LLM APIs you can pay per call

What an LLM API does

An LLM API accepts a prompt or conversation and returns generated text. The useful contract includes more than a model name: message roles, input limits, maximum output, streaming, tool calls and structured output determine whether your application can use it. Decide which features your caller requires before comparing a general gateway with a specialized generation route.

A completion endpoint is different from a model catalog, an embedding operation or a credit purchase. A gateway can expose those products on neighboring paths, but their prices buy different outputs. Even a route named chat/completions can represent a prepaid top-up in a provider's listing. The response contract and billing unit decide whether it belongs in this comparison.

The page distinguishes chat-style text, Responses-shaped requests, native Messages requests, independent prompt batches, structured decisions and attested generation. Those categories can overlap when one endpoint returns both text and a provenance record. The same classified paths drive the table and the quoted prices so a cheaper adjacent tool cannot silently become the completion reference.

Compare the returned data

Chat includes general prompt or assistant-reply contracts, including model-specific paths. A video-focused compatible answer route retains its narrower provider-described scope. Native Messages requires an explicit protocol contract; the word messages in a path does not establish it. Batch means generated results for independent prompts, while attestation describes an accompanying provenance record rather than an accuracy guarantee.

Each price is the cheapest endpoint of that service for that kind of data: the same number wherever this page quotes it.

Compare 24 services for this task

Rows prefer a classified chat quote when present. A provider with a different text-generation output keeps the price of its matching task. The listed reference may be a minimum or model-dependent amount; inspect the current challenge for your exact body before treating it as a workload total.

$/CALL is the price of the endpoint named under it: the cheapest chat and prompt text generation endpoint of the service or, for a service without one, its cheapest text generation endpoint. Never the cheapest tool of a gateway. Other kinds of data have their own price in the tables below. TEXT GENERATION ENDPOINTS counts the text generation endpoints out of all the endpoints the service lists. BUYERS 30D counts distinct wallets that paid the whole service on-chain in 30 days, not only its text generation endpoints. "Generalist" marks a gateway that sells more than three unrelated kinds of tools: it is never called the cheapest.

How to choose an LLM endpoint

Match the protocol your client sends

List the request fields and response features used by your client. Test a plain conversation before adding streaming, tools or a JSON schema. An OpenAI-compatible description can cover a useful subset without establishing every feature expected by an SDK. Check content blocks, stop reasons and error shape instead of relying on the provider's label alone.

Evaluate the model on your task

Use a small reviewed set of prompts with explicit acceptance criteria. Include an answerable question, an ambiguous request and a case where the model should state uncertainty. Preserve the model identifier and output settings with the result. A provider's model-name claim is attributed catalog information; the directory does not independently identify the weights that served a response.

Bound input and output

Estimate the longest request your workflow can send, including system instructions, retrieved context and tool definitions. Set an output budget that fits the task. A small reference quote may apply to a minimum request, while a different model or maximum-output reservation changes the payment requirement. Validate the requested limits before asking a payment-enabled client to retry.

Handle streaming and retries deliberately

Decide how your caller represents a partial answer and how it detects completion. A disconnected stream can leave useful text without a final usage record. Before retrying, check whether the first request could have completed and been charged. Preserve request identifiers and outcomes so a network failure does not produce unbounded repeat generation.

Keep provenance separate from correctness

A signed response record can make a particular generation traceable. It does not validate source facts or establish suitability for a decision. When a response uses retrieved evidence, keep the source passages and evaluate the answer against them. Model routing, geographic hosting and privacy claims should remain attributed until your requirements have been independently checked.

Choose by the job you need to complete

Select the required response contract below. The same endpoint can support more than one job when its description establishes both outputs. A structured decision or native protocol should be selected for that contract, rather than inferred from a general chat label.

Generate an assistant reply

Check message roles, output limits, model selection and streaming behavior. An OpenAI-compatible label should be verified against the specific features your client uses. Keep unsupported fields explicit rather than silently dropping a tool definition or structured-output requirement.

Candidates today, most buyers first: Cluster Protocol ($0.003/call, generalist, 719 buyers) · Agent402.tools ($0.01/call, generalist, 276 buyers) · glim.sh ($0.001/call, 156 buyers) · NetIntel ($0.005/call, generalist, 100 buyers) · MapleAI ($0.01/call, 81 buyers) · Rubric Protocol — Post-Quantum AI Attestation ($0.01/call, 52 buyers) · Spraay Gateway ($0.005/call, generalist, 38 buyers) · Losbeto — LLM Gateway & Market Intelligence ($0.005/call, generalist, 32 buyers) · GlianaAI ($0.002/call, generalist, 27 buyers) · Gondola ($0.01/call, 22 buyers) · XFuel ($0.002/call, 21 buyers) · LoboMint Shop ($0.005/call, 9 buyers) · GadgetHumans API ($0.005/call, 8 buyers) · true402 ($0.0001/call, generalist, 7 buyers) · Synap by Linkrra ($0.001/call, 7 buyers) · RedRobotRouges IA Inference ($0.001/call, 2 buyers) · x402vid — AI Intelligence for YouTube Videos ($0.05/call, 2 buyers) · SAGG Gateway ($0.000326/call) · BridgeNode ($0.002/call) · x402 LLM Gateway ($0.01/call) · YiduoChan API Credits ($0.01/call) · GPUOps AI Inference ($0.01/call) · AskClaude ($0.01/call) · Prism Network ($0.012144/call)

Use a Responses-shaped request

Match the documented input shape to your client and check which upstream operation implements it. A translated request can differ from a native Responses service in state, tools and streaming. Preserve those limits in your integration contract.

Candidates today, most buyers first: MapleAI ($0.01/call, 81 buyers)

Use a native Messages request

Require an explicit native protocol description. A route named messages can still accept an OpenAI-style body, so the path alone is not sufficient. Test role mapping, content blocks and stop reasons against your actual client.

Candidates today, most buyers first: Gondola ($0.01/call, 22 buyers)

Generate from independent prompts

Define whether each prompt is independent and how errors are reported per item. Keep item identifiers with returned results. A payments-batch operation from a gateway is not an LLM prompt batch, even when both use the word batch.

Candidates today, most buyers first: Prism Network ($0.012144/call)

Receive a bounded structured decision

Use the provider-defined decision schema only when its fields and allowed values fit your workflow. A specialized decision output should not be treated as an unrestricted chat endpoint. Evaluate an ambiguous case and inspect the evidence behind a recommendation.

Candidates today, most buyers first: MapleAI ($0.001/call, 81 buyers)

Keep generation provenance

Preserve the prompt and response hash, model version and verification reference described by the provider. A signed generation record can help trace an output. It does not establish the factual correctness of the generated answer.

Candidates today, most buyers first: Rubric Protocol — Post-Quantum AI Attestation ($0.01/call, 52 buyers)

What the providers describe

The following notes were checked against complete public endpoint lists and descriptions captured on October 21, 2026. Model identity, compatibility, hosting and billing behavior remain provider declarations. Reference prices come from classified endpoint fields.

Synap by Linkrra

Synap by Linkrra describes a model-routing chat gateway with token-metered usage and a minimum reference. The classified chat quote is $0.001. A wallet top-up and a voice endpoint are separate products, so their amounts do not price this task. Its catalog method and stored request schema conflict; the example section preserves that disagreement instead of presenting a runnable GET completion call.

1 text generation endpoint of 3 · $0.001/call (GET /api/billing/x402/chat/completions) · Base · uptime 30d 100% · 230 ms · 7 buyers in 30 days

Cluster Protocol

Cluster Protocol lists chat completions at $0.003 alongside embeddings, reranking, agent jobs, audio and image routes. The chat quote stays attached to the completion path. A lower embedding reference buys vectors, not an assistant reply, and a managed agent job can involve a different amount of work. Check model, token limits and supported response features for the request you intend to send.

1 text generation endpoint of 9 · $0.003/call (POST /v1/chat/completions) · Base · uptime 30d 100% · 264 ms · 719 buyers in 30 days

MapleAI

MapleAI describes chat at $0.01, a Responses-shaped route at $0.01 and a structured SystemOne decision route at $0.001. The provider calls its Responses support alpha and says it translates to upstream chat. Its model-dependent prompt and maximum-output accounting means the reference is not a universal total for an arbitrary request. Image generation and prepaid code purchases remain outside text-generation prices.

4 text generation endpoints of 8 · $0.01/call (POST /api/v1/chat/completions) · Base/POL/SOL · uptime 30d 99.9% · 195 ms · 81 buyers in 30 days

Gondola

Gondola describes OpenAI-style chat and native Anthropic Messages requests, with chat at $0.01 and Messages at $0.01. It attributes its model supply to Venice. A balance created by an overpayment is prepaid value under the provider's account model; it does not turn subsequent generated text into an unrestricted free tier. Verify the protocol and credit behavior that your client actually uses.

2 text generation endpoints of 3 · $0.01/call (POST /v1/chat/completions) · Base · uptime 30d 99.5% · 381 ms · 22 buyers in 30 days

NetIntel

NetIntel lists model-specific routes and a general chat gateway, with chat references starting at $0.005. Its route named /messages explicitly describes an OpenAI-compatible body and response, so it belongs to chat rather than the native Messages group. The provider says the request's model determines the exact quote. A free model catalog is discovery metadata and does not establish free inference on the completion endpoint.

18 text generation endpoints of 124 · $0.005/call (POST /api/v1/chat/completions) · Base/SOL · uptime 30d 100% · 107 ms · 100 buyers in 30 days

Prism Network

Prism Network describes single-prompt generation at $0.012144 and an independent-prompt batch at $0.012144, with a Merkle receipt for the batch. Its shell-command execution route is a different output and does not define inference price. Check result ordering and per-prompt failures before sending a collection; a receipt can identify a job without establishing that each generated statement is correct.

2 text generation endpoints of 3 · $0.012144/call (GET /inference/v1/inference) · Base · uptime 30d 100% · 223 ms · 0 buyers in 30 days

Rubric Protocol — Post-Quantum AI Attestation

Rubric Protocol — Post-Quantum AI Attestation describes a generated completion with a signed, Hedera-anchored record containing prompt and response hashes, model version and time. The relevant attested-generation quote is $0.01. Other attestation routes certify different records and are excluded from this generation task. Keep the difference between provenance and truth visible when a consumer inspects the signature.

1 text generation endpoint of 10 · $0.01/call (GET /v1/x402/attested-inference) · Base/SOL · uptime 30d 100% · 212 ms · 52 buyers in 30 days

true402

true402 describes chat completions with a reference floor of $0.0001 and an exact-body challenge for the requested generation. That catalog value is useful for identifying the route; it should not be multiplied into a guaranteed price for a long prompt or changed model. Its utility tools return different outputs and cannot set the quote for an assistant reply.

1 text generation endpoint of 27 · $0.0001/call (POST /v1/chat/completions) · Base/SOL · uptime 30d 99.5% · 131 ms · 7 buyers in 30 days

Reference cost and request budgets

The median selected reference is $0.005. At an unchanged request shape, 100,000 reference calls amount to $500.00.

Each service at the price of its main text generation endpoint, the one named in the table above.

Prompt length, output reservation, selected model and routing can affect a provider's exact challenge. Use the volume table as a consistent reference calculation, then estimate your actual request distribution. Compare cost per accepted answer after retries and validation rather than maximizing generated tokens without a quality criterion.

Keep credit top-ups separate from inference consumption in your records. A wallet deposit, an API-key purchase and one completion have different units. A credit balance can fund future requests without making their underlying compute free. When a route describes a floor, obtain the exact request requirement before authorizing its payment.

What the captured 402 establishes

The catalog lists Synap by Linkrra at GET /api/billing/x402/chat/completions, while the stored 402 gives its example input as POST with a JSON body containing model, stream, messages and max_tokens in extensions.bazaar.info.input; its input schema does not restrict the method. The catalog method and the captured example input disagree. The captured message array illustrates a request shape, not a paid assistant response. The catalog reference quote for this method and path is $0.001; it is not a verified quote for a different method or request body.

Inspect the full Payment evidence on the service page and reconcile the current route contract before sending a completion request. This page does not show a runnable GET curl for that POST body or assign its catalog reference to a verified body-specific quote. Preserve the resource, network, amount and input together when evaluating a fresh challenge. A stored 402 does not establish model quality, streaming behavior or successful paid generation.

How to read this comparison

Candidates start from the directory's LLM-inference capability. Endpoint review retains documented generation outputs and excludes inactive or degraded listings, prepaid access sales, embeddings, reranking, image or audio generation and bounded utilities that do not establish a prompt-completion contract. A finance tool that uses a model internally is not automatically a general-purpose LLM endpoint.

The comparison date is October 26, 2026. Median monitoring uptime is 100%; this describes observed service checks rather than a paid model benchmark. The directory does not infer output quality, context length or successful streaming from an HTTP availability observation.

Keep a versioned integration record with the provider, route, protocol, model, input normalization and result validation. Review it when the provider changes an upstream model or request schema. A compatibility alias can be useful, but it should not silently inherit features that its own description does not promise. Re-run the task evaluation after a material routing change rather than assuming a familiar path means the behavior is unchanged.

Questions

How to pay

Each service answers an unpaid call with HTTP 402 and its price; your agent signs a USDC payment on the listed network and retries. No API key, no account, no subscription. See what x402 is.