ZeroGPU Review (2026): Pricing, Features & Honest Verdict

ZeroGPU is an OpenAI-compatible inference API that routes AI workloads to small specialized models on an edge network. Verdict: real savings on some models, exact price parity on others, and embeddings priced 25x the market. Best for: teams running high-volume classification. Price: Pay per token from $0.02/1M (free plan: no). Rating: 6.5/10.

What is ZeroGPU?

ZeroGPU, Inc. sells inference on the argument that most AI workloads do not need a frontier model. Content classification, intent extraction, moderation, PII redaction, sentiment scoring: these are narrow, repetitive jobs, and a 1.2-billion-parameter model fine-tuned for the task can beat a general-purpose giant on both latency and cost. The company builds its own family of small models it calls ZLMs, serves them alongside open-weight models from Qwen, DeepSeek, GLM, Llama, LiquidAI and OpenAI’s open releases, and routes each request to whichever compute is cheapest, including a network it puts at “100K+ Edge devices” across eleven cities from San Francisco to Seoul.

It launched in June 2026 and took second place for the day on Product Hunt. Three makers are credited: KP, Maddy Arvapally and Joshua Goikhman. The company is backed into the Nvidia and Cloudflare startup programs, both linked from its own footer, and its terms of service were last revised on 28 February 2026. The headline claims are “50-70% Lower inference cost vs. traditional GPU clouds” and “Up to 10x Faster inference for specialized AI workloads,” each carrying a footnote that performance varies by workload.

What are ZeroGPU’s key features?

A one-line swap from your existing provider

The API is OpenAI-compatible, so adoption is a base URL change and a model name. You keep your existing SDK, your existing request shape and your existing error handling. This is the single most important thing about the product commercially, because it turns “evaluate a new inference vendor” into “point one workload somewhere else for an afternoon.” Requests go to api.zerogpu.ai/v1/chat/completions with an API key and a project ID header.

Purpose-built small models for the boring jobs

The ZLM family is where the thesis lives. There are IAB content classifiers with 400-token contexts, a domain classifier at 100 tokens, a moderation model at 4,000 tokens and a follow-up question generator. Alongside them sit DeBERTa-v3-small for zero-shot classification against labels you define at request time, and two GLiNER models for entity and PII extraction. These are CPU-friendly models doing one job each, and at $0.02 per million input tokens they are priced to be thrown at production traffic.

An agent-native storefront that sells without a signup

This is the most unusual thing here. ZeroGPU publishes a machine-readable storefront at agents.zerogpu.ai/llms.txt where an AI agent can read the price list, register a credential, buy credits and start calling, all without a human. Payment settles over x402 or MPP in USDC on Base, or by card through a hosted checkout the agent hands to its operator. An autonomous agent with no wallet can create a managed one at no cost. Whether you think that is the future or a solution looking for a problem, nobody else in the inference market has shipped it.

Cost and latency visibility per request

The platform reports latency per request and per model and inference cost down to the token, across projects. For the routing use case this is not a nice-to-have. The entire pitch is that you should send 70-80% of production tasks to a cheaper model, and you cannot make that call without per-model cost and quality numbers in front of you.

How much does ZeroGPU cost?

ZeroGPU is pay-per-token with no subscription, starting at $0.02 per million input tokens and $0.05 per million output on the small classification and extraction models. There is one plan, called Credits, with a $0.01 minimum purchase and a 50-cent minimum by card. The larger models climb from there: gpt-oss-120b is $0.15 in and $0.60 out, qwen3-30b-a3b-fp8 is $0.05 and $0.30, deepseek-v4-flash is $0.07 and $0.14, and glm-5.2 is $1.10 and $3.50.

Check those against a mainstream competitor and the “50-70% lower cost” line gets complicated. Together AI prices gpt-oss-120b at exactly $0.15 input and $0.60 output. That is the same model at the same price to the cent, so on the flagship open-weight model ZeroGPU is at parity, not a discount. On DeepSeek V4 Flash the story flips: Together charges $0.14 and $0.28, and ZeroGPU charges $0.07 and $0.14, precisely half on both meters. The savings are real but they are model-specific, and the site never says which comparison the 50-70% figure comes from.

The embeddings are the problem. ZeroGPU charges $0.50 per million input tokens for all-minilm-l6-v2 and bge-small-en-v1.5. OpenAI charges $0.02 per million for text-embedding-3-small and $0.13 for text-embedding-3-large. Together AI charges $0.02 for multilingual e5 large instruct. So ZeroGPU is asking 25 times the going rate to serve two small open-source embedding models, one of which has a 256-token context against OpenAI’s 8,191. It is also 25 times ZeroGPU’s own price for running a language model on the same infrastructure. We can find no reading of that price that makes sense for a buyer.

One tier is dominated by the one next to it, too. qwen3-30b-a3b-fp8 costs $0.05 in and $0.30 out with a 32,768-token context. deepseek-v4-flash costs $0.07 and $0.14 with a 1,048,576-token context. Deepseek is more expensive on input by two cents and cheaper on output by sixteen, so it wins on total cost for any request where input is less than eight times output, which covers almost every real chat or extraction workload. It also has 32 times the context. There is close to no reason to pick the Qwen model.

On free usage, the machine-readable price manifest lists includedUnits: 0 on every meter for every model. The documentation does mention a free allowance, but it “is reserved for claimed credentials with a verified email,” so there is nothing free until you attach a human mailbox. Treat this as pay-as-you-go from the first token and budget accordingly.

Who is ZeroGPU best for?

Use ZeroGPU if you are running high-volume, narrow inference and paying frontier prices for it. Ad tech intent classification, content moderation queues, PII redaction on inbound documents, lead enrichment, support ticket triage. If you are currently sending a million short classification calls a month to a large general model, moving them to a $0.02 classifier is the kind of change that shows up on the bill immediately, and the OpenAI-compatible endpoint means the migration is cheap to try.

Use it too if you are building autonomous agents that need to buy their own compute. The agent storefront is a real, working implementation of something most vendors are still writing blog posts about.

Skip ZeroGPU for embeddings. At $0.50 per million tokens against $0.02 from both OpenAI and Together, there is no workload where this is the right choice. Skip it if you need a frontier model for reasoning, since glm-5.2 at $1.10 and $3.50 is the ceiling here and that is not what this platform is for. And skip it if you need a published human-readable pricing page, because /pricing returns a 404 and the prices live in the docs and an agent manifest instead.

What are the best ZeroGPU alternatives?

Together AI is the direct comparison and the one we would default to. It serves gpt-oss-120b at the identical $0.15 and $0.60, DeepSeek V4 Flash at $0.14 and $0.28, and embeddings at $0.02 per million. It has a bigger catalog, a real pricing page and fine-tuning. ZeroGPU beats it on DeepSeek by exactly half, and loses on embeddings by 25x.

Groq is the choice when latency is the constraint rather than cost. Its custom inference hardware is the fastest way to serve open-weight models today, and if your workload is user-facing rather than batch, the speed difference is more valuable than a few cents per million tokens. It does not publish per-model pricing on its main pricing page, so you will need to price your specific model through the console.

OpenRouter solves a different half of the problem. Instead of one vendor’s catalog it gives you one API key across dozens of providers with live per-model pricing, so you can route to whoever is cheapest today without integrating each one. If your objection to ZeroGPU is lock-in to a small catalog, this is the answer.

Final verdict: is ZeroGPU worth it?

The thesis is right. Most production AI spend goes on tasks that a small specialized model handles as well as a large one, and a vendor that makes routing those tasks a one-line change is doing useful work. The ZLM catalog is properly aimed at the jobs that actually generate volume, the per-token prices on the small models are aggressive, and the agent storefront is a working answer to a problem the rest of the market is still writing about. On DeepSeek V4 Flash the platform is half the price of Together AI, which is a serious number.

The pricing is where it loses points. There is no human pricing page, the headline savings claim has no published basis, gpt-oss-120b costs exactly what Together charges, one model tier is dominated by its neighbour, and the embedding models are priced at 25 times what both OpenAI and Together charge for equivalent work. None of that is disqualifying if you use ZeroGPU for what it is good at, which is cheap classification and extraction at volume. Point one high-volume classification workload at it, measure the bill against your current provider for a fortnight, and keep your embeddings where they are.