Flat Rate LLM API: What “Unlimited” Actually Means, and When It Pays

A flat rate LLM API sounds like the answer to agent bills you cannot predict: one monthly price, no per-token meter. We read the pricing pages of four providers that sell something like it on 2026-10-03. Every plan we checked has a cap of some kind: requests, concurrency, credits or usage terms. The one plan advertising unlimited tokens, Featherless Chat at $25 a month, says in its own FAQ that it does not cover app or API traffic. Below are the real limits, and a break-even calculation built from token counts we measured, not guessed.

Search interest in a flat rate LLM API makes sense. Agent sessions run thousands of calls, and a per-token bill is hard to forecast. The big labs do not sell one. OpenAI, Anthropic and Google bill their APIs per token, and their monthly subscriptions are for their own apps. So the flat plans come from smaller inference providers. We read their pricing pages directly. Every figure in the next section is as it appeared on 2026-10-03.

Four providers, read on 2026-10-03

Only two of these are flat in the sense people mean: a fixed price with no per-token meter. Those are Featherless Chat and Synthetic. Chutes is a subscription with a quota and then a discount. Ollama's plans are prepaid credits with a bonus. Featherless Developer is per-token billing sold in monthly units. That is a useful finding in itself. Most “subscription” LLM APIs are prepaid metering.

The unlimited plan is not for APIs

Featherless Chat is the only plan above that advertises unlimited tokens. Its pricing page says the plan is limited to human-typed interactive chat. Its FAQ is more specific. Under what is not covered, it lists app or API traffic and reselling, among other uses. It names SillyTavern, RisuAI and Wyvern as the frontends it works with. So if you want to point a coding agent, a backend job or a product at a flat-rate endpoint, the unlimited plan is the one you cannot use.

That leaves Synthetic's pack as the only plan in our list sold for API traffic at a fixed price. Its cap is easy to work out. 500 requests per 5 hours is at most 2,400 requests a day, or 72,000 in 30 days if you used every window around the clock. Nobody does, so treat 72,000 as a ceiling, not an estimate.

Break-even from measured tokens

Whether $30 for up to 72,000 requests is a bargain depends on how much one request costs per-token. Most break-even calculators guess at tokens per request. We have measured values. Every model on our benchmark runs the same nine coding tasks, and we log the tokens each one uses. Multiply by today's list price and you get a real per-request cost for a short coding request:

For short requests the answer depends on the model. On Kimi K3, a $30 flat plan pays for itself after about 5,300 requests, well under the 72,000 ceiling. On gpt-oss-120b, $30 of per-token usage buys 558,047 requests, almost eight times the plan's maximum. A flat plan for a cheap open model is paying for convenience. DeepSeek V4.1 Flash sits in between. Its list price has doubled since we benchmarked it in September, when it was priced at $0.15 / $0.60 (2026-09-15), and at 41,470 requests the flat plan only wins for heavy users. Prices move like this all the time, which is why we date every price.

Agent-sized requests change the answer

Our benchmark requests are small, a prompt of a hundred or so tokens. An agent turn is not. It resends the conversation and tool output, and the input grows into the tens of thousands of tokens. So here is the same arithmetic for a hypothetical agent turn of 30,000 input and 1,500 output tokens at the same list prices. These turn sizes are an assumption, not a measurement. Substitute your own.

Now the request-count plan looks very different. A request cap does not care how big each request is. If your agent makes a few hundred large turns a month on a frontier-priced open model, a $30 request-capped plan beats per-token pricing easily. This is the same reason coding subscriptions exist. Per-token pricing still includes prompt caching discounts, which this simple arithmetic ignores. Our prompt caching guide shows how much that narrows the gap.

The real ceiling is concurrency

Every plan above limits how many requests can run at once. Featherless says it plainly: concurrency is the only ceiling. Synthetic allows one concurrent request per model per pack, Ollama allows 1, 3 or 10 depending on plan, and Featherless Chat allows 4 units. For a chat window that is fine. For an agent that fans out parallel tool calls or sub-agents, one slot per model turns parallel work into a queue. Ollama's page says requests over the limit are queued and rejected once the queue is full. Before you buy, check whether your workload is limited by tokens or by parallelism. A flat plan only fixes the first.

Coding plans are a different product

The model labs' own flat-fee offers are coding plans: a subscription priced for use inside coding tools. Examples are Z.ai's GLM Coding Plan, whose docs say it starts at $18 a month for Lite, and Kimi's membership tiers, from $19 to $199 a month. These plans are a good deal for their intended use, and our pages on them include break-even estimates. Check each plan's terms before you wire it into a product backend, because these plans are built for tool use, not general API traffic. Claude's subscriptions have the same split, covered in our Claude Code weekly limit guide.

How to decide

1. Get your real per-request cost. Take a week of logs, divide the bill by the number of requests, and compare it with the table above.
2. Count requests, not tokens. A request-capped plan rewards big requests and punishes chatty ones.
3. Check concurrency against your peak, not your average.
4. Read the usage terms. Unlimited plans may exclude API or automated traffic, as Featherless Chat does.
5. Price the cheapest capable model per-token first. If gpt-oss-120b or DeepSeek V4.1 Flash does the job, per-token is already close to flat. Our cheapest LLM API guide ranks the options, and the free LLM API guide covers the $0 tier.

Where the numbers come from

Plan prices and limits come from the providers' own pricing pages, fetched on 2026-10-03: featherless.ai/pricing, synthetic.new/pricing, chutes.ai/pricing and ollama.com/pricing. We quote only what was on the page. Tokens per request are the totals our benchmark measured across its nine tasks, divided by nine. They were measured through OpenRouter, not the DataLLM Lab gateway. Kimi K3 was measured on 2026-07-30, GLM-5.3 on 2026-08-22, Qwen3.8 Max 0902 and DeepSeek V4.1 Flash on 2026-09-16, and gpt-oss-120b on 2026-07-30. Prices are OpenRouter list prices on 2026-10-03. All costs here are derived from those two inputs, not taken from a bill. The agent-turn table uses an assumed turn size, as stated.

What this cannot tell you

• Plan quality. We did not buy these plans or test their speed, uptime or model versions. This is a reading of their terms against our token data.
• Chutes' quota. Its pricing page did not state the daily request numbers, so we cannot calculate a break-even for it.
• Your request size. Our measured requests are short coding prompts. Use your own logs.
• Changes. Flat-rate terms change more often than per-token prices. Re-read the page before you commit.

Evidence: dated measurements and catalogue fields; method, latency statistics and limitations. Measurements describe these runs, not all providers or future versions.

Primary sources, checked October 3, 2026: Featherless pricing; Synthetic pricing; Chutes pricing; Ollama pricing. Dated measurements above may differ from the current documentation.