Best software · Developer & data
10 Best LLM APIs of 2026
This list is for developers and engineering leads picking a hosted model API for a product, an agent or an internal tool. It covers first-party model APIs (OpenAI, Google Gemini, Anthropic, Mistral AI, DeepSeek), inference platforms that serve open-weight models (Groq, Fireworks AI, Together AI) and the two big cloud marketplaces (Amazon Bedrock, Azure OpenAI). We compared documented features, context windows and published per-token prices for 10 APIs. Prices are per 1 million input tokens unless noted, and output tokens cost more on every service. Scores weigh features at 40%, ease of use at 30% and value at 30%.
By the Mojolin research desk · Updated November 9, 2026 · 10 tools compared · Prices checked on vendor pages, latest November 9, 2026
Questions or corrections: editorial@mojolin.com
Mojolin is self-funded How we score
OpenAI API is our top pick: its four GPT-6 family models run from $0.10 to $10 per 1M input tokens, every one has a 1.05M-token context window, and functions, web search, file search and computer use are built in. Google Gemini API is the runner-up and the best place to start for free, with a free tier and Gemini 3.1 Flash-Lite at $0.25 per 1M input tokens. Anthropic Claude API is the pick for long agent runs, with a 1M-token window on all four current models and Claude Haiku 5.5 from $0.10. For the lowest bills on open-weight models, Groq, DeepSeek, Fireworks AI and Together AI start between $0.05 and $0.15. Teams with AWS or Azure commitments should look at Bedrock or Azure OpenAI first.
Top picks
Best overall
Google Gemini API
9.1 / 10
Developers who want to prototype for free and run high-volume work on low-cost Flash and Flash-Lite models.
Read the review →Runner-up
OpenAI API
9.1 / 10
Developers who want a full range of models from cheap to frontier, a 1.05M-token context window and built-in tools behind one API.
Read the review →Also great
Anthropic Claude API
9.0 / 10
Teams building coding agents, long-running agent workflows and long-document analysis that need a 1M-token context window on every current model.
Read the review →Comparison table
| # | Tool | Pricing | Score | |
|---|---|---|---|---|
| 1 | Google Gemini API | Developers who want to prototype for free and run high-volume work on low-cost Flash and Flash-Lite models. | Free tier; paid from $0.25 per 1M input tokens and $1.50 per 1M output tokens (Gemini 3.1 Flash-Lite) | 9.1 |
| 2 | OpenAI API | Developers who want a full range of models from cheap to frontier, a 1.05M-token context window and built-in tools behind one API. | Pay as you go; from $0.10 per 1M input tokens and $0.50 per 1M output tokens (GPT-6 Luna) | 9.1 |
| 3 | Anthropic Claude API | Teams building coding agents, long-running agent workflows and long-document analysis that need a 1M-token context window on every current model. | Pay as you go; from $0.10 per 1M input tokens and $0.50 per 1M output tokens (Claude Haiku 5.5, prompts up to 100,000 tokens) | 9.0 |
| 4 |  DeepSeek API | Cost-sensitive teams that want 1M-token context and long outputs at very low prices and can schedule work off-peak. | Pay as you go from a prepaid balance; deepseek-flash from $0.15 per 1M input tokens off-peak and $0.30 at peak | 8.7 |
| 5 |  Groq | Developers building real-time apps and agents that need very fast responses from open-weight models at low cost. | Free tier; Developer tier pay as you go from $0.075 per 1M input tokens (GPT OSS 20B) | 8.6 |
| 6 |  Mistral AI | Teams that want low-cost European models, OCR and transcription from one vendor, with regional inference options. | Pay as you go; from $0.10 per 1M input and output tokens (Ministral 3 3B); Mistral Large 4 at $0.70 in, $2.10 out | 8.6 |
| 7 |  Fireworks AI | Teams running open-weight models in production that want per-request speed modes, discounts for batch and spot capacity, and fine-tuning. | Prepaid credits; serverless from $0.05 per 1M input tokens | 8.5 |
| 8 |  Together AI | Teams that want a large catalog of open-weight chat, image, video and audio models plus a path to dedicated GPUs. | Pay as you go; chat models from $0.15 per 1M input tokens (GLM-5.3-Flash, GPT OSS 120B) | 8.4 |
| 9 |  Amazon Bedrock | Companies already on AWS that want many model providers under AWS billing, IAM and regional controls. | See vendor pricing (USD check pending) | 8.1 |
| 10 |  Azure OpenAI | Organizations on Microsoft Azure that want OpenAI models with Azure billing, data zone deployments and enterprise controls. | See vendor pricing (USD check pending) | 7.9 |
Scoring weights
Every tool gets three editorial scores out of 10, from vendor documentation, product pages and the published price. The overall score is the weighted average below, the same method on every list. Full method
The full ranking
Rank 1 · Best overall
Google Gemini API
Google's developer API for Gemini models, with a free tier in Google AI Studio and paid pay-as-you-go access.
ai.google.dev9.1
out of 10
Best for: Developers who want to prototype for free and run high-volume work on low-cost Flash and Flash-Lite models.
The Gemini API is the easiest model API to start on without paying. The free tier includes free input and output tokens on selected models plus Google AI Studio, though Google uses free-tier content to improve its products. Paid accounts get higher rate limits, context caching, a Batch API at half price and content kept out of training.
The lineup runs from Gemini 4 Argon, a reasoning model with a shared 1M-token context window, down to Flash and Flash-Lite models built for volume. Gemini 3.1 Flash-Lite costs $0.25 in and $1.50 out per 1M tokens, and Gemini 3.8 Flash is $0.75 in and $3.75 out through the end of 2026. Several prices are promotional: Argon goes from $2 to $4 per 1M input tokens on December 1, 2026, and 3.8 Flash doubles on January 1, 2027. Grounding with Google Search gives 5,000 free requests a month on Gemini 3 and newer, then $14 per 1,000. The older Gemini 2.5 Flash-Lite is cheaper but limited to existing users.
Standout feature
A real free tier plus grounding in Google Search makes it the cheapest place to build and test a model-backed feature.
Use cases
Indie developers
Prototype a chat feature on the free tier in Google AI Studio
Outcome: A working demo before any billing is set up
Content platforms
Translate and tag large volumes of text with Gemini 3.1 Flash-Lite
Outcome: Input billed at $0.25 per 1M tokens
Research assistants
Answer questions grounded in Google Search results
Outcome: Cited, current answers with 5,000 free grounded requests a month
Key features
- Free tier with free input and output tokens on selected models and Google AI Studio access
- Gemini 4 Argon flagship reasoning model with a 1M-token context window
- Flash and Flash-Lite models for high-volume agentic tasks, translation and data processing
- Batch API at a 50% cost reduction and context caching on the paid tier
- Grounding with Google Search and Google Maps, 5,000 free requests per month on Gemini 3 and newer
Pros
- Free tier lets you build and test before adding a card
- Gemini 3.1 Flash-Lite costs $0.25 in and $1.50 out per 1M tokens
- Native grounding in Google Search results
Cons
- Free tier content is used to improve Google's products
- Gemini 4 Argon input doubles from $2 to $4 on December 1, 2026
- Gemini 4 Argon is not available on the free tier
Rank 2 · Runner-up
OpenAI API
OpenAI's developer platform for the GPT-6 model family, with built-in tools, agents, realtime voice, embeddings and fine-tuning.
openai.com9.1
out of 10
Best for: Developers who want a full range of models from cheap to frontier, a 1.05M-token context window and built-in tools behind one API.
OpenAI's API covers the widest spread of price and capability on this list. The four flagship models step down from GPT-6 Astra at $10 in and $50 out per 1M tokens to GPT-6.1 Sol at $2 and $10, GPT-6.1 Terra at $1 and $5, and GPT-6 Luna at $0.10 and $0.50. All four list a 1.05M-token context window, 128K max output, and built-in functions, web search, file search and computer use, so you can move a request from Luna to Astra without changing your tool setup.
Cost controls are well documented. Batch and Flex processing halve the Standard price, cached input is billed at roughly a tenth of the input rate, and you can choose 30-minute or 12-hour cache writes. For latency-sensitive work, Fast and Ultrafast tiers trade higher prices for speed. Around the models sit the Agents API, Realtime voice, embeddings, moderation and fine-tuning. There is no free tier, and the top model's output price adds up quickly on long generations.
Standout feature
Every flagship model, from the $0.10 Luna to the $10 Astra, shares a 1.05M-token context window and the same built-in tools.
Use cases
SaaS product teams
Add an assistant that searches the web and the customer's files
Outcome: Built-in web and file search without a separate retrieval stack
Data teams
Classify millions of support tickets overnight with GPT-6 Luna on Batch
Outcome: Input billed at $0.05 per 1M tokens
Agent builders
Run long computer-use tasks with GPT-6 Astra
Outcome: A 1.05M-token window keeps the whole session in context
Key features
- Four GPT-6 family models from GPT-6 Luna at $0.10 to GPT-6 Astra at $10 per 1M input tokens
- 1.05M-token context window and 128K max output on every flagship model
- Built-in tools for functions, web search, file search and computer use
- Batch and Flex processing at half the Standard price, plus Fast and Ultrafast tiers
- Prompt caching with 30-minute and 12-hour cache writes
Pros
- Cheapest model starts at $0.10 per 1M input tokens with cached input at $0.01
- Same long context window across cheap and frontier models
- Agents API, Realtime API, embeddings and fine-tuning on one platform
Cons
- No free API tier listed on the pricing page
- GPT-6 Astra output costs $50 per 1M tokens
- Fast and Ultrafast tiers cost two to six times the Standard rate
Rank 3 · Also great
Anthropic Claude API
Anthropic's API for Claude models, from Claude Haiku 5.5 for routing and extraction to Claude Fable 5.1 for long-horizon agent work.
anthropic.com9.0
out of 10
Best for: Teams building coding agents, long-running agent workflows and long-document analysis that need a 1M-token context window on every current model.
Anthropic's API is built around four current Claude models that all share a 1M-token context window and 128K max output. Claude Fable 5.1 costs $10 in and $50 out per 1M tokens and targets demanding reasoning and long-horizon agentic work, Opus 5.5 is $4 and $20 for long coding and knowledge work, Sonnet 5.5 is $2 and $10, and Haiku 5.5 is $0.10 and $0.50 for classification, extraction and routing on prompts up to 100,000 tokens.
Cost levers are documented in detail. Prompt caching offers 5-minute and 1-hour writes with cheap cache hits, the Batch API takes 50% off input and output, and US-only inference costs 1.1 times the standard rate. The same models are available through Amazon Bedrock, Google Cloud and Microsoft Foundry for teams that want cloud billing. Two pricing details deserve attention: Haiku's input price rises fivefold above 100,000 tokens, and Claude 4.7 and later models use a tokenizer that produces about 30% more tokens for the same text.
Standout feature
All four current Claude models, including the $0.10 Haiku 5.5, support a 1M-token context window.
Use cases
Developer tool companies
Run a coding agent over a large repository with Claude Opus 5.5
Outcome: Long sessions fit in a 1M-token window
Ops teams
Route and extract fields from inbound email with Claude Haiku 5.5
Outcome: Input billed at $0.10 per 1M tokens on short prompts
Legal and research teams
Analyze long contracts with a cached system prompt
Outcome: Repeated context billed at the cache-hit rate
Key features
- Four current models: Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5
- 1M-token context window and 128K max output on all four current models
- Prompt caching with 5-minute and 1-hour writes and low-cost cache hits
- Batch API with a 50% discount on input and output tokens
- Also sold through Amazon Bedrock, Google Cloud and Microsoft Foundry
Pros
- Claude Haiku 5.5 starts at $0.10 per 1M input tokens
- Every current model has a 1M-token window
- Clear published prices for caching, batch and data residency
Cons
- No free API tier
- Claude Haiku 5.5 input rises to $0.50 per 1M tokens for prompts over 100,000 tokens
- Newer models use a tokenizer that produces about 30% more tokens for the same text
Rank 4
DeepSeek API
DeepSeek's API for its Flash and Pro models, compatible with both the OpenAI and Anthropic API formats.
deepseek.com8.7
out of 10
Best for: Cost-sensitive teams that want 1M-token context and long outputs at very low prices and can schedule work off-peak.
DeepSeek offers the lowest prices of any first-party model maker on this list, with two models and a simple table. deepseek-flash, now served by DeepSeek-V4.1-Flash-1017, costs $0.15 per 1M input tokens and $0.60 per 1M output tokens off-peak, and deepseek-v4-pro costs $0.66 and $1.98. Both list a 1M-token context length and up to 384K output tokens, which is far more output room than most competitors.
The API speaks both the OpenAI and Anthropic formats, so coding agents and SDKs built for those vendors can point at DeepSeek with a base URL change. JSON output, tool calls, the Responses API and chat prefix completion are supported, and vision works on Flash. The catch is time-of-day pricing: rates double during peak hours on weekdays (01:00 to 04:00 and 06:00 to 10:00 UTC). Usage is deducted from a topped-up balance, and concurrency limits are 2,500 for Flash and 500 for Pro.
Standout feature
Up to 384K output tokens per request at off-peak prices from $0.15 per 1M input tokens.
Use cases
Startups
Swap a coding agent's backend to DeepSeek through the Anthropic-compatible endpoint
Outcome: Lower token bills without rewriting the agent
Data teams
Generate long structured reports overnight
Outcome: Off-peak rates and up to 384K output tokens per request
Search products
Pair the model with DeepSeek's Search and Fetch APIs
Outcome: Web results at $0.75 per 1,000 requests
Key features
- Two models: deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813)
- 1M-token context length and up to 384K output tokens
- OpenAI and Anthropic compatible endpoints, so existing SDKs work with a new base URL
- JSON output, tool calls, Responses API and chat prefix completion
- Search and Fetch APIs at $0.75 per 1,000 successful requests
Pros
- Off-peak rates are half the peak rates
- Cache hits cost $0.003 per 1M input tokens off-peak on deepseek-flash
- Drop-in compatibility with OpenAI and Anthropic SDKs
Cons
- Peak hours on weekdays double the price
- Only two current models
- Vision is supported on deepseek-flash only
Rank 5
Groq
GroqCloud inference platform that serves open-weight models on Groq's LPU hardware at high token speeds.
groq.com8.6
out of 10
Best for: Developers building real-time apps and agents that need very fast responses from open-weight models at low cost.
Groq's pitch is speed. Its models page lists GPT OSS 20B at about 1,000 tokens per second and GPT OSS 120B at about 500, served on Groq's own LPU hardware. Prices are low: GPT OSS 20B costs $0.075 in and $0.30 out per 1M tokens, and GPT OSS 120B costs $0.15 and $0.60. The GPT OSS models come with built-in browser search and code execution, and the API is OpenAI compatible.
You can start on the free tier and move to the Developer tier by adding a payment method. Developer adds higher rate limits, chat support, a Flex service tier, batch processing and spend limits, and new users are billed at progressive thresholds of $1, $10, $100, $500 and $1,000 before monthly invoicing. The trade-off is range. The self-serve production catalog is short, context windows are 131,072 tokens, and several larger models, including Llama 3.3 70B, are marked Contact Sales.
Standout feature
GPT OSS 20B runs at a listed 1,000 tokens per second for $0.075 per 1M input tokens.
Use cases
Voice app developers
Generate replies fast enough for spoken conversation
Outcome: Short response times from high token speeds
Agent builders
Run GPT OSS 120B with built-in browser search
Outcome: A searching agent without a separate search API
Students and hobbyists
Test open models on the free tier
Outcome: Experiments at no cost before upgrading
Key features
- Listed speeds of about 1,000 tokens per second on GPT OSS 20B and 500 on GPT OSS 120B
- OpenAI-compatible API with a Responses API
- Built-in browser search and code execution on GPT OSS models
- Flex service tier, batch processing and spend limits on the Developer tier
- Speech to text with Whisper Large V3 Turbo at $0.04 per hour
Pros
- GPT OSS 20B costs $0.075 per 1M input tokens
- Free tier, with the option to downgrade back from Developer at any time
- Progressive billing thresholds from $1 limit surprise invoices
Cons
- Production models list a 131,072-token context window
- Small catalog of self-serve production models
- Several models such as Llama 3.3 70B are listed as Contact Sales
Rank 6
Mistral AI
Mistral AI's API for its own generalist, code, OCR and audio models plus selected third-party open-weight models.
mistral.ai8.6
out of 10
Best for: Teams that want low-cost European models, OCR and transcription from one vendor, with regional inference options.
Mistral's API is a good fit for teams that want one vendor for chat, code, documents and speech. The generalist line covers Mistral Large 4 at $0.70 in and $2.10 out per 1M tokens, Mistral Small 4 at $0.15 and $0.60, and Ministral 3 models from 3B at $0.10 in and out to 14B at $0.20. Mistral also serves third-party open-weight models, including Z.ai GLM 5.3 Flash and DeepSeek V4.1 Flash, each with a 1M-token context window.
Beyond chat, OCR 4.1 costs $4 per 1,000 pages, Voxtral Mini Transcribe 3 is $0.0035 per minute, Codestral handles code, and Mistral Moderation 2 is free. Pricing pages offer Standard, Batch and Priority tiers, regional inference and prices in USD or EUR, which helps European buyers. The pricing page does not list context windows for Mistral's own models, and there is no free API tier shown there, so check model cards before you plan long-document work.
Standout feature
Mistral Large 4 costs $0.70 per 1M input tokens, a low price for a vendor's flagship model.
Use cases
European SaaS teams
Run chat features with regional inference
Outcome: Model calls processed in the chosen region
Back-office teams
Extract text from scanned invoices with OCR 4.1
Outcome: Structured output at $4 per 1,000 pages
Media companies
Transcribe interviews with Voxtral
Outcome: Transcripts from $0.0035 per minute
Key features
- Mistral Large 4, Mistral Small 4 and Ministral 3 models in 3B, 8B and 14B sizes
- Third-party models such as Z.ai GLM 5.3 and DeepSeek V4.1 Flash with 1M-token context windows
- Codestral for code at $0.30 in and $0.90 out per 1M tokens
- OCR 4.1 at $4 per 1,000 pages and Voxtral transcription from $0.0035 per minute
- Standard, Batch and Priority tiers with regional inference
Pros
- Mistral Large 4 costs $0.70 in and $2.10 out per 1M tokens
- Prices published in both USD and EUR
- OCR, speech and moderation models alongside chat models
Cons
- No free API tier shown on the pricing page
- Context windows for Mistral's own models are not on the pricing page
- Fewer built-in tools than OpenAI or Gemini
Rank 7
Fireworks AI
Inference and training platform for open-weight models, with serverless, reserved and in-your-cloud deployment options.
fireworks.ai8.5
out of 10
Best for: Teams running open-weight models in production that want per-request speed modes, discounts for batch and spot capacity, and fine-tuning.
Fireworks AI focuses on serving open-weight models fast and cheaply, and its pricing page shows more ways to cut the bill than most. Serverless prices start at $0.05 per 1M input tokens for NVIDIA Nemotron 3.5 Lightning 30B A3B, GPT OSS 120B costs $0.15 in and $0.60 out, and larger models such as Kimi K3 are $3 and $15. Each request can pick Standard, Priority, Fast, Batch or Spot mode on the same endpoint. Batch takes 50% off, Spot takes 30% off on select models, and cached input is billed at half price or less with no write fee.
For steady traffic, Reserved Throughput adds a throughput SLA and 99.9% uptime terms, and Fireworks Sovereign runs managed inference on your own GPUs in private preview. Managed training starts at $0.50 per 1M training tokens. Usage draws down prepaid credits with spend limits. The catalog is open-weight only, and US-only serverless carries a 50% premium on select models.
Standout feature
Standard, Priority, Fast, Batch and Spot modes can be chosen per request on the same endpoint.
Use cases
AI product teams
Send interactive chats on Fast mode and background jobs on Batch
Outcome: One endpoint with different prices per request
ML engineers
Fine-tune an open model with SFT and serve it serverless
Outcome: A custom model without managing GPUs
Agent platforms
Keep long agent sessions on cached context
Outcome: Repeated input billed at half price or less
Key features
- Serverless open-weight models including Kimi K3, GLM 5.3, DeepSeek V4.1 Flash and OpenAI GPT OSS 120B
- Standard, Priority, Fast, Batch and Spot modes switchable per request on the same endpoint
- Batch at 50% off and Spot at 30% off Standard on select models
- Cached input billed at half price or less with no cache write fee
- Managed training from $0.50 per 1M training tokens with SFT, DPO and ORPO
Pros
- Serverless prices start at $0.05 per 1M input tokens
- Reserved Throughput adds a throughput SLA and 99.9% uptime terms
- Monthly spend limits and auto reload on prepaid credits
Cons
- No first-party frontier models from OpenAI, Google or Anthropic
- Spot capacity can drop to zero and needs retry logic
- US-only serverless costs a 50% premium on select models
Rank 8
Together AI
AI cloud for open-source models, with serverless and batch inference, dedicated endpoints, fine-tuning and GPU clusters.
together.ai8.4
out of 10
Best for: Teams that want a large catalog of open-weight chat, image, video and audio models plus a path to dedicated GPUs.
Together AI is a broad platform for open-source models. Its serverless catalog lists chat models from GLM-5.3-Flash at $0.15 in and $0.50 out per 1M tokens and gpt-oss-120B at $0.15 and $0.60, up to Kimi K3 at $3 and $15, with cached input prices shown for many of them. The same account covers image models such as FLUX.2, video models, text to speech, transcription with Whisper Large v3, embeddings, rerank and moderation.
When serverless is not enough, Together offers batch inference, provisioned throughput with SLAs, dedicated model inference, dedicated containers for custom models and GPU clusters up to GB300. Managed fine-tuning supports SFT and DPO, and a post-training service supports custom RL and SFT loops. The breadth is the strength and the weakness: the pricing page lists dozens of models and products but not context windows, so plan on checking model pages. Like other open-model hosts, it does not serve closed frontier models.
Standout feature
Chat, image, video, audio and embedding models plus GPU clusters on one account.
Use cases
Product teams
Compare several open chat models on the same API before choosing one
Outcome: Model swaps without new contracts
Creative apps
Generate images and short videos alongside chat
Outcome: One bill for text, image and video models
Scaling startups
Move a high-traffic model from serverless to a dedicated endpoint
Outcome: Predictable capacity with an SLA
Key features
- Serverless chat models including Kimi K3, GLM-5.3, DeepSeek V4.1 Flash, Qwen3.7 and gpt-oss-120B
- Batch inference, provisioned throughput with SLAs and dedicated model inference
- Image, video, audio, transcription, embedding and rerank models on the same account
- Fine-tuning with managed SFT and DPO jobs and a post-training service
- GPU clusters with H100, H200, B200, B300, GB200 and GB300 hardware
Pros
- Wide catalog across text, image, video and audio
- Cached input prices published for many chat models
- Clear path from serverless to dedicated endpoints and GPU clusters
Cons
- No first-party frontier models from OpenAI, Google or Anthropic
- Context windows are not listed on the pricing page
- Many products can make it harder to find the right option
Rank 9
Amazon Bedrock
AWS managed service for foundation models from Amazon and other providers, billed through your AWS account.
aws.amazon.com8.1
out of 10
Best for: Companies already on AWS that want many model providers under AWS billing, IAM and regional controls.
Amazon Bedrock is the widest marketplace on this list. Its pricing page lists models from AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Google, Luma AI, Meta, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, Qwen, Stability AI, TwelveLabs, Writer, xAI and Z AI, all billed through your AWS account. That is the main reason to choose it: one contract, AWS IAM and regional controls, and existing committed spend.
Bedrock supports Standard, Flex, Priority and Reserved service tiers, batch inference at 50% below on-demand prices for select models, and Provisioned Throughput. You can fine-tune, run preference optimization or distill models, and bring your own weights through Custom Model Import. Per-model prices are set per region and load through the page's region selector, so we could not confirm USD rates from our location and list pricing as pending. Expect more setup than a direct API key, and check whether the model version you want is live in your region.
Standout feature
Models from 19 providers, including Anthropic, OpenAI and Google, under one AWS account.
Use cases
Enterprise platform teams
Offer several model providers to internal teams through AWS IAM
Outcome: Central access control and one AWS bill
Regulated companies
Keep inference in a chosen AWS region
Outcome: Model calls stay inside existing cloud governance
ML teams
Distill a large model into a smaller custom one
Outcome: Lower per-request cost for a narrow task
Key features
- Models from Amazon, Anthropic, OpenAI, Google, Meta, Mistral AI, DeepSeek, Cohere, Qwen, xAI and more
- Standard, Flex, Priority and Reserved service tiers
- Batch inference at 50% lower prices than on-demand for select models
- Fine-tuning, preference optimization and distillation for custom models
- Custom Model Import and Provisioned Throughput
Pros
- Largest provider list on this page under one bill
- Uses existing AWS accounts, IAM and committed spend
- Batch inference halves on-demand prices on select models
Cons
- Per-model prices load through a region selector, which makes comparison slower
- More setup than a single-vendor API key
- Newest models may arrive on first-party APIs before Bedrock
Rank 10
Azure OpenAI
Microsoft's Azure service for OpenAI models, deployed in your Azure subscription with Global and Data Zone options.
azure.microsoft.com7.9
out of 10
Best for: Organizations on Microsoft Azure that want OpenAI models with Azure billing, data zone deployments and enterprise controls.
Azure OpenAI brings the GPT-6 and GPT-6.1 series into an Azure subscription. The pricing page lists GPT-6 Astra, Sol and Luna and GPT-6.1 Sol and Terra, each with Global and Data Zone deployments and separate short-context and long-context rates. Priority Processing and Flex pricing are available on select models, and as of November 1, 2026, explicitly requested 12-hour prompt cache writes are billed at their own rate.
The case for Azure over OpenAI's own API is governance. Deployments live in your Azure tenant, Data Zone options keep processing in a geography, and spend counts toward Azure commitments. US government entities can buy through Azure Government. The page renders prices by script and currency, so we could not confirm USD rates from our location and list pricing as pending. Setup takes longer than a direct API key, and you will need to pick a deployment type and region per model.
Standout feature
OpenAI's GPT-6 models with Data Zone deployments inside your own Azure subscription.
Use cases
Enterprise IT
Deploy GPT-6 models inside the company Azure tenant
Outcome: Model access under existing Azure policies and billing
EU and US regulated teams
Use Data Zone deployments for residency requirements
Outcome: Processing kept within the chosen geography
Public sector agencies
Buy model access through Azure Government
Outcome: A procurement path that fits government rules
Key features
- GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol and GPT-6.1 Terra models
- Global and Data Zone deployment types
- Priority Processing and Flex pricing on select models
- Separate short-context and long-context pricing
- Prompt caching with 30-minute and 12-hour cache writes
Pros
- OpenAI models inside an Azure subscription and its governance
- Data Zone deployments for data residency
- Azure Government purchase options for US public sector
Cons
- Prices load by script and currency, so they are slower to compare
- More setup than a direct OpenAI API key
- Long-context requests are priced separately from short ones
Frequently asked questions
+What is the best LLM API in 2026?
OpenAI API is our top pick because its GPT-6 models range from $0.10 to $10 per 1M input tokens, all with a 1.05M-token context window and built-in tools. Google Gemini API and Anthropic Claude API are the strongest alternatives.
+Which LLM API is cheapest?
Among the prices we verified, Fireworks AI serverless starts at $0.05 per 1M input tokens and Groq at $0.075 for GPT OSS 20B. OpenAI GPT-6 Luna, Claude Haiku 5.5 and Ministral 3 3B start at $0.10, and DeepSeek charges $0.15 off-peak.
+Is there a free LLM API?
Google Gemini API has a free tier with free input and output tokens on selected models, though free-tier content is used to improve Google's products. Groq also has a free tier with lower rate limits than its paid Developer tier.
+Which LLM API has the largest context window?
OpenAI lists 1.05M tokens on its GPT-6 models. Anthropic's four current Claude models, DeepSeek's models and Gemini 4 Argon list 1M tokens, and DeepSeek allows up to 384K output tokens.
+Should I use a model maker's API or a cloud marketplace?
Direct APIs from OpenAI, Google and Anthropic get new models first and are quickest to set up. Amazon Bedrock and Azure OpenAI fit companies that need cloud billing, IAM, data residency and committed spend under one contract.
+How can I lower my LLM API bill?
Use prompt caching for repeated context and batch processing for work that can wait; OpenAI, Anthropic, Gemini, Fireworks AI and Bedrock list 50% batch discounts. Route simple requests to a small model and keep the large model for hard ones.
About this list
Updated November 9, 2026 by the Mojolin research desk. Mojolin is self-funded
Questions or corrections: editorial@mojolin.com
More best software lists