Frequently asked questions
Unlimited and flat-rate subscriptions for frontier open-source LLMs — Kimi K2.7, GLM 5.2, MiniMax M3, DeepSeek V4 Flash, and MiMo v2.5. OpenAI- and Anthropic-SDK compatible. Pay by card or USDC on Base.
What is cheapestinference?
cheapestinference is an inference service where you take out one or more unlimited subscriptions — each covering a daily 8-hour time block — entirely via API. During every block you've reserved, usage of all served models is unlimited, for a fixed monthly fee instead of per-token billing. It is drop-in compatible with the OpenAI and Anthropic SDKs.
What does "unlimited-token" mean?
It means that during the 8-hour block (or blocks) you've reserved, your usage is completely unlimited — no tokens to count and no overage charges. The only restriction is that each key runs one request at a time (fair-use concurrency); take out additional subscriptions to run requests in parallel. Within your reserved hours you can send as much as you want.
Is there a flat-rate LLM API with unlimited tokens?
Yes — a flat-rate LLM API is exactly what cheapestinference sells: a fixed monthly subscription instead of per-token billing. Each subscription reserves a daily 8-hour time block with unlimited tokens — no metering, no overage charges — across every model in your pool, from $14.99/month for the Core Pool. Live pricing per pool is on /pools.
What is the cheapest LLM API?
It depends on your volume, so here are both honest answers. Per token, list prices currently run from $0.28 per 1M output tokens (DeepSeek V4 Flash) up to $50 (Claude Fable 5) — we track every major model month over month in the LLM Price Tracker. But past roughly 10M tokens per month, flat-rate wins: cheapestinference sells unlimited-token subscriptions from $14.99/month (Core Pool — DeepSeek V4 Flash and MiMo v2.5, both with ~1M-token context windows), and unlike per-token inference providers the price doesn't move with usage. The exact break-even math is published in the tracker and refreshed monthly.
Which models does cheapestinference support?
Three pools, each with its own model line-up. The Flagship Pool serves Kimi K3 — Moonshot's current flagship — with very limited seats. The Frontier Pool serves Kimi K2.7 (Moonshot), GLM 5.2 (Zhipu / Z.ai), and MiniMax M3 (MiniMax), whose 1M-token context window handles entire codebases and long agent runs. The Core Pool serves DeepSeek V4 Flash and MiMo v2.5 (Xiaomi) — both also with ~1M-token context windows — at a fraction of the Frontier price. Inference is unlimited across every model in your pool — each reserved block gives uncapped usage of the pool's full set. The line-up evolves as new open-weights models ship; the live menu is always on /pools.
Is there an unlimited Kimi K3 API?
Yes. The Flagship Pool serves Kimi K3 — Moonshot's flagship, scoring 57 on the Artificial Analysis Intelligence Index, 4 points off the top model and the narrowest open-vs-closed gap since February — with no token caps during your reserved hours, from $149/month for a daily 8-hour block (all three blocks = 24/7). Model id kimi-k3, OpenAI- and Anthropic-compatible, works in Claude Code, Cline, and Roo Code. Seats are very limited; live availability is on /pools.
Do the models support tool calling and agent workflows?
Yes. Every model on both pools supports function / tool calling, and we verify it against the live API with real agent workloads — it's the same machinery that lets Claude Code, Cline, and Roo Code run on cheapestinference. Send a standard tools array (OpenAI format) or tools blocks (Anthropic format) and you get structured tool calls back. If a tool-calling flow misbehaves in your client, tell us at support@cheapestinference.com.
Are the models quantized?
Not by us — we add no quantization, distillation, or compression of our own on top of any model we serve. What is worth knowing: several of today's leading open-weights models are natively low-precision by design. Kimi K2.7 is trained by Moonshot with native INT4 quantization-aware training, and DeepSeek V4 Flash's official release is FP4+FP8 mixed — in both cases that low-precision format is the reference implementation, the exact weights their makers benchmark and publish. Low precision from the factory is a design choice, not a cut-down copy.
Is a flat-rate subscription slower than a per-token API?
No — the subscription changes billing, not performance. Generation speed depends on the model you pick (size and architecture), and every model supports streaming, so tokens arrive as they are generated. The one thing flat pricing does change: each key runs one request at a time, so you scale throughput with additional subscriptions (combined keys), not by paying more per token.
How do I get more parallel requests?
Take out additional Unlimited subscriptions and combine them into a single API key — a combined key. Your coverage becomes the union of the subscriptions' hours, and where their hours overlap, parallel capacity stacks: two subscriptions over your peak hours means two requests running side by side under one credential. Each subscription keeps its own monthly allowance and renewal, and it works across pools — requests are served by the subscription whose pool offers the model you ask for. Create one from the Keys page by selecting several subscriptions, or via the Management API.
How is this different from OpenRouter, Together AI, or Fireworks?
Those providers charge per token, so your cost rises with usage and can spike unpredictably. cheapestinference charges a fixed monthly fee, so your cost is the same whether you send one request or millions during your reserved hours. We also offer multi-key subscriptions through the Management API, so platforms can provision and manage unlimited subscriptions on behalf of their own customers.
How are you so cheap?
Because we run like a cooperative, not a reseller. Instead of marking up per-token capacity, members pool into shared 8-hour time blocks, and we balance traffic across pools so reserved capacity stays busy instead of sitting idle. Every new pool we open gives us more room to shift load between pools and fill the gaps — the more the network is used, the higher the utilization, and the lower the median cost per request. We pass that efficiency straight back as a flat monthly fee instead of a per-token meter, so during your reserved hours you get unlimited usage at a price per-token billing can't match.
Can I use it for unlimited AI coding — Claude Code, Cursor, Cline, Roo Code?
Yes — coding agents are the main use case. Claude Code works natively: set ANTHROPIC_BASE_URL to https://api.cheapestinference.com/anthropic with your subscriber key (full walkthrough in the Claude Code setup guide). Cursor, Cline, Roo Code, Continue and any client that accepts a custom OpenAI base URL point at https://api.cheapestinference.com/v1. Either way you run models like Kimi K2.7 or GLM 5.2 on a flat monthly fee — unlimited tokens during your reserved hours.
How do I lower my Cline or Roo Code API costs?
Cline and Roo Code bill through whatever per-token API you plug in, so heavy agent sessions get expensive fast. Both accept a custom OpenAI-compatible provider: set the base URL to https://api.cheapestinference.com/v1 with your subscriber key and your coding agent runs on a flat monthly fee — unlimited tokens during your reserved hours, no matter what a long session burns. Current pool prices are on /pools.
Do you offer a GLM 5.2 API subscription?
Yes. GLM 5.2, Z.ai's open-weights flagship, is served in the Frontier Pool: a flat-rate subscription with unlimited-token usage during your reserved daily 8-hour block, instead of per-token billing. Live pricing is on /pools.
How can I get MiniMax M3 API access?
MiniMax M3 — with its 1M-token context window — is served in the Frontier Pool. Subscribe to a daily time block and you get unlimited MiniMax M3 usage during those hours, for a fixed monthly price (/pools).
Is there an unlimited DeepSeek API plan?
Yes. DeepSeek V4 Flash is served in the Core Pool, our most affordable tier, alongside Xiaomi's MiMo v2.5: a flat monthly subscription with unlimited-token usage during your reserved 8-hour block. See live pricing on /pools.
Will there be an unlimited Kimi K3 API subscription?
Yes — we have already secured the capacity to serve Kimi K3, and it goes live the moment Moonshot publishes the weights (promised by July 27, 2026) and the license permits commercial serving. K3 scores 57 on the Artificial Analysis Intelligence Index — above Claude Opus 4.8 — with a ~2.8T-parameter MoE architecture and a 1M-token context window. Depending on how it performs on real workloads it will join an existing pool or debut in a dedicated pool with its own pricing; the announcement lands on /pools. Until then, here's everything we know.
Can I pay with crypto / USDC?
Yes. cheapestinference accepts USDC on Base L2 via any wagmi-compatible wallet (MetaMask, Coinbase Wallet) as well as credit and debit cards through Stripe. Card subscriptions auto-renew each cycle until you cancel; USDC and x402 subscriptions are one-time (30 days, no auto-renewal) and you renew manually.
Is my data private — do you store, share, or train on my prompts?
Privacy is the default, not an upsell. Prompts, completions, attachments, and tool outputs are processed in memory and discarded when the request finishes — there is no prompt database, and nothing you send is ever used to train, fine-tune, or evaluate any model. We never sell, share, or repurpose your content, and you keep full ownership of everything you send and receive. The only thing we keep is usage metadata (token and request counts, cost per key) for billing — never the content. Enterprises that need a formal Data Processing Agreement or written zero-retention terms can request them at privacy@cheapestinference.com.
What compliance and data rights do you support?
We are built to be GDPR-friendly — EU users can request data export or deletion at any time, data is handled in secure, audited infrastructure, and per-key usage is fully isolated: even a platform provisioning keys for its own customers can never see those customers' request content, because nothing is stored. Full details, retention windows, and contact info live in our Privacy & Data Handling policy and Terms.
Still have a question? Email contact@cheapestinference.com or read the documentation.