AI inference providers explained: the top platforms in 2026 and how they compare

Pick a model and the next question arrives before the first one is settled: where does it run? The same open-weight text model is sold by at least three of the platforms below, at input prices that differ by nearly four to one. The picture model you just chose sits on three catalogs with three different price units. The word for all of them is “inference provider,” and it is one word for five different businesses.

This guide maps the five shapes, then walks the top platform in each, using numbers read from the vendors’ own pricing pages on September 10, 2026. By the end you should know which key, or keys, to hold. The short answer for most people who generate text, images, audio and video: one media host, one router, and a local runtime for everything that can stay on the machine.

A provider is where the model runs, and that is a separate choice

A model vendor trains the weights: OpenAI, Anthropic, Google, Meta, DeepSeek, ByteDance. An inference provider owns the GPUs those weights run on and sells the output, by the token, by the image or by the second. Sometimes vendor and provider are the same company. OpenAI hosts GPT-5.6 itself. Often they are not. Meta trains Llama and hosts none of it for you, while Together, Fireworks, Groq, DeepInfra and Amazon all do. ByteDance’s Seedream image model is on Runware, fal and Replicate, priced three ways.

A gateway, or router, is a third thing: a switchboard that holds keys to many providers and resells them behind one API. We covered how gateways work, what they quietly cost and who now owns them in the gateways and routers post. Here OpenRouter is treated as one provider shape among five, and that story is not re-argued.

The shapes matter because they price differently, host different things and fail differently. A media host bills per output and does not charge for a failed generation. A GPU host bills per token, with a cheaper rate for input it has seen before. A hyperscaler bills through a cloud account your finance team already pays. The hero above lays out the five. The rest of this post walks them in the order most readers meet them.

Media hosts bill by the picture or the second

If you generate images, video or audio, read the price unit before the price. A per-image price tells you what a picture costs before you make it. A per-second GPU price tells you only after you know how long the model ran, and for a diffusion model that varies with resolution, step count and whether the model was warm. Both units exist on these three platforms, sometimes on the same page.

The four illustrations in this post came from Runware’s API while it was being written: Seedream 5 Pro at 2560 by 1440, first result each, $0.39 for all four, 46 to 59 seconds per image. That is the whole point of per-output pricing. The number was known before the request went out.

Open-weight GPU hosts turn text into a commodity

Open-weight text is the most competitive market in AI. The same DeepSeek V4 Flash checkpoint is on all four of these platforms, and its input price runs from $0.06 to $0.22 per million tokens depending on which one you ask. Every one of them speaks the OpenAI request shape, so switching is a base URL and a model name. That is why prices converge, and why the differences that remain are speed tiers, cache discounts and billing terms.

A router sells breadth for a 5.5% fee

OpenRouter owns no GPUs. It holds keys to the hosts above and to the first-party APIs, and resells all of them behind one OpenAI-shaped key. Stripe’s August 19, 2026 announcement of its acquisition put the count at “400+ models from more than 80 providers.”

First-party APIs sell exclusivity, hyperscalers sell procurement

Some models you cannot get anywhere else on launch day, and the first weeks of a frontier model are often the only weeks that matter for a product decision. That is the first-party business. The prices are public and the discounts are structural: cached input and batch processing, not negotiation.

• OpenAI. GPT-5.6 Terra at $2 in and $12 out per million tokens, GPT-5.6 Sol at $4 and $20, GPT-6 Astra at $10 and $50, with cached input at a tenth of the input price. Image output for gpt-image-2 is $30 per million output tokens. Batch runs at half price (pricing).
• Anthropic. Haiku 4.5 at $1 and $5, Sonnet 5 at $2 and $10, Opus 5 at $5 and $25, Fable 5.1 at $10 and $50. Cache reads cut input to $0.10 to $0.50 per million; batch is 50% off (pricing).
• Google. Gemini 3.8 Flash at $0.75 and $3.75 through December 31, 2026, then $1.50 and $7.50. Gemini 2.5 Flash-Lite at $0.10 and $0.40. A free tier exists, with the note that content “may be used to improve our products.” Veo 3.1 video runs $0.05 to $0.60 per second, and Gemini 2.5 Flash Image is $0.039 per picture (pricing).

Hyperscaler gateways sell the same models to a different buyer. Amazon Bedrock’s model catalog lists eighteen providers, from Anthropic and OpenAI to DeepSeek, Qwen, Mistral, xAI and Z.AI, billed on demand per token, at half price in batch, or as provisioned throughput on one-month and six-month terms (pricing). Google’s Vertex AI, which the docs now file under its Gemini Enterprise Agent Platform, hosts Claude Fable 5.1, Opus 5 and Sonnet 5 beside Gemini with Standard, Priority and Flex pay-as-you-go lanes and provisioned throughput (docs). Microsoft Foundry claims “over 11,000” models including OpenAI, Anthropic, Meta, Google and xAI, billed per service (product page).

What you buy from a hyperscaler is not a model. It is one invoice, one identity system, a data-residency choice and compliance paperwork someone already signed. What you give up is the router’s breadth and the GPU hosts’ bargains. The price you see is the list price, and the model you want sometimes arrives there after it arrives at source.

The whole map on one table

Read the table by the column that matters to you. If you generate pictures, the “bills by” column decides. If you run one open text model at volume, “weights” and “free” narrow it to four rows and the price cards above settle it. If you need a signature from procurement, only the last row applies.

The scenario list below is the same table read backwards, from the need to the key. It assumes you are a builder or a power user paying your own bill; a company with a cloud contract should start from the last line.

Two things the table cannot show. First, per-output prices and per-token prices are not comparable until you know your task, which is the argument of our post on what one AI image costs. Second, every number here moves. Google’s Gemini 3.8 Flash doubles on January 1, 2027; Fireworks raised on-demand GPU rates on September 1; Together’s H100 price is a promotion. Treat the cards as a snapshot with links, not a contract.

Three keys and a local runtime cover nearly everything

Here is how the map turns into a decision. Hold one media host key, because per-output pricing is the honest unit for pictures, video and sound. Hold one router key, because open text is a commodity and the router lets you follow the price without rewriting anything. Run a local runtime for everything small enough to stay on the machine, which is more than most people expect; the cloud API versus local runtime post draws that line. Add a first-party key only when a model is exclusive, and a hyperscaler only when procurement demands it.

That is the shape CSuite is built around. The desktop app ships Runware, Replicate and OpenRouter as cloud platforms next to local Ollama, Hugging Face and GGML runtimes. A model appears once in the catalog with the platforms able to run it listed beside it, so the same Seedream or DeepSeek entry runs under whichever key you hold. The keys never leave the machine: they sit in the app’s local database, encrypted through the operating system’s keyring, the Keychain on macOS, DPAPI on Windows and Secret Service on Linux.

The provider question does not go away after you answer it. Prices move, catalogs change, and a company that owned no GPUs last year is owned by a payments company this year. But it gets easier once you stop treating “provider” as one word. There are five businesses in it, and you only need to be a customer of two.