Best forThe broadest model lineup — text, images, speech, embeddings — and the largest ecosystem.
Trade-offFrequent model and API changes to keep up with, and no option to run its hosted models elsewhere.
Used by 736 makers' products for this
Of the 281 products launched in 2026 that we know the LLM API of, none we track use DeepInfra. The most picked now is OpenAI API (70%, up from 84%).
3 of the 31 are open source or self-hostable (LiteLLM, Portkey and Bifrost) and 9 have a free tier. At 10,000 users, the cheapest is Mistral AI at $6.00 a month on Ministral 3 (3B). DeepInfra also answers open model hosting & GPUs: those alternatives are below too.
Not a ranking: each option is described by the situation it fits, with its trade-off.
The share of makers' products we track that use each for LLM API, by the year they launched on Product Hunt. The stack is the one we see today, not the one it launched with; a product can use more than one.
| Option | Launched ≤2023 · 103 | Launched 2024 · 222 | Launched 2025 · 273 | Launched 2026 · 281 |
|---|---|---|---|---|
| 84% | 80% | 70% | 70% | |
| 22% | 24% | 26% | 19% | |
| 8% | 8% | 14% | 11% | |
| 5% | 2% | 6% | 10% | |
| 4% | 3% | 6% | 2% | |
| 2% | 1% | 3% | 2% |
How many makers' products we track use each for LLM API, and what each costs at 10,000 users.
| Option | Kind | Makers' products | At 10,000 users | Free tier | Open source | |
|---|---|---|---|---|---|---|
| DeepInfra | Open models, hosted | 2 | — | No | No | |
| Model provider | 736 | $10/mo · GPT-6 Luna | No | No | ||
| Model provider | 252 | $10/mo · Claude Haiku 5.5 | No | No | ||
| Model provider | 99 | $28/mo · Gemini 3.1 Flash-Lite | Yes | No | ||
| Gateway (many providers, one API) | 65 | — | Yes | No | ||
| Open models, hosted | 39 | $6.75/mo · GPT OSS 20B | Yes | No | ||
| Gateway (many providers, one API) | 21 | — | Yes | Yes | ||
| Model provider | 15 | $6.00/mo · Ministral 3 (3B) | Yes | No | ||
| Model provider | 11 | $27/mo · deepseek-flash | No | No | ||
| Through your cloud | 22 | — | No | No | ||
| Model provider | 6 | — | No | No | ||
| MiniMax | Model provider | 6 | — | No | No | |
| Through your cloud | 3 | — | No | No | ||
| Gateway (many providers, one API) | 2 | — | Yes | Yes | ||
| Eden AI | Gateway (many providers, one API) | 2 | — | No | No | |
| Through your cloud | 1 | — | No | No | ||
| Model provider | 0 | — | No | No | ||
| Gateway (many providers, one API) | 0 | — | Yes | No | ||
| Gateway (many providers, one API) | 0 | — | Yes | Yes | ||
| Qwen API | Model provider | 0 | — | No | No | |
| Kimi API | Model provider | 0 | — | No | No | |
| GLM API | Model provider | 0 | — | Yes | No | |
| Doubao API | Model provider | 0 | — | No | No | |
| ERNIE API | Model provider | 0 | — | — | No | |
| Hunyuan API | Model provider | 0 | — | — | No | |
| Open models, hosted | 0 | — | — | No | ||
| Open models, hosted | 0 | — | — | No | ||
| Cerebras Inference API | Open models, hosted | 0 | — | No | No | |
| Meta Model API | Model provider | 0 | — | No | No | |
| Xiaomi MiMo API | Model provider | 0 | — | No | No | |
| Model provider | 0 | — | — | No | ||
| Model provider | 0 | — | — | No |
How many makers' products we track use each for open model hosting & GPUs.
| Option | Kind | Makers' products | Free tier | Open source | |
|---|---|---|---|---|---|
| DeepInfra | Serverless model API | 0 | No | No | |
| Dedicated model deployments | 77 | — | No | ||
| Your code on GPUs | 60 | — | No | ||
| Dedicated model deployments | 23 | — | Self-hostable | ||
| Your code on GPUs | 19 | — | No | ||
| Serverless model API | 12 | — | No | ||
| Serverless model API | 5 | — | No | ||
| Open-source serving engine | 0 | — | Yes | ||
| Your code on GPUs | 5 | No | No | ||
| Serverless model API | 1 | — | No | ||
| Serverless model API | 1 | No | No | ||
| Serverless model API | 0 | — | No | ||
| Cerebras Inference API | Serverless model API | 0 | No | No | |
| Serverless model API | 0 | Yes | No | ||
| Serverless model API | 0 | No | No | ||
| Your code on GPUs | 0 | No | No | ||
| Your code on GPUs | 0 | No | No | ||
| Your code on GPUs | 0 | No | No | ||
| Your code on GPUs | 0 | No | No | ||
| Dedicated model deployments | 0 | — | No |
Which provider, and whether to go through a gateway. Going direct to a provider gets its newest models and features first and the lowest latency; a gateway gives you one key and one bill for hundreds of models, with fallbacks when a provider is down, for a small markup or a proxy you run. The code that calls the model — an AI SDK or agent framework — is its own decision.
DeepInfra (Open models, hosted): Low per-token prices on a large catalog of open-weight models, plus embeddings, speech and image models, behind an OpenAI-compatible API. Trade-off: No free tier — a card or prepaid balance comes first — and a default cap of 200 concurrent requests per model until you ask for more.
Best forThe broadest model lineup — text, images, speech, embeddings — and the largest ecosystem.
Trade-offFrequent model and API changes to keep up with, and no option to run its hosted models elsewhere.
Used by 736 makers' products for this
Best forCoding, long documents and agent-style tool use.
Trade-offNo image generation or embedding models of its own, so those need a second provider.
Used by 252 makers' products for this
Best forVery long context and multimodal input, with a free tier to start.
Trade-offFree-tier prompts may be used to improve Google's products, so paid tier is the one for user data.
Used by 99 makers' products for this
Best forTrying many models from many providers with one API key and one bill, without opening an account at each.
Trade-offA fee on top of provider prices, and your traffic and data pass through a third party.
Used by 65 makers' products for this
Best forVery low-latency inference on open models, for real-time features.
Trade-offOnly the open models it chooses to host, and no frontier closed models.
Used by 39 makers' products for this
Best forRunning your own OpenAI-compatible proxy in front of your own provider keys, with retries, routing and per-key budgets.
Trade-offYou deploy, monitor and upgrade it yourself; SSO and audit logs are in the paid enterprise tier.
Used by 21 makers' products for this
Best forA European provider whose API covers chat, coding, OCR and speech models.
Trade-offA smaller ecosystem of third-party integrations than the largest providers.
Used by 15 makers' products for this
Best forLow per-token prices on strong reasoning and coding models, with OpenAI- and Anthropic-format endpoints.
Trade-offServed from China, which matters for some customers' data requirements; peak-hour prices are double off-peak.
Used by 11 makers' products for this
Best forOpenAI models inside an Azure account, with regional and EU/US data-zone deployments.
Trade-offMore setup than the OpenAI API (Azure resources, deployments per model), and new models can arrive later.
Used by 22 makers' products for this
Best forGrok models with built-in web and X search, plus image, video and voice APIs.
Trade-offReal-time data only arrives when you enable the search tools, which are billed as tool calls.
Used by 6 makers' products for this
Best forLow per-token prices on open-weight coding and agent models, with speech, image, video and music models in the same account.
Trade-offA smaller ecosystem of integrations and guides than the largest providers; its open-weight models need multi-GPU servers if you want to self-host them.
Used by 6 makers' products for this
Best forGemini and partner models such as Claude on Google Cloud, with its regions, IAM and billing.
Trade-offMore setup than the Gemini API (a Cloud project, service accounts, regions), for the same models.
Used by 3 makers' products for this
Best forA gateway that also handles guardrails, prompt management and request logs in one dashboard.
Trade-offThe free tier caps logged requests; the full control layer is a paid hosted product.
Used by 2 makers' products for this
Best forOne API key and bill for LLMs and non-chat AI — OCR, speech, translation, vision — across many providers, with an EU endpoint.
Trade-offA 5.5% fee on credits on top of provider prices, and your traffic and data pass through a third party.
Used by 2 makers' products for this
Best forClaude, Llama, Mistral, Amazon Nova and other models inside an AWS account, billed with the rest of your AWS usage.
Trade-offIAM, regions and per-model access requests add setup, and new models can arrive later than on the maker's own API.
Used by 1 maker's product for this
Best forOpen-source image, video and audio models without running GPUs yourself.
Trade-offBilled per second of compute, and cold starts on less-used models add latency.
No maker product tracked for this yet
Best forOne key and one bill for hundreds of models at provider list prices, with fallbacks and budgets, built into the AI SDK.
Trade-offTied to a Vercel team and prepaid credits, and bringing your own provider keys needs the paid tier.
No maker product tracked for this yet
Best forA self-hosted, open-source gateway written in Go that routes one OpenAI-compatible API to many providers with failover, load balancing and caching.
Trade-offYounger and with a smaller community than LiteLLM; clustering, guardrails and some governance features are in the paid enterprise edition.
No maker product tracked for this yet
Best forQwen models plus DeepSeek, Kimi and GLM under one Alibaba Cloud account, with OpenAI- and Anthropic-format endpoints.
Trade-offKeys, endpoints and model lists are separate per region, and batch and fine-tuning are only in Singapore and Beijing.
No maker product tracked for this yet
Best forKimi K3, an open-weight model with 1M-token context, and coding models behind OpenAI- and Anthropic-format endpoints.
Trade-offRate limits start very low and rise only with total top-ups; inputs may be used to improve its services unless you arrange otherwise.
No maker product tracked for this yet
Best forGLM coding and agent models through Z.ai, with free Flash models and Anthropic-format access for Claude Code.
Trade-offCaching is still in beta with unpublished rules, and the international and mainland platforms are separate accounts.
No maker product tracked for this yet
Best forByteDance's Seed models alongside its Seedream image and Seedance video models, with an OpenAI-compatible API.
Trade-offNewer Doubao models launch on the mainland platform first; prices step up for longer prompts.
No maker product tracked for this yet
Best forBaidu's ERNIE models through Qianfan, with an OpenAI-compatible endpoint.
Trade-offData may be processed in any country with Qianfan facilities, and the terms tell you not to send confidential information.
No maker product tracked for this yet
Best forTencent's HY models plus DeepSeek, GLM, Kimi and MiniMax on one TokenHub key, in OpenAI and Anthropic formats.
Trade-offHY vision and translation models are only on the mainland site, and processing can leave your chosen region for operational reasons.
No maker product tracked for this yet
Best forPer-token access to a wide catalog of open-weight models, with fine-tuning and dedicated GPUs from the same vendor.
Trade-offSeveral billing models to compare — per token, per GPU-hour, dedicated and reserved — and only the models it chooses to host.
No maker product tracked for this yet
Best forServerless and dedicated inference on popular open-weight models, with fine-tuning and batch jobs at half the serverless price.
Trade-offKeeping requests in the US means choosing US-only model variants priced at 1.5× standard, and priority or fast variants cost more again.
No maker product tracked for this yet
Best forVery fast output — thousands of tokens per second — on a small set of open-weight models.
Trade-offOnly two models on self-serve pay-as-you-go (others need a dedicated contract), a 131k-token context cap, and no discount on cached input.
No maker product tracked for this yet
Best forMeta's Muse Spark models for agents and coding, with a 1M-token context, multimodal input and OpenAI- and Anthropic-format endpoints.
Trade-offClosed weights and a young API (public preview from July 2026); the cheap contributor tier gives Meta the right to train on your prompts and outputs.
No maker product tracked for this yet
Best forVery low per-token prices on open-weight models with a 1M-token context and image, audio and video input, through OpenAI- and Anthropic-format endpoints.
Trade-offIts published terms and data handling are mainland China's (overseas terms weren't found), and older models are retired about a month after a new version ships.
No maker product tracked for this yet
Best forReka's own multimodal models plus hosted open models (DeepSeek, GLM, Qwen) through one OpenAI-compatible API, with fine-tuning.
Trade-offA small provider whose company focus has moved to physical AI and demo-led sales; only the developer platform is self-serve.
No maker product tracked for this yet
Best forThe Yi family of models from 01.AI through its own API.
Trade-offThe model service was discontinued (the API answers HTTP 410); 01.AI now sells enterprise products.
No maker product tracked for this yet
Per-token API or your own code on rented GPUs. Serverless model APIs host popular open models behind an OpenAI-compatible endpoint, so you pay per token and run nothing. Renting GPU compute lets you run any model or pipeline you like, but you package it, handle cold starts and pay for time, busy or idle.
DeepInfra (Serverless model API): Hundreds of open models — chat, embeddings, speech, image and video — through one OpenAI-compatible API, plus dedicated GPUs for your own models. Trade-off: Hosted only; dedicated GPUs bill per GPU-hour while they run.
Best forDeploying almost any model from the Hugging Face Hub to a dedicated endpoint, or calling popular models through Inference Providers, which routes to partner hosts under one API.
Trade-offDedicated endpoints bill per hour while running; Inference Providers only covers models the partner hosts serve, at their prices.
Used by 77 makers' products for this
Best forRunning your own Python code — inference, batch jobs, sandboxes — on GPUs defined in code, scaling to zero between calls.
Trade-offCold starts on large models, and your code is shaped around Modal's decorators.
Used by 60 makers' products for this
Best forDeploying a specific model as a production endpoint with autoscaling and monitoring.
Trade-offBuilt for production workloads; more setup than calling a shared serverless API.
Used by 23 makers' products for this
Best forRenting GPU machines by the hour or running serverless GPU workers, with control over the hardware.
Trade-offMore hands-on setup and operations than a managed platform.
Used by 19 makers' products for this
Best forCalling many popular open models through one OpenAI-compatible API, with fine-tuning on the same platform.
Trade-offYou're limited to the models they host unless you pay for dedicated capacity.
Used by 12 makers' products for this
Best forLow-latency serving of open models, including your own fine-tuned versions.
Trade-offHosted only; dedicated deployments are needed for steady high volume.
Used by 5 makers' products for this
Best forServing open-weight models on your own or rented GPUs behind an OpenAI-compatible API, with high throughput for many concurrent requests.
Trade-offYou own the GPUs, scaling, upgrades and on-call; tuning memory and batching settings takes real effort.
No maker product tracked for this yet
Best forHourly NVIDIA GPU instances and multi-GPU clusters with no egress fees, for training or self-hosted serving.
Trade-offMachines only — its serverless Inference API has been wound down, so packaging, serving and scaling are yours.
Used by 5 makers' products for this
Best forFast open models for coding agents (Kimi K3, GLM-5.3, Qwen, MiniMax, DeepSeek) plus its own Fast Apply, WarpGrep and Compact models, behind one OpenAI-compatible API.
Trade-offA small set of models chosen for agents; it started as a code-edit model company and has since repositioned as an inference provider.
Used by 1 maker's product for this
Best forA wide per-token catalog of open models for text, image, audio and video, with agent sandboxes and GPU instances in the same account.
Trade-offHosted only; sandboxes and GPU instances bill for time, and the catalog is whatever Novita chooses to host.
Used by 1 maker's product for this
Best forVery fast open-model inference on SambaNova's own chips through the SambaCloud API.
Trade-offHosted only on its own hardware, so check the open model you need is in its catalog; it also sells on-prem racks to data centers.
No maker product tracked for this yet
Best forVery fast inference on a small set of open-weight models, on Cerebras's wafer-scale chips, through an OpenAI-compatible API.
Trade-offOnly a handful of models, and reserved capacity is by contract.
No maker product tracked for this yet
Best forVery fast inference for popular open models on Groq's LPU hardware, with a rate-limited free tier.
Trade-offLimited to the models Groq hosts, and the free tier is rate-limited.
No maker product tracked for this yet
Best forOpen-weight models through an OpenAI-compatible per-token API (Token Factory), with dedicated endpoints, batch inference and fine-tuning on the same account.
Trade-offA card is required at sign-up and the trial is $1 of credit; you only get the models Nebius hosts.
No maker product tracked for this yet
Best forRenting H100, H200 or B200 GPUs by the hour with no commitment, plus a small per-token API for open text models.
Trade-offThe per-token API covers only a short list of text models and needs a $5 deposit; you run everything on the GPUs yourself.
No maker product tracked for this yet
Best forThe cheapest GPU hours, from a marketplace of hosts with per-second billing and cheaper interruptible instances.
Trade-offMachines vary by host and prices move with demand; interruptible instances can be stopped, so plan for restarts.
No maker product tracked for this yet
Best forLarge NVIDIA GPU clusters for training and serving, with per-token serverless inference for open models through CoreWeave Forge.
Trade-offBuilt for large teams — the biggest systems go through sales, and it is more than a small product needs.
No maker product tracked for this yet
Best forRunning Ray jobs — batch inference, distributed training, data processing and Ray Serve endpoints — on managed clusters or in your own cloud.
Trade-offYou have to adopt Ray; billed per instance-hour in Anyscale Credits rather than per token.
No maker product tracked for this yet
Best forTrying open models on NVIDIA's hosted endpoints for free, then running the same model as a NIM container on your own GPUs.
Trade-offFree access covers development and testing only; production needs an NVIDIA AI Enterprise license or a partner's endpoint.
No maker product tracked for this yet
It depends on what you need it for; we don't rank them. Among products launched in 2026, the most picked for LLM API instead is OpenAI API (70%). You want the strongest general models and the widest ecosystem of examples and integrations: OpenAI API, Claude and Gemini API. Token cost dominates your budget, for example high-volume batch or agent workloads: DeepSeek, Groq and Mistral AI. You want to try or mix many models with one key and one bill, and fall back when a provider is down: OpenRouter.
Yes. For LLM API, Gemini API, OpenRouter, Groq, LiteLLM, Mistral AI, Portkey and 3 more have a free tier a small product can start on.
Yes. For LLM API: LiteLLM, Portkey and Bifrost.
At 10,000 users, the cheapest is Mistral AI at $6.00 a month on Ministral 3 (3B). From each vendor's pricing page.