial Intelligence as controller, with its registered address in China, and lists training and improving its technology among the uses of personal data. For public or non-sensitive content the low price is attractive. For client or regulated data, many teams run the open weights themselves or use a host.
What is the difference between Together AI, Fireworks AI and Groq?
All three host open-weight models for a per-token fee, but they differ in emphasis. Groq stresses speed on a short catalog, Together AI offers the widest model list plus fine-tuning and reserved GPUs, and Fireworks AI pairs serverless tokens with per-second GPU deployments and zero data retention by default for open models. Pick by latency, catalog and data terms.
Should I use Amazon Bedrock or call the model vendor directly?
Use Bedrock when your company already buys from AWS and wants one bill, IAM controls and compliance coverage across several model vendors. Go direct when you want the newest features first or a simple price list. Anthropic lists Bedrock pricing separately and notes regional endpoints carry a 10% premium over global ones on newer Claude models.