Every serverless AI inference platform sells you the same dream: call the API, pay per token, scale to zero. Inferock is the first one we have reviewed that puts a receipt on that dream — an independent, per-call record of what each LLM API call cost, how long it took, and whether it actually worked — plus credits back for failed calls that billed you anyway. Verdict up front: if you are shipping AI features where failed or unusable model calls are quietly inflating your bill, Inferock’s Watch tier is free and worth an afternoon of your time. Serve, the pay-per-token tier, deserves a proper trial before you route production traffic through it. This review walks through what Inferock does, what it costs, where it is honest about its limits, and who should use it. It is written for technical founders and product teams — the people who will have to justify the inference bill at the end of the month.
Inferock is a young San Francisco company (2–10 people, per its LinkedIn) that launched this serverless product on Product Hunt on 14 October 2026, finishing the day as the #2 Product of the Day. It was the company’s second launch — the first, Inferock Bench in August 2026, was a local proxy generating an independent receipt for every LLM API call. The October launch turns that measurement idea into a full inference service with an unusual promise: when a call fails and you are billed for it, you should have a clear record, and a path to get something back.
What this serverless AI inference platform actually does
Strip away the marketing and Inferock does three things. First, it routes your model calls — through your own provider accounts, its managed serverless tier, or dedicated capacity. Second, it measures each response independently: quality findings against your app’s requirements, timing, token usage, and available cost data. Third, it writes you a receipt for every call, so you can trace each failure back to the exact call and the evidence recorded.
The measurement layer is the real product. Most teams rely on the provider’s dashboard and their own application logs, which disagree constantly — retries, streaming chunks and partial failures make provider-side bills hard to reconcile with what your users actually got. Inferock sits in the middle and keeps its own books: provider status, failure class, and tokens billed versus tokens actually delivered, per call.
The three routes: Watch, Serve and Dedicated
Inferock offers three ways to run, and the billing model and credit terms change between them:
Inferock Watch — bring your own keys
Watch is the observability play. You connect your own OpenAI, Anthropic, Gemini, or OpenRouter-shaped API keys, and Inferock measures calls in the path. Your providers bill you separately; Watch itself is free on your own keys, with no Inferock service credit in this mode — the bill stays between you and your providers. The loss reporting shows objective failures, thresholded quality signals, and review-only flags. This is the tier to start with: it costs nothing, touches no procurement, and tells you within days whether failed calls are a real line on your bill.
Inferock Serve — managed serverless inference
Serve is the serverless AI inference platform proper. You prepay through Inferock, no provider API key required, and get pay-per-token access to models with the same independent measurement attached. Models and service terms are confirmed during onboarding — availability, charging start and terms are agreed before activation. Signup credits are offered for evaluation. On Serve, eligible objective failures earn bounded service credits under the agreed terms, and credits from Watch-tier failure records can be used toward Serve usage.
Inferock Dedicated — reserved capacity
Dedicated is reserved capacity for predictable, high-volume workloads: price and terms agreed before provisioning, contact-sales onboarding. For most startups, Watch and Serve are the relevant tiers.
Pricing: what we could verify
Inferock does not publish a public per-token price list. Here is what we could verify from primary sources — the official site and the company’s own Product Hunt and LinkedIn statements:
- Watch: free on your own provider keys. You keep paying OpenAI/Anthropic/Gemini directly; Inferock takes no cut and grants no service credits in this mode.
- Serve: prepay, pay-per-token. Exact per-model rates are confirmed during onboarding, not published. Signup credits are offered for evaluation.
- Dedicated: custom quote after a workload discussion.
- Failed-call credits: automatic on Serve for objective failures — errors, timeouts, and empty answers that billed tokens. Watch-tier eligible failures earn credits toward Serve usage rather than cash refunds.
The honest reading: the credit mechanism is a loyalty and acquisition device, not a service-level agreement. You do not get money back; you get credits that keep you inside Inferock’s ecosystem. Still genuinely useful — most inference spend is sticky anyway — but do not confuse it with a refund policy.
Who it’s for — and who should skip it
- Teams shipping AI features with retry logic. Retried failed calls bill you twice for one answer, and your provider bill will not show it clearly. Watch surfaces this within days.
- Agencies and product teams on client budgets. When you must explain an inference line item to a non-technical client, an independent per-call receipt beats any provider dashboard. We have written before about what AI integration work actually costs to run — the hidden part is always the unusable-call tax.
- Startups deciding between direct provider contracts and aggregators. Watch gives you real failure and cost data on your own keys before you commit to any managed tier.
- Skip it if your AI usage is a few hundred calls a day of straightforward chat completions. Your failure rate is probably too low for the receipt to earn its keep, and the setup cost outweighs the insight.
- Skip it (for now) if you need certified confidential inference today. Inferock’s end-to-end encrypted route is described as intended architecture that requires qualification evidence for each exact model and configuration — it is a roadmap, not a feature, and it does not apply to ordinary Watch connections.
Watch-outs and honest limitations
No tool review is complete without the parts the vendor would rather you skim:
- “Completed but unusable” is not credited yet. On the Product Hunt launch thread, the makers confirmed the receipt tracks what the gateway can see — errors, timeouts, and empty answers that billed tokens. Valid JSON that fails your own application’s field checks looks complete to the gateway, so it is neither credited nor flagged separately today — though the makers called app-side unusability marking a sensible next step. If your biggest cost is outputs that parse fine but are wrong, the credit mechanism does not cover that yet.
- Telemetry trust is still maturing. A launch-thread commenter flagged that the terms were vague about payload telemetry — fair criticism for a product sitting in the middle of your prompts. The makers acknowledged it and promised audits and refined terms. Until those exist, treat the privacy story as a promise in progress.
- No public per-token pricing. Serve’s rates are confirmed during onboarding. That is normal for an early-stage infrastructure product, but it means you cannot comparison-shop without a sales conversation.
- The 20% savings claim is theirs, not ours. The company says customers it is working with are seeing more than 20% cuts in AI spend — plausible for retry-heavy workloads, but a vendor number measured on friendly traffic. Verify it against your own receipts before you repeat it.
How it compares to the alternatives
Inferock sits between three established categories — know which problem you actually have before picking one:
- Direct provider APIs (OpenAI, Anthropic, Google). Cheapest per token at scale, zero middlemen, but you own all the observability and you eat every failed call. Inferock Watch pairs well with direct APIs — you keep the pricing, you gain the receipts.
- Aggregators like OpenRouter. One key, hundreds of models, automatic fallbacks. The model is breadth and routing, not accountability — no independent receipt, no failed-call credits. If model choice matters more than cost forensics, OpenRouter wins.
- Observability platforms (LangSmith, Langfuse, Helicone). These trace and evaluate your AI pipeline in depth, often with richer evaluation tooling than Inferock’s checks. What they do not do is attach a remedy: a beautiful trace of a failed call still leaves the money on the table. Inferock’s differentiator is that the receipt comes with a defined remedy for eligible failures.
The redundancy risk is real: teams already on Langfuse or Helicone may find Inferock’s measurement duplicative and only care about Serve’s credits. Teams with no observability at all get the most value — Watch is free, and free beats a second paid dashboard.
From our experience: the unusable-call tax
After nine years of building client software — 300+ projects, a good share of them AI features in the last three — the most expensive line on an inference bill is rarely the token price. It is the denominator. Inferock’s own site illustrates this honestly with a worked example: at $10 per 1,000 billed calls, if 10% of those calls are unusable, each usable response actually costs $11.11 — an 11.1% tax that never appears on any provider invoice.
Two pieces of practical advice. First, measure your own unusable-call share before you optimise token prices — teams usually negotiate the visible 5% discount while ignoring the invisible 20% waste. Second, Watch pays for itself fastest as a diagnostic: run it for two weeks on your own keys, find out where your retries and empty responses actually come from, fix the prompts and timeouts, and keep the savings forever even if you never touch Serve. And if you want more tools like this one researched properly, our AI Tools Directory is a good place to continue.
Frequently asked questions
Is Inferock free to use?
Inferock Watch is free when you bring your own provider keys — you keep paying OpenAI, Anthropic, Gemini or OpenRouter directly. Inferock Serve, the managed serverless tier, is prepaid with pay-per-token billing and signup credits for evaluation; exact rates are confirmed during onboarding.
What counts as a failed call for the credit-back?
Objective failures the gateway can verify: errors, timeouts, and empty answers that billed tokens. These are flagged and credited automatically on Serve. Outputs that complete but fail your own application’s checks are recorded as failed checks on the receipt but are not credit-eligible today.
Which models can I run on Inferock Serve?
Serve offers managed access to open models, with the exact model list and service terms confirmed during onboarding. Watch works with your own keys across OpenAI, Anthropic, Gemini, and OpenRouter-shaped endpoints.
Does Inferock see the content of my prompts?
As a gateway, Inferock measures calls in the path — treat it like any other intermediary until its confidential-inference route is qualified. The company describes end-to-end encrypted inference as intended architecture requiring evidence for each exact model and configuration, and has promised refined terms and audits after launch feedback. Review the current terms before sending sensitive data.
Should my startup use Inferock or just call providers directly?
Start with Watch on your own keys — it is free and tells you within days whether failed and retried calls are a real cost for you. If the receipts show meaningful waste, trial Serve with the signup credits before committing production traffic. If your usage is light and simple, direct provider calls with basic logging are enough.
Inferock is that rare AI infrastructure launch with a genuinely new idea: not cheaper tokens, but accountability for the tokens you already bought. The credit mechanism is narrower than the headline suggests, the privacy story is still maturing, and pricing needs a sales conversation — but the free Watch tier makes the core claim testable in an afternoon, and that is more than most launches offer. If you are evaluating inference infrastructure for a product you are building, talk to us — we help teams get the cost model right before a single token is billed.