DeepSeek: per-call chat completions over x402

DeepSeek at deepseek.x402.press offers OpenAI-compatible chat completions to agents with no API key and no account. Two models, paid per call, from $0.007 on Base. Built for developers wiring a model into an agent that settles in stablecoins.

What it is

DeepSeek is a hosted endpoint at https://deepseek.x402.press. It answers POST /v1/chat/completions in the OpenAI request and response shape, so an agent built against an OpenAI client needs its base URL changed and a payment step added, and little else.

There is no sign-up, no API key and no monthly bill. A call is paid for with an x402 payment against the HTTP 402 challenge the endpoint returns. The completion comes back on the retried request that carries that payment.

For an operations lead the difference is administrative rather than technical. There are no credentials to issue, store or rotate, and no account balance to reconcile. The agent holds a small stablecoin balance on a supported chain and pays from it as it works. Nothing is bundled with the call: no trial credit, no seat, no dashboard quota. You pay per request and you can stop between requests.

What you get

Two models

deepseek-v4-flash and deepseek-v4-pro. Both accept the same request shape and are chosen per call in the model field. The list is readable at any time from GET /v1/models, which is free. Read that list at start-up rather than hardcoding model names.

The request shape

The endpoint takes the OpenAI chat completions body. model and messages are the parts you always send. The endpoint forwards temperature, max_tokens, top_p, frequency_penalty, presence_penalty and stop to the model, so sampling and stopping behaviour you have already tuned carries over unchanged.

{
  "model": "deepseek-v4-flash",
  "messages": [
    {"role": "system", "content": "Classify each support ticket."},
    {"role": "user", "content": "My invoice shows a charge I do not recognise."}
  ],
  "temperature": 0.2,
  "max_tokens": 200
}

The response comes back in the OpenAI shape, so the parsing code you already have will read it.

The free calls

Three things on this service cost nothing. GET /v1/models lists what the endpoint serves. Recent feedback can be read. Feedback can be submitted against a response_id with a rating from 1 to 5 and an optional comment. None of these is billed.

Feedback is worth wiring up. It records whether a particular response was good, keyed to the response rather than to the request. It is a record of quality, not a support line: it opens no ticket and promises no reply timescale.

What it costs

Every supported chain quotes two amounts for a call, under two schemes. Under exact you pay a fixed prepaid amount for the completion. Under upto you authorise a metered ceiling, and the charge settles on measured input and output tokens. The exact amount is fixed within a prompt band, not regardless of prompt: a short request pays the short-prompt amount and a long one pays the higher amount, which the upto ceiling already authorises. Avalanche quotes an exact amount only.

Table 1. deepseek-v4-flash on Base: the same call at two prompt lengths, same max_tokens. The band’s boundary was not measured. Source: measured 30 October 2026.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

The ceiling is identical in both rows because both rows authorise the same max_tokens; the prompt band moves the exact amount, not the ceiling.

Table 2. deepseek-v4-flash: price per POST /v1/chat/completions in US dollars, by settlement chain and payment scheme. Source: live 402 handshake, 30 October 2026, outside peak windows.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

Table 3. deepseek-v4-pro: the same call, priced per model. Same source.

The upto amount is the ceiling at max_tokens: 1000; it rises with the output allowance you authorise (see below).

The model named in the request sets the price: Table 3 puts deepseek-v4-pro at about 1.9 times the flash price under exact, and about 2.7 times under upto at max_tokens: 1000.

Two things move those amounts. The first is the hour: all of them are higher inside DeepSeek’s peak windows, 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. That is seven hours of each weekday, 35 of the 168 hours in a week, so about one hour in five. The figures above were read outside those windows. The second is the scheme, the prompt band the call falls into, and — under upto — the max_tokens you authorise.

Base is the default. It is cheap, it is the only chain here that takes both USDC and EURC, and the gap with Ethereum mainnet makes the choice worth stating. At $3.00 against $0.00700 under exact, a call on Ethereum mainnet costs roughly 430 times what the same call costs on Base, and the ratio holds under upto at a 1,000-token ceiling: $4.092858 against $0.00955. An agent paying on mainnet has not bought a better completion, only a more expensive one. Avalanche is cheaper again at $0.00200 exact, and sensible if the agent already holds a balance there.

Treat the price tables as reference points, not quotes. The amount the endpoint asks for in the 402 response is the amount you pay. Under upto, take the ceiling from that response rather than from the table.

What the ceiling actually does

The upto amount is not a second price. It is the most the call can cost at the max_tokens you authorise, and it moves with that allowance, not with the work the call does. Table 4 measures it on Base with one short prompt, changing nothing but max_tokens.

Table 4. deepseek-v4-flash and deepseek-v4-pro on Base: the upto ceiling at four output allowances, one short prompt, exact unchanged throughout. Source: measured 30 October 2026.

Set max_tokens deliberately. A call that leaves the field out authorises the model’s maximum output, and the flash ceiling on Base rises from $0.00955 at 1,000 tokens to $0.160787 with the field unset — nearly seventeen times the figure in Table 2 — on a prompt that has not changed. The ceiling is a limit on what a call can cost you, not a forecast of what it will cost.

How an agent uses it

Illustrative worked example. Both examples use the Base prices in Table 2 for deepseek-v4-flash, outside peak windows, and assume short prompts: $0.00700 is the short-prompt amount under exact. A long prompt moves exact to $0.00955, the same figure as the upto ceiling at max_tokens: 1000, so the upper figure in each example also covers a long prompt under exact. Both bounds assume max_tokens: 1000: with max_tokens unset the flash ceiling on Base is $0.160787 a call, and they would not hold. Peak windows, or deepseek-v4-pro, cost more.

Example A: support triage, 800 calls a day

A support agent reads each incoming ticket, classifies it and drafts a first reply. On Base, at the exact price of $0.00700 a call:

• 800 calls a day: 800 × $0.00700 = $5.60 a day
• 24,000 calls a month: 24,000 × $0.00700 = $168.00 a month
• At $0.00955, the long-prompt exact amount or the upto ceiling at max_tokens: 1000: 800 × $0.00955 = $7.64 a day, and 24,000 × $0.00955 = $229.20 a month

So the same 24,000 calls cost between $168.00 and $229.20 a month, depending on the prompt length under exact and on measured usage under upto. On deepseek-v4-pro the exact price gives $318.00 a month at that volume.

The same workload on Avalanche would be $1.60 a day, or $48.00 a month. On Ethereum mainnet it would be $2,400.00 a day, or $72,000.00 a month. The agent does identical work in all three cases.

$168.00 a month is the floor for that agent and $229.20 the ceiling on Base at max_tokens: 1000. Budget against the ceiling for the max_tokens you have set, then compare actual spend with call volume after the first week.

Example B: research, 40 calls per report

A research agent makes 40 calls to produce one report: one to plan, several to summarise sources, several to check its own claims, one to write.

• 40 calls × $0.00700 = $0.28 per report
• 10 reports a week = 400 calls = $2.80 a week
• Four weeks = 1,600 calls = $11.20 a month
• At $0.00955, the long-prompt exact amount or the upto ceiling at max_tokens: 1000: 40 × $0.00955 = $0.382 per report, and 1,600 × $0.00955 = $15.28 a month

On Avalanche the same work costs $0.08 a report, or $3.20 a month. On Ethereum mainnet a single report costs $120.00.

At $0.28 a report the model call is small enough that it should not shape the design. Spend the effort on the prompts and the sources, not on shaving calls. One caution on the arithmetic: under upto the settled charge tracks measured tokens, so a research agent that lets its context grow with every call pays more for its later calls. $11.20 is the cost of forty short calls, not forty calls with a long accumulated context. Under exact the amount is fixed for the band, so a long accumulated context pays the upper figure, but it does not rise with each added token.

Where it stops

Five limits are worth knowing before the first deployment.

The amount is not a constant. It differs by chain, by model, by scheme, by prompt band and, under upto, by the max_tokens you authorise, and it is higher inside the peak windows. A budget built on one call count and one price will be understated.

Not every chain offers both schemes: Avalanche quotes an exact amount only.

No performance figures are published. This service publishes no latency, no throughput and no uptime figure, and no commitment to streaming. This article asserts none of them. If those properties decide whether your agent is usable, measure them against your own traffic before you depend on them, and expect to measure again after any change in load.

Model availability can change. The models offered are read from GET /v1/models for exactly that reason, and an agent that hardcodes a model name will break when that name goes away.

This is a pay-per-call endpoint, not a contracted service level. There is no SLA attached to a call. It is priced for a call that can fail and be retried, and it should be built into a system that can absorb that.

When not to use it

Three cases where a different instrument is the right one.

You need a contractual uptime commitment. If your process requires a service level with defined measurement, credits or remedies, this endpoint carries none. Use it where a failed call is a retry, not where a failed call is an incident with a contractual consequence.

You need data processing terms signed by a vendor. There is no account and no contract behind a call. If a reviewer, a client or a regulator expects a signed agreement governing the data you send, this endpoint will not satisfy that expectation, and it should not be used for data that requires one.

Your volume is high enough that a committed contract is the right instrument. Per-call pricing is the correct tool at $11 or $168 a month. At a volume where the monthly bill becomes a line someone negotiates, the usual instrument is a committed rate with agreed terms. We do not publish such a contract for this endpoint. If you need one, this endpoint is not the way to obtain it.

How to start

Two requests, with one payment in between.

1. Call GET /v1/models. It is free. It returns the model names the endpoint serves right now, and it confirms that your base URL and network path are correct before any money is involved.
2. Send POST /v1/chat/completions with a model and a messages array. The first attempt returns HTTP 402, with the amount due and the payment details for the chain you are using. On chains that offer metered pricing the response quotes a ceiling alongside the exact amount; on Avalanche there is an exact amount only.
3. Pay the quoted amount in USDC, or in EURC on Base, from a wallet that holds a small balance, then present the payment. The completion comes back on the retried request that carries it, not on the first one.

That is the whole of the handshake: ask, be told the price, pay, receive. The agent needs a wallet and a float of a few dollars in stablecoin, which at $0.00700 a call is several hundred calls. There is nothing to register and nothing to rotate.

One point is worth enforcing in code. Take the amount from the 402 response and pay that. Do not compute the charge from Table 2 and treat any difference as a fault. The tables are for budgeting; the handshake is for paying.

Method note

Prices read from the live 402 handshake on 30 October 2026, outside peak windows; the catalogue figure differs and is not used. The upto ceilings were measured the same day by varying max_tokens — 1,000, 8,000 and 32,000, then unset — on Base with the prompt held fixed.

Filed under Service Reviews · Tags: x402, inference, deepseek, per-call-billing

Related reviews: Qwen review · HyperFusion review · SpiralStake review · Wolfram Alpha review · ElevenLabs review · Mage review

Background: what x402 is, and how a payment works

All seven reviews: the services reviewed

Marketplace: DeepSeek on x402.press — prices per chain, operations and the metered ceiling.