Inference providers: the honest 2026 comparison

The five kinds of inference provider

"Inference provider" covers companies with genuinely different business models, and comparing them on a single price column produces nonsense. Sort them first, then compare inside a tier.

The tier boundary that catches people out is the third one. Modal and RunPod do not sell you a model. They sell you a container with a GPU attached, priced by the second, and everything above that line is your problem: weights, server, batching, autoscaling, and the on-call rotation when it falls over. That is the right trade for a fine-tune or an embedding batch and the wrong one for "I want a chat completions endpoint by lunchtime."

What the same model costs at each provider

Three models are served by nearly everyone in the open tier, which makes them a fair yardstick. All figures below were read from each vendor's own pricing page in August 2026, and they are per million tokens, input first.

DeepInfra is the exception, undercutting by roughly 5 percent on Kimi K3 and by about 35 percent on DeepSeek V4 Flash. That discount is real, and so is the reason for it: DeepInfra competes on price where the others compete on latency and support. If your workload is a batch job that nobody is watching, the cheap row is the right row. If it is an interactive agent, buy the latency.

The frontier tier does not have a price column

Claude and GPT models are single-source. There is no shopping around, only choosing which frontier model you can afford.

Two footnotes that change the arithmetic more than the headline rates do. Anthropic prices the full 1M-token context window at the standard rate on Claude 4.6 and later, so a 900k-token request costs the same per token as a 9k one. OpenAI does not: on the GPT-5.6 family, requests past 272K tokens pay double on input and 1.5x on output, so gpt-5.6-sol becomes $10.00 / $45.00. For a long-running coding agent that accumulates context, that surcharge is a bigger line item than the choice of provider.

Scoring a provider on the things that are not price

Since price has converged, here are the five axes that still separate them, and how each tier actually performs on each.

Speed is the one with published numbers

Artificial Analysis measures every provider on the same 60-prompt set and publishes both throughput and time to first token. On the August 2026 board, Groq led output speed at 457 tokens per second (Qwen3.6 27B), with Nebius at 453 on Nemotron 3 Ultra, Google at 372 on Gemini 3.7 Flash, and Databricks at 347 on GLM 5.2. The lowest time-to-first-token figures cluster near half a second, with Baseten posting 0.49 seconds on Inkling Small.

Those are two different races, and which one you care about depends entirely on the shape of your workload. See tokens per second versus time to first token for why an agent that runs twenty tool calls per task is bound by the second number and a chat UI is bound by the first.

Free tiers are mostly a formality

Fireworks gives $1 in credits and then requires a payment method. Together publishes no ongoing free credit on its pricing page as of August 2026. Modal is the genuine outlier: the Starter plan is $0 per month and includes $30 of compute credits every month, renewed, which is enough to run real embedding jobs indefinitely. OpenRouter grants 50 free-model requests per day, rising to 1,000 per day once you have purchased at least $10 in credits.

Which one for a coding agent

Coding agents are an unusual inference workload and most provider marketing is not written for them. A single task is a long chain of short model calls separated by tool execution: read a file, think, run a test, read the failure, patch. Three consequences follow.

• Per-turn overhead dominates. Twenty turns at 0.8s of time-to-first-token is 16 seconds of pure waiting, which the user experiences as the agent being slow no matter what the tok/s figure says.
• Cache pricing is the real bill. The system prompt, the repo context, and the transcript are resent every turn. A provider charging 10 percent of input for cache reads (Anthropic) versus 20 percent (some open shops) changes the monthly total more than a 5 percent list-price gap.
• Frontier quality still wins on hard edits. A cheap model that makes a wrong patch costs you a whole turn plus the human review, which is more expensive than the tokens it saved.

Practically: run the cheap open models for mechanical work (renames, test scaffolding, docstrings) and a frontier model for anything requiring judgment. That is exactly what a router does, and it is why model routing pays off faster on agent traffic than on chat traffic. Which cheap model is good enough is an empirical question rather than a price question: our coding-model leaderboard puts each model's measured pass rate next to its cost per task, and every model card carries the rows behind it.

When you do not need one of these at all

Three cases where hiring an inference provider is the wrong move.

• You already pay a subscription. Claude Max and similar plans include a large usage allowance at a flat price. Buying tokens on top of an underused subscription is paying twice. Check the crossover arithmetic before adding an API bill.
• Your volume is trivial. Under roughly a million tokens a month, every option costs less than a lunch and the correct optimization is to stop optimizing.
• The model fits on hardware you own. A 4B model doing classification runs fine on a laptop or a single old GPU. Per-token pricing on tiny models exists because it is convenient, not because it is cheap: Fireworks charges $0.10 per million for anything under 4B parameters, which is real money at scale for a job an idle machine could do for free.

Questions people ask

Keep reading

Continue through Inference providers with Fireworks AI review: what it is good at, and what it is not.