Liquid Inference Review (2026): Pricing, Features & Honest Verdict
TLDR
Liquid Inference is an auction-style LLM router: one endpoint URL, OpenAI or Anthropic format, and providers compete on price for each request. Sound idea, excellent docs, but the book is thin and mostly operator-affiliated. Best for: cost-driven developers on open-weight models. Price: usage from $0.01 per million output tokens plus a 5% fee, free plan no. Rating: 6.5/10.
What is Liquid Inference?
We spotted Liquid Inference on Product Hunt, where it launched on 8 October 2026 as an LLM router where providers compete for every prompt. The idea is an exchange rather than a gateway. Each provider posts a per-token price for each model it serves and can change that price whenever it likes. You set rules such as model, region, speed and a minimum quality, and the cheapest offer that passes them answers.
The product is operated by Architect Financial Technologies, Inc., a Delaware corporation, and the marketing site signs off “by the AI Exchange”, a sister site that describes itself as a U.S. derivatives exchange for GPU compute. The legal entity shows up in the terms, the data processing agreement and the privacy notice, all hosted on the app domain at inference.ai.exchange. The marketing site itself is a one-page app whose footer links for About, Contact, Privacy and Terms all point nowhere, which is a sloppy first impression for a company handling prepaid money.
What Are Liquid Inference’s Key Features?
A price limit before the first token
Every request returns two headers before any output: a job id and the maximum charge for that job. The exchange fixes that figure from the winning quote and your rules, and you are charged only for the tokens the model produces. Failed jobs cost nothing, and a provider that misses its promised time to first token or completion deadline triggers a credit for the whole job. Each finished job gets an itemised receipt you can read by id with a read-scoped key.
Purchasing rules you send in the request body
An optional exchange object rides along with the normal chat payload, so an unchanged SDK can send it. You can limit the charge in dollars, the price per million tokens in and out, the time to first token, the region, which providers or model creators may compete, and whether the provider is allowed to keep the prompt. Patterns like gpt-* or *-mini work for exclusions. Rules can be saved as named presets and even tied to a specific key so a leaked key cannot overspend.
Drop-in OpenAI and Anthropic compatibility
The endpoint URL accepts Chat Completions, the stateless Responses API, Anthropic Messages and a token-count route, plus images, speech, video and embeddings. The docs carry connection guides for Anthropic’s and OpenAI’s coding CLIs, several open-source coding agents, the official Python and TypeScript SDKs and the popular JavaScript and Python agent frameworks. One detail they get right: the Anthropic client adds /v1/messages itself, so you point it at the host without the /v1 suffix. An OpenAPI 3.1 document is served without authentication.
Scoped keys with spending limits
Keys come in three scopes (spend, read, list), are shown once, can carry a USD spending limit per period and an expiry from 1 to 3,650 days. When a key hits its limit the exchange answers 402 even if the workspace balance is higher. Prompt caching follows the Anthropic cache_control convention with five-minute or one-hour retention, and a response cache lives in memory on one router.
What Is It Like to Use Liquid Inference?
We did not open an account, so this describes the flow the vendor’s docs and sign-up page show. Sign-up asks for an email, a password and an optional referral code, with no card. Every new workspace gets a starting balance that the operator sets, so the first requests are free. The Product Hunt launch post said the first 500 email sign-ups get $20 of inference; the docs never print a figure, though the admin console has a campaigns view for “larger signup credits for the first accounts”.
From there you issue a key, swap one URL in your client and send a request. The response carries an exchange block naming the job, the provider, the maximum charge and the real charge. Streaming works in both formats, and the last event carries the finalised charge. The auto mode, still in beta, lets the exchange pick the model by comparing a published Artificial Analysis score against the estimated charge, but it refuses to run without some price bound.
Providers see a mirror of this. Registration is free, admission is manual, an agent holds a WebSocket session open and republishes quotes, and the exchange sends its own test jobs to every listed deployment on a schedule and pays for them. Sellers get a performance page with a one-hour delay that explains why they lost auctions without ever showing a rival’s bid.
Where Does Liquid Inference Fall Short?
The auction is more promise than practice today. We pulled the public book on 12 October 2026: 823 resting offers across 756 model markets, but 715 of those markets have exactly one independent provider, and only six are flagged liquid (DeepSeek V4.1 Flash with six bidders, two GLM 5.3 variants, one Kimi model, one Qwen model and one more DeepSeek). The sign-up page itself says “Max bidders 6”. On most names there is nobody to compete with.
Worse for the independence story, 719 of the 756 markets carry a single affiliation group named grp_architect. The provider docs define that field as the group a provider “counts under for market depth”, and the exchange derives it from the service each provider declares it depends on. Read plainly, the operator or parties affiliated with it supply most of what is on sale. Region rules are mostly moot too: 810 of the 823 offers sit in the US, with one EU offer.
Volume is small. The public lifetime figures showed 24,161 jobs, 1.56 billion input tokens and 13.8 million output tokens since the exchange opened. The marketing site promises a public download of “live and past prices for every model”, yet the market-data docs say quote history and price trends are console-only internal endpoints outside the OpenAPI document. The public feed shows current offers without provider names.
The terms need an editor. Section 15.3 is a broken sentence ending “or unused Credits are forfeited, except as required by law”, yet section 8.4 cites it as the clause describing a refund of remaining credits on termination. Credit purchases and the exchange fee are otherwise non-refundable. We also counted 25 offers whose names include “abliterated” or “derestricted”, community fine-tunes with safety training removed, on the same book as a terms clause that forbids bypassing a model’s safety controls.
How Much Does Liquid Inference Cost?
There is no subscription. You prepay credits by card, USDC or bank transfer, and each job is charged at the winning quote plus the exchange’s fee, which the console receipt labels as 5% of the provider’s settled amount. No minimum deposit is published. At the time of reading, the cheapest output offers were GPT OSS 120B at $0.0034 per million output tokens (the marketing site rounds it to $0.01) with input at $0.05, and DeepSeek V4 Flash at $0.005 out and $0.05 in. Those prices move whenever a provider re-quotes.
Closed families are pricier and single-sourced: the lowest Gemini offer was $0.39 per million output tokens, the lowest Anthropic and xAI offers $0.50, Mistral $0.11, Kimi $0.10 and the OpenAI O-series $4.39. Each of those sat with one provider, so you are paying list price on a reseller’s book, not an auction price.
Against rivals, the fee is the difference. The largest credit-based router charges 5.5% with an $0.80 minimum on every credit purchase, so Liquid Inference’s 5% on usage is a hair cheaper and only applies to what you spend. A serverless host lists gpt-oss-120B at $0.15 in and $0.60 out per million tokens, and another lists DeepSeek V4 Flash at $0.09 in and $0.18 out. Liquid Inference’s resting quotes undercut both by a wide margin, but with one or two bidders behind them you should test throughput before moving traffic.
Referrals pay in credit, not cash: the launch post said 20% of a referred user’s fees for the first tier and 10% for the second, while the docs say the rates and period are whatever the Referrals page shows and can change. A referred person gets 10% off the fee by default. Credits are paid once a day, only on money the referred person actually deposited, and cannot be withdrawn.
| Plan | Price | Plan Features | Best For |
|---|---|---|---|
| Pay as you go | Winning quote plus 5% fee | Prepaid credits by card, USDC or bank transfer; from $0.0034 per 1M output tokens on GPT OSS 120B at the time of reading; no minimum deposit published | Developers routing open-weight models at volume |
| Starting balance | $0 | One-off credit set by the operator on every new workspace, no card; the launch post said $20 for the first 500 email sign-ups | Testing the endpoint before depositing |
| Provider (sell side) | $0 registration | Manual admission, WebSocket quoting agent, public dispatch endpoint; the exchange fee is charged to buyers, not sellers | Teams with idle inference hardware |
Who is Liquid Inference Best For?
Use Liquid Inference if you run high-volume batch or agent workloads on open-weight models like DeepSeek, GLM, Kimi or GPT OSS, you can tolerate variable latency, and you want a hard dollar limit per request enforced by the gateway rather than by your own code. The scoped keys with spending limits are a real safety feature for unattended agents, and the receipts simplify cost attribution.
Skip Liquid Inference if you need frontier Anthropic, Google or OpenAI models, because those sit at one provider each and the auction adds nothing. Skip it if you need EU-only processing, because the book has one EU offer. And skip it for regulated data: the terms bar health, payment, credential and government-ID data outright, and the no-retention guarantee rests on provider attestations that Architect says it does not control.
Best Liquid Inference Alternatives
OpenRouter
The established aggregator with hundreds of models and provider-level routing. Its own FAQ says it charges a 5.5% fee with an $0.80 minimum when you buy credits, and 5% on crypto payments; unused credits are refundable but the fee is not. You pay provider list prices, so it beats Liquid Inference on breadth and depth of providers, not on the per-token figure.
Together AI
A serverless host that runs the models itself rather than brokering them. Its pricing page lists gpt-oss-120B at $0.15 per million input tokens and $0.60 per million output tokens, and DeepSeek V4 Flash 0731 at $0.14 in and $0.28 out. No platform fee, predictable throughput, and fine-tuning on the same stack.
DeepInfra
Another first-party host with aggressive open-model pricing. Its page shows DeepSeek V4 Flash at $0.09 per million input tokens ($0.018 cached) and $0.18 out, and DeepSeek-V3.2 at $0.26 in and $0.38 out. Prices are fixed rather than auctioned, which is exactly what some teams want.
Final Verdict: Is Liquid Inference Worth It?
Liquid Inference has the best-written router documentation we have read this year, a real company behind it, a per-request price limit that nobody else exposes this cleanly, and resting quotes on open-weight models that are a fraction of what first-party hosts charge. If your traffic fits the six models with real competition, the economics are hard to argue with, and a 5% usage fee is fair for the brokering.
But an exchange needs sellers, and today most of this one is affiliated with the house. Until independent providers show up on more than a handful of models, treat it as a cheap reseller with excellent plumbing rather than a market. Fund it with a small deposit, point a non-critical workload at it, watch the receipts, and revisit when the depth numbers on the sign-up page move. 6.5/10.
Liquid Inference Pros & Cons
What We Like
- A maximum charge is fixed and returned in a header before the first token, and you pay only for produced tokens.
- OpenAI Chat Completions, Responses and Anthropic Messages on one endpoint URL, with connection guides for the major coding CLIs, open-source coding agents and the official SDKs.
- Scoped keys with USD spending limits and expiry, plus named presets that can be tied to a key.
- Resting quotes on open-weight models such as GPT OSS 120B and DeepSeek V4 Flash were a fraction of first-party host prices at the time of reading.
- Itemised receipts per job, a public OpenAPI 3.1 document and a data processing agreement with EU and UK transfer clauses.
What Could Be Better
- 715 of 756 markets on the public book have one independent provider and only six are flagged liquid, so the auction rarely has anyone to bid against.
- 719 of 756 markets carry the operator's own affiliation group, and 810 of 823 offers are in the US, with a single EU offer.
- The marketing site promises public download of past prices, but the docs say quote history is console-only; its footer legal links go nowhere.
- Credits and the 5% fee are non-refundable, and the terms clause on credits at account closure is a broken sentence that contradicts the refund cross-reference.
Liquid Inference FAQ
What is Liquid Inference?
Liquid Inference is an auction exchange for LLM inference run by Architect Financial Technologies, Inc. You send an OpenAI or Anthropic format request to one endpoint URL, providers quote per-token prices for the model, and the cheapest offer that meets your rules on model, region, speed and quality answers. A maximum charge is fixed before the first token.
How much does Liquid Inference cost?
There is no subscription. You prepay credits by card, USDC or bank transfer and each job is charged at the winning quote plus a fee the console shows as 5% of the provider's settled amount. At the time of reading the cheapest output offers were GPT OSS 120B at $0.0034 and DeepSeek V4 Flash at $0.005 per million output tokens, with input priced separately; prices change whenever providers re-quote.
Is Liquid Inference worth it?
For high-volume open-weight workloads that can tolerate variable latency, yes, with a small deposit and a non-critical workload first. The docs, receipts and per-request price limit are excellent. But 715 of 756 markets have one provider and most of the book is affiliated with the operator, so treat it as a cheap reseller rather than a market for now. We rate it 6.5/10.
What are the best Liquid Inference alternatives?
OpenRouter is the broadest aggregator and charges 5.5% with an $0.80 minimum when you buy credits. Together AI hosts models itself and lists gpt-oss-120B at $0.15 in and $0.60 out per million tokens. DeepInfra lists DeepSeek V4 Flash at $0.09 in and $0.18 out. All three offer fixed prices and deeper provider coverage.
Does Liquid Inference offer a free plan?
No. Every new workspace gets a one-off starting balance set by the operator so you can send first requests without a card, and the launch post said the first 500 email sign-ups get $20 of inference, but after that it is prepaid usage only. Referral rewards are paid as non-withdrawable credit.
Who is Liquid Inference best for?
Developers and small teams running batch or agent workloads on open-weight models like DeepSeek, GLM, Kimi or GPT OSS who want the lowest per-token price and a hard dollar limit per request enforced by the gateway. It is a poor fit for frontier Anthropic, Google or OpenAI models, EU-only processing, or regulated data.






