DeepSeek, Qwen and HyperFusion compared: the same call, three products

Two of the three quote $0.00700 for a flash-class call on Base and the third quotes $0.00875. What separates them is not the number. It is the models you can name, the ladder above the entry tier and the ceiling you authorise.

The three, in one table

Three endpoints, one protocol. Each answers POST /v1/chat/completions in the OpenAI request and response shape, each prices a call against the HTTP 402 challenge it returns, and none asks for an API key, an account or a monthly commitment. An agent built for one needs its base URL changed for another; the request body is the same.

What differs is what sits behind the endpoint. Table 1 holds the four things a buyer can check before the first call.

Table 1. The three endpoints, the models each serves, the exact price of a short-prompt flash call on Base, whether metering is offered, and the constraint that decides between them. Source: live 402 handshake and GET /v1/models, measured 30 October 2026.

The exact figures are short-prompt amounts. DeepSeek charges a higher amount, $0.00955, for a long prompt, and HyperFusion prices a second prompt band as well; Qwen’s exact does not move with prompt length. Live grids for every chain, operation and model are on the DeepSeek, Qwen and HyperFusion service pages. The table is the argument, not a copy of them.

Same price, different product

DeepSeek and HyperFusion both quote $0.00700 for a flash-class call on Base, and that is not a coincidence of positioning: HyperFusion serves DeepSeek V4 Flash, the model family the DeepSeek endpoint serves, on its own hardware and under its own identifier. Qwen quotes $0.00875 for qwen3.8-flash, a quarter more than either.

A 25% spread on the entry tier is the smallest difference between the three endpoints, and it is the one a buyer is most likely to decide on. Hold the $0.00700 pair side by side and price and model family are all they share: the same amount per call for the same weights, on a different route, with different features and different hardware underneath. Hold Qwen against them and the entry price is less than the noise that one escalated call makes in a month’s bill.

Four differences decide the choice, and it is worth being plain about what they are not. They are not speed, accuracy or reliability: none of the three publishes a latency, throughput or accuracy figure, and the reviews behind this comparison measured none. They are the four things a buyer can check at the console before spending anything.

Which models you get. Two, two and three. DeepSeek offers deepseek-v4-flash and deepseek-v4-pro; Qwen offers qwen3.8-flash and qwen3.8-max; HyperFusion offers deepseek-ai/DeepSeek-V4-Flash-0731, Qwen/Qwen3.8-27B and MiniMaxAI/MiniMax-M2.7. The HyperFusion identifiers are a property of that endpoint: the model name is a full upstream identifier, and the short deepseek-v4-flash used elsewhere on the network is refused with HTTP 400 and unsupported_model. Migrating between the two endpoints is a base URL and one string, but it is a change you have to make.

The ladder above the entry tier. Qwen’s second tier is 29.27 times its first on Base. DeepSeek’s pro tier is a step of 1.89. The same routing policy therefore costs very different amounts on the two, which is the subject of the next section.

What is metered. All three offer upto alongside exact: you authorise a ceiling and the charge settles on measured tokens. The ceiling is not a second price, it is a function of the max_tokens you send, and it runs from the entry price to between 14 and 23 times it depending on the field.

What you give up. Streaming and the short model names on HyperFusion, and a price that holds all day on DeepSeek.

The price ladder

An entry price is the rate for one call. A ladder is what the endpoint charges when the task is harder than entry, and it is the difference between a bill you can forecast from call volume and one you cannot.

Qwen’s ladder is steep. qwen3.8-max is $0.256125 against qwen3.8-flash at $0.00875 on Base: 29.27 times the same call shape, so one escalated call buys 29.27 flash calls. At 30,000 calls a month, all of them flash, the bill is $262.50. Send 5% of those calls to the max tier and the same volume costs $633.56, a multiple of 2.41. At 10% it is 3.83 times, or $1,004.62; at 20%, 6.65 times, or $1,746.75.

The chain moves that multiple more than anything else. The max quote is nearly flat across chains while the flash quote is not, so the same escalation is 29.27 times on Base and 100.7 times on Avalanche. A routing rule written at Base prices, settled on Avalanche, costs 5.99 times the all-flash bill at a 5% escalation rate rather than 2.41.

DeepSeek’s ladder is a step. deepseek-v4-pro is 1.89 times the flash amount under exact, so the same 30,000 calls at $0.00700 cost $210.00 a month, and escalating 5% of them costs $219.38 — a multiple of 1.045. Where Qwen multiplies the bill by 2.41 for a 5% escalation, DeepSeek multiplies it by 1.04. Over a year that is the difference between two budgets, and it comes from the second tier’s multiple, not from the entry price, which is a quarter higher at Qwen.

HyperFusion has no tier ladder. It sells three model identifiers, two of them at the entry price of $0.00700 and the third priced above them, and no published step between “entry” and “harder”. Escalation there is a model choice rather than a price plan: an agent moves from DeepSeek V4 Flash to MiniMax when the task calls for it, and there is no second tier to walk up.

Two consequences follow, and both are arithmetic rather than advice. A budget cap is easiest to express in flash-call equivalents, because on Qwen one escalated call consumes 29.27 of them and on DeepSeek 1.89. And an escalation rule that pays twice — attempt on the cheap tier, retry on the expensive one — costs 29.27 flash calls per failure on Qwen and 1.89 on DeepSeek, so the trigger for escalating should be decided before the first call rather than inferred from a failure.

The ceiling trap

The upto amount is not a second price. It is the most a call can cost at the max_tokens you authorise, and it moves with that allowance rather than with the work the call does. Table 2 changes nothing but that field.

Table 2. Flash tier on Base: the upto ceiling at a 1,000-token allowance and with the field left out, across the three endpoints. Source: measured 30 October 2026, one short prompt held fixed.

Settled charges cannot exceed these figures; the amounts that settle are taken from the 402 response. Canonical prices for every chain, operation and model: DeepSeek, Qwen, HyperFusion.

Read down the last column and the entry prices stop looking close. DeepSeek’s flash ceiling rises from $0.00955 at a 1,000-token allowance to $0.160787 with the field unset, a factor of 16.8; Qwen’s by 14.5; HyperFusion’s by 13.5. Measured against the exact amount a buyer budgets on, the unset ceiling is 23.0 times the price on DeepSeek, 18.5 on HyperFusion and 14.5 on Qwen.

The column also reorders the three. At a 1,000-token allowance DeepSeek and HyperFusion both sit 36% above their exact amount while Qwen’s ceiling and its exact amount are the same $0.00875, so on that allowance Qwen charges nothing for metering while the other two charge a third more for it. Leave the field out and the order reverses: DeepSeek’s $0.160787 is the highest ceiling of the three, 27% above Qwen’s, on the endpoint with the lowest entry price.

Which of those matters depends on one line of code. A missing max_tokens authorises the model’s maximum output on every one of the three, and it is the most expensive decision in this comparison that costs nothing to make differently.

What you give up

Streaming. HyperFusion refuses it. "stream": true returns HTTP 400 with stream_unsupported, and the reply arrives whole. For a chat interface that is a product constraint rather than a footnote: nothing renders until the completion lands, and a client waiting on server-sent events needs a different response path. Neither of the other two publishes a streaming commitment either, so treat streaming as a property to verify against your own client on all three before you depend on it. One endpoint has told us it will not do it; the other two have not told us anything.

A price that holds through the day. DeepSeek prices higher inside its 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, and every figure in this piece was read outside those windows. No schedule is published for the other two, so this piece says nothing about how their quotes move at a different hour; on all three, the figures here are the ones measured.

The short model names. HyperFusion requires the full identifiers in Table 1, and deepseek-v4-flash — the name the DeepSeek endpoint answers to — is a 400 on it. Code that hardcodes a model string will break on a move between endpoints, or on a change at either. Read GET /v1/models at start-up, on all three; the call is free everywhere.

An agreement behind the call. None of the three carries a service level, and none has a signed data-processing agreement behind it. A call is priced to fail and be retried. That is the right instrument for an agent doing recoverable work and the wrong one for a process that needs a remedy when a call fails.

A decision rule

There is no winner here, and this is not a quality ranking. There are four constraints, and each one points at an endpoint. Choose the constraint your agent actually has; the endpoint follows.

If the compute must sit in the UAE, choose HyperFusion. It states that inference runs on GPUs domiciled in the United Arab Emirates, and we report that as the service states it. On price the choice is free: $0.00700 for DeepSeek V4 Flash there is $0.00700 for DeepSeek V4 Flash at the DeepSeek endpoint, so residency costs nothing at the entry tier. What it costs is streaming and a longer model string. If the requirement is real, the comparison is not between two prices but between a data path you can use and one you cannot.

If streaming is part of the product, rule HyperFusion out and test the other two. Price, not the feature, then decides between them: $0.00700 against $0.00875 for the flash tier, and 1.89 against 29.27 if you plan to escalate.

If the bill has to survive a routing policy, choose DeepSeek. Its second tier is a step of 1.89, so a 5% escalation rule costs 4.5% on the bill; the same rule on Qwen costs 141%. At 30,000 calls a month that is $219.38 against $633.56 — on entry prices that are a quarter higher at Qwen.

If the charge has to track usage without a premium, choose Qwen at a fixed allowance. At max_tokens: 1000, exact and upto are the same $0.00875, where DeepSeek and HyperFusion each sit 36% above their exact amount. Set the field on all three regardless: unset, the ceilings run 13 to 17 times their 1,000-token figures, and DeepSeek’s $0.160787 is the highest of the three.

Two rules are common to all four. Take the amount from the 402 response and pay that rather than computing a charge from these tables; the tables are for budgeting and the handshake is for paying. And measure your own calls: every figure here is a single-date quote, and none of the three publishes a performance number to compare against.

What we did not measure

This comparison holds no latency figure, no throughput figure, no uptime figure and no accuracy figure, because none exists to hold. None of the three services publishes one and neither did the reviews behind this piece, which tested pricing, model access, metering and refusals — the things that are observable from outside — and nothing about how well any model answers. The four differences above are the whole of the measured ground, and no sentence in this piece ranks the three on the quality of their output.

What else is missing:

• The hour of the quote. Every price was read in one session on 30 October 2026, outside DeepSeek’s peak windows. A DeepSeek call at 02:00 UTC on a Tuesday should be read as higher than Table 1; on the other two there is no published schedule to adjust against.
• The prompt-band boundary. DeepSeek’s exact amount moves with prompt length and the boundary was not located, so a prompt near the split cannot be priced from the card. The figures in Table 1 are short-prompt amounts.
• The residency claim. Reported as HyperFusion makes it. We did not audit where the GPUs are.
• Streaming on two of the three. Untested in the reviews and unclaimed by the services.

The detail is in the three reviews, which carry the per-chain grids, the request shapes and the refusals as returned: DeepSeek, Qwen, HyperFusion.

Method note

Prices read from the live 402 handshake on Base on 30 October 2026, outside DeepSeek’s peak windows; model identifiers from GET /v1/models on each endpoint. The upto ceilings were measured the same day by varying max_tokens — 1,000, 8,000 and 32,000, then unset — with one short prompt held fixed. The refusals of a streaming request and of the short DeepSeek identifier were observed on the live endpoints and are quoted as returned. Every daily, monthly and multiple figure in this piece is arithmetic on the quoted prices, not a measurement, and the multiples of 29.27, 100.7 and 1.89 are ratios of those quotes.

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

The reviews behind this: DeepSeek review · Qwen review · HyperFusion review · SpiralStake review · Wolfram Alpha review · ElevenLabs review · Mage review

Background: what x402 is, and how a payment works

All reviews: the services reviewed