China Model API Relay Market Monthly · 2026-9 (first issue)

This Issue's Judgements

This is the first issue of the Sinan monthly report. The data covers only two weeks (September 2 to 15). We do not pretend it represents a month, but a two-week cross-section is already enough to see the structure of this market. The seven points below are the judgements we draw from 1803 relay sites and 52,000 effective-price quotes; each is supported by data later in the text.

1. The essence of the relay market is a "discount channel for overseas closed models", not a "model supermarket". For overseas closed models (GPT, Claude, Gemini, Grok), the market median effective price at relay sites is only 15% of the official price; for Chinese-vendor models (DeepSeek, Qwen, GLM, Kimi, MiniMax), the median is 97% of the official price, and half are above it. For Chinese-vendor models, relay sites offer no price advantage.
2. Price is not the competitive variable in this market; the top-up rate is. Of the 353 sites with public quotes, half charge "1 yuan RMB for 1 US dollar of credit", which uniformly sells overseas flagship models at 15% of the official price; another 30% charge close to 1:1 against the US dollar, selling at around 100% of the official price. The price distribution is bimodal, with almost nobody in the middle.
3. 48% of flagship quotes are below the "cost floor". Under the assumption of "compliant procurement from the vendor with an enterprise discount of no more than 50%", these prices do not add up. We do not speculate about any individual site's upstream, but below we list three mechanisms that could explain it, and the ways to verify them.
4. There was no real price trend in the two weeks. The chained index level of the price index went from 100 to 100.9, essentially flat; the jump in the "discount ratio" from 50% to 77% was entirely caused by a change in sample composition, not by price cuts or increases. Any issue of this report that looks only at snapshot figures will reach the wrong conclusion.
5. The premium on Chinese-vendor models at relay sites is converging. Among price changes involving Chinese-vendor models during the period, there were 33 cuts and 12 increases; for overseas models, 24 cuts and 19 increases. Relay prices for Chinese-vendor models are moving back toward the official price. This is the hypothesis most worth tracking from this issue.
6. Subscription resale (the Sub2API family) now splits the market evenly with pay-as-you-go. 776 sites use the Sub2API panel to sell subscription seats for coding agents such as Claude Code and Codex, with an open-registration rate of 85% and prices hidden behind login. This is the biggest structural change in this market in 2026, and also the hardest part to measure.
7. The image and video market reproduces the same dividing line. Veo, Sora and Grok video are below the official price at relay sites; Wan, Kling and Vidu are above it; images likewise. Overseas cheap, Chinese-vendor expensive, in the same direction as text models.

I. Scale and Structure: Who Is Selling, and How

Two panels, two businesses. The technical base of relay sites comes in only two major families. The one-api family (including branches such as new-api) has 919 sites and bills by usage: the panel lists a nominal price in US dollars, and users buy credit in RMB at a "top-up rate". The Sub2API family has 776 sites and sells subscription seats: it splits the monthly subscriptions of coding agents such as Claude Code and Codex and resells them, with package prices placed behind login and not made public. The remaining roughly 100 sites are self-built aggregators and sites that expose only an OpenAI-compatible interface.

The rise of the Sub2API family is the most important structural fact of this issue. It means the commodity in this market is shifting from "tokens" to "seats": buyers pay not per million tokens, but per month, for a given concurrency and rate limit. The open-registration rate of these sites is 85% (659/776), higher than the one-api family (about 78% among sites already classified), which suggests they are still in an expansion phase, competing for new users. At present we can only measure their uptime and registration, not their prices; next issue we will read package prices with registered accounts.

Registration status is this market's "open for business" sign. Of the 1803 sites, 1062 have open registration, 201 are closed or invitation-only, and 491 we have not yet classified (mostly sites in the one-api family whose registration interface is non-standard). Closed registration has two entirely different meanings: one is "full, serving existing users only", the other is "preparing to shut down". We distinguish them by uptime and quote changes over several consecutive days, and draw no conclusion for now.

Login methods reveal the community roots of this market. 166 sites support login with a linux.do account (124 in the one-api family, 42 in the Sub2API family), the most concentrated third-party login apart from email and GitHub (332). linux.do is where this market's users gather, and the traffic source for cold-starting new sites. GitHub login on 332 sites indicates that the buyers are mainly developers. WeChat login exists on only 9 sites; this market is deliberately keeping its distance from China's mainstream identity systems.

Domain choice reflects operators' expectations about "the long term". .com 530, .top 177, .ai 151, .cc 143, .xyz 104, and .cn only 92. Domains such as .top, .cc and .xyz, which cost around ten-odd yuan a year, together exceed 80% of the .com count. A low-cost domain does not indicate site quality, but it does indicate the builder's expected lifespan and sunk cost for the site. We have not yet systematically collected ICP filing information this issue and will add it next issue; the fact that only 92 sites use a .cn domain already shows that the overwhelming majority of sites have not designed themselves around compliant operation within mainland China.

Sources of inclusion. Of the 1803 sites this issue, 1387 came from our bulk crawl of public directories, 161 from multi-source cross-referencing, 66 were identified by content fingerprint (pages that simultaneously show model names and prices from four or more vendors), 53 from the two third-party directories aiapirank and hvoyai, 64 from directories and code on GitHub, and 13 from links between sites. Through these channels we also removed 114 tool sites that had been misclassified and 3 vendor official domains. In the next issue these figures will appear as "increments" rather than "stock".

Panel family is determined by public API fingerprint; Sub2API-family plan prices sit behind login, so there is no public quote this issue.

II. Prices: The Three Tiers Are Three Different Markets

First, how to read this chart. Discount ratio = market median effective price ÷ official price. The all-market 77% is the median of the discount ratios across models, and is affected by which models are in the sample; the index level is chained on shared models, unaffected by sample entry and exit, and reflects only the rise and fall of prices themselves. Over the two weeks the index level went from 100 to 100.9; market prices did not move. The discount ratio going from 50% to 77% was caused by the one-time inclusion on September 5 of nearly 600 sites charging at 1:1 against the US dollar. The price changes in these two weeks were almost entirely "we saw more sites", not "sites changed their prices".

Overseas closed models: 15% is a line of convention, not a competitive outcome. The market median effective price for six overseas models, GPT-6 Astra, Claude Opus 5, Claude Fable 5, GPT-5.6 Terra, Grok 4.6 and Claude Sonnet 5, is in every case exactly 15% of the official price. This is not a coincidence: half of relay sites charge 1 yuan RMB for 1 US dollar of credit, the nominal price copies the official US dollar price, and converted at an exchange rate of 6.74 that comes to 14.8%. Of the 303 sites selling overseas flagship or mid-tier models, 81 have an effective price falling exactly between 14% and 15.5%. This price is a tacit understanding among operators, not the result of each one calculating its own costs.

Cost floor: 48% of flagship quotes do not add up. For closed models, any channel must ultimately pay the vendor at API prices, and publicly verifiable enterprise discounts do not exceed 50%. Under this assumption, 48% of the 538 flagship-tier quotes are below the cost floor (under 15% of the official price), and a further 16% are between 15% and 40%; combined, nearly two-thirds are below the level that compliant procurement can explain. We do not speculate about any individual site's upstream, but only three mechanisms can simultaneously explain "price exactly 15%" and "long-term existence" in this market: first, subscription seat pooling, which splits monthly-billed subscriptions such as Claude Code Max into API credit for resale, with a marginal cost close to zero; the rise of the Sub2API family is consistent with this; second, differences in regional pricing and payment channels; third, sources we cannot verify. These three mechanisms are indistinguishable on price alone; only consistency probes and capability sampling can separate them, and that is exactly what we will add next issue (see the note in Part III).

Chinese-vendor models: relay sites are more expensive than official, and the markup follows a pattern. The median discount ratio for 13 Chinese-vendor models at relay sites is 97%, with 5 of them above the official price: DeepSeek V4 Flash 869%, GLM-5.2 166%, DeepSeek V4 Pro 136%, GLM-5.3 Flash 123%, GLM-5.3 121%. This is not site operators being greedy; it is a by-product of the panel mechanism: panels in the one-api family set each model's nominal price as "multiplier × 2 US dollars", the multiplier tables mostly copy OpenAI's price structure, and this is then multiplied by the 1-yuan-per-dollar top-up rate. For a model like DeepSeek whose official price is only 0.094 US dollars, the price computed from this multiplier table lands far above official. The conclusion for buyers is direct: buy Chinese-vendor open models directly from the vendor; relay sites have no price advantage there. The entire price advantage of relay sites lies in overseas closed models.

The premium on Chinese-vendor models is converging. During the period there were 45 price changes involving Chinese-vendor models that were consistent across two consecutive crawls, of which 33 were cuts; overseas models had 43, with rises and falls roughly offsetting. The three models with the largest declines in market median price were Claude Sonnet 5 (−33%), GLM-5.2 (−20%) and Qwen 3.7 Max (−20%), the latter two being Chinese-vendor models returning toward the official price. We list "the relay premium on Chinese-vendor models converging to within 110%" as the first hypothesis to watch next issue.

Cross-site dispersion: the same model varies 1.3x in price, and that is where price comparison pays off. Take, for each model, the spread between the 25th and 75th percentiles of effective prices across sites divided by the median, then take the median across models: the all-market figure is 127%. In other words, even after you have excluded the most expensive and cheapest quarter of sites, among the remaining half the price of the same model still differs by 1.3x. This number is the direct answer to "is it worth spending time comparing prices".

Fast tier: 38% of quotes are significantly above the official price. For models with an official price below 2 US dollars (GPT-5.6 Luna, DeepSeek V4 Flash, MiniMax M3, etc.), the median effective price at relay sites is 91% of the official price, almost no discount; moreover 38% of quotes fall into "significantly above the public price". For these models, what relay sites sell is the convenience of "one account, one bill, many vendors' models", not price.

Discount ratio = relay market median effective price ÷ official reference price; domestic / overseas by the vendor's country of registration.

Band definitions are in section 3 of the definitions page: <15% below cost floor · 15–40% below common bulk discount · 40–80% within range · 80–125% near public price · 125–200% above · >200% well above.

Each site's most common top-up rate; 382 sites with public quotes. 1.0 = ¥1 buys $1 of quota (effective price about 15% of official); 6.8–7.3 ≈ real exchange rate 1:1.

Largest moves in market median this period

III. Uptime and Testing: What We Measured, and What We Did Not

The divergence in uptime is not in "whether it opens", but between 99% and 99.9%. Of the 234 sites that passed the threshold (≥24 probes in 7 days, selling ≥10 models), 87 had 100% uptime over 7 days, 86 were between 99% and 99.9%, and 61 were below 99%. Below 99% means more than 1.7 cumulative hours in a week during which the interface would not open. The three least stable, 32ai.uk (60.8%), bluebirdapi.com (67.5%) and llms.best (86.7%), were "intermittent" rather than "fully offline".

A mistake that must be stated clearly. From 18:00 on September 11 to 02:00 on September 12, our own probe node lost network connectivity for 9 rounds, and all 1801 sites were recorded as 0% uptime in those 9 rounds. This pushed every site's 7-day uptime down to around 93%, emptied the response rankings, and led the first draft of this report to state "0 sites at 100% uptime". On September 15 we added a rule: if all-market uptime in a single round is below 50%, it is judged a failure on our side, and that round is not counted against any site. The corrected figures are those above. The rule is written into Section 7 of the definitions document.

The consistency probe produced no valid samples this issue, and the cause is on our side. The consistency probe uses Keys we register ourselves at each site to make real calls to the model and compare the returned token counts. This issue we sent probes to 13 sites, 88 site×model pairs, and all of them failed because the balance on our 7 Keys was exhausted. Capability sampling had 6 valid results, none below the median for the same model. This means this issue's report has no evidence at all on "which of the three mechanisms described in Part II is at work", only prices. This is the largest gap in this report; we will fill it next issue after topping up.

Data verification: 5 quotes released, 9 still unverified. Of the 9 unverified, 5 are "unit alerts" (the panel text says per second, we parsed it as per call, mainly Grok and Kling video models), 3 are "price outliers" (more than half below the second-cheapest site in the same family, one site only), and 1 is a "gap at the top of the rankings". Unverified quotes do not enter any comparison or ranking.

IV. Image and Video: The Same Dividing Line

Video. 8 model families have an official reference price, with 817 comparable quotes. By "family median effective price ÷ official reference": Sora (OpenAI) 97%, Grok Imagine video (xAI) 232%, Veo (Google) 445%; Wan (Alibaba) 333%, Kling (Kuaishou) 212%, Vidu (Shengshu) 167%, Hailuo (MiniMax) 297%, Seedance (ByteDance) 152%. The numbers make it look as if both Chinese-vendor and overseas models are above official, but the distributions are entirely different: of Veo's 166 comparable quotes, 115 are below the volume-discount line, and of Sora's 32, 20 are below the cost floor; whereas of Wan's 191, 79 are above the public price. In other words, overseas video models at relay sites are "many low prices plus a few high ones", while Chinese-vendor video models are "generally above official", fully consistent with the split in text models.

Image. Nano Banana (Google) 84%, Grok Imagine image 103%, Seedream (ByteDance) 91%, Qwen-Image (Alibaba) 139%, Wan image 260%, Kling image 24%. The Kling image sample is only 25 quotes, 10 of them below the cost floor, so we do not interpret it for now.

Two holes not yet filled. The first is GPT Image: 191 sites sell it, making it the most widely covered image model, but it is priced per token, which is not comparable with a per-image reference, so there is no ratio this issue. The second is the three families HappyHorse, PixVerse and Midjourney, sold by 94 sites combined, which have no official public price per second or per image; we do not use third-party prices as references. These two holes are themselves an observation: the most widely sold image and video products at relay sites happen to be the ones whose prices are hardest to check.

The reference price uses the lowest official tier (resolution / duration); relay quotes are converted to the same tier; families without a public official price get no ratio.

V. The Compute Cost Layer: Official Prices for Chinese-Vendor Models Are Already Below the Single-Card Cost Line

End-of-period single-card hourly prices. H100 SXM: Vast.ai on-demand median 3.00 US dollars, RunPod secure cloud 3.49 US dollars; H200 3.64 / 4.59; A100 80G 0.72 / 1.59; RTX 4090 0.47 / 0.74; RTX 5090 0.54 / 0.99.

Converting this into token cost. Under the assumptions in Section 13 of the definitions document (H100 at high concurrency 3000 tokens/second, 4090 at 800 tokens/second), the cost floor for renting cards yourself to run open models is roughly 0.28 US dollars per million tokens (H100) or 0.16 US dollars (4090). This is only an order of magnitude; change the throughput and the conclusion changes.

This order of magnitude explains one phenomenon from Part II. The official price of DeepSeek V4 Flash is 0.094 US dollars per million tokens, below both cost lines above; GLM-5.3 Flash is officially 0.25 US dollars, also near the line. When a vendor's official price is already below the cost line of single-card rental, relay sites have no room at all to undercut it by deploying themselves; they can only sell at the panel multiplier, and the result is 869% and 123%. Chinese vendors pushing official prices below the compute cost line amounts to squeezing the relay layer out of their own distribution system. This is a vendor's channel-strategy choice; for buyers it means buying directly from the vendor, and for relay sites it means Chinese-vendor models can only be a bundled add-on.

The cost line is a rough estimate under the assumptions in section 13 of the methodology doc (H100 $3.21/h, 3000 tokens/s), for order of magnitude only.

VI. What These Numbers Mean for Different People

We do not endorse any individual site; the following only translates the measurement results into the facts each party needs to face.

Buyers. If you buy overseas closed models, relay sites offer a 5x to 7x price difference, but the closer the price gets to the 15% line, the more you need to ask "where does this price come from", because under the assumption of compliant procurement it does not add up; the consistency probe and capability sampling are checks you can do yourself, and our test page and command-line tool are free. If you buy Chinese-vendor open models, buy directly from the vendor; relay sites have no price advantage and are more expensive most of the time. The same model differs 1.3x across sites, so ten minutes of price comparison is worth it.

Site operators. Prices have already converged; half of sites sit on the same line, and price cannot create separation. Only two things can: verifiable stability (7-day uptime above 99.9%, which only 173 sites achieved this issue) and verifiable consistency. The free verification we offer site operators turns these two into a public mark; the more sites apply, the more "not having applied" becomes a signal in itself.

Model vendors. Chinese vendors' official pricing is already too low for the relay layer to resell, which means the responsibility for reaching downstream channels rests entirely with the vendors themselves; overseas vendors' products have a distribution layer in the Chinese market that they cannot see, priced at 15%, with a scale of 303 of the 1803 sites.

Researchers and media. Please cite the index level rather than the discount-ratio snapshot, and cite the pair "Chinese-vendor 97% / overseas 15%" rather than the single figure "all-market 77%"; the former is stable, the latter moves with the sample. The scale of the Sub2API family is the fact most worth recording from this issue; it does not appear in any public statistics.

VII. Points to Watch Next Issue

1. Whether the premium on Chinese-vendor models at relay sites converges to within 110% (median 97% this issue, 5 models above 120%).
2. After the probe Keys are topped up, whether the consistency probe shows a systematic difference between sites priced at 15% and sites priced at 100%. This is the only way to distinguish the three mechanisms in Part II.
3. Package prices in the Sub2API family: read after login, to build a second price table of "how much per month, how much concurrency".
4. The 30-day survival rate of the nearly 600 sites bulk-included on September 5, as the first estimate of this market's "monthly churn rate".
5. Whether GPT Image's per-token pricing can be aligned with a per-image reference, bringing the most widely covered image model into the ratios.
6. The transmission speed of official price changes: after a vendor adjusts prices, how many days it takes relay sites' median price to follow. No vendor adjusted prices this issue, so it could not be measured.

Definitions and Corrections

All figures come from the site's daily automated measurement: effective price = panel nominal price × top-up rate ÷ exchange rate; discount ratio = median effective price ÷ official reference price; the index level is chained with 2026-09-02 as 100. Chinese-vendor and overseas are divided by the model vendor's place of registration. The cost-floor assumptions are in Section 13 of the definitions document. Three method corrections were made this issue: excluding the rounds when our node lost connectivity (Part III); removing 114 tool sites misclassified as relay sites and 3 vendor official domains; correcting the error of parsing per-second video models as per-call, voiding 144 old rows. The correction record is permanently public on the corrections log page.

Note on finalisation (added 2026-10-07). This issue was due to be finalised on 1 October, but the report script only recomputed the current month and lacked the step that finalises the previous month on the first day of the next; we found this in our own review on 7 October. To avoid mixing October data into September, the final version keeps the last computation of 30 September 04:55 (Beijing time) unchanged (period 2 to 30 September) rather than recomputing it; the GitHub Release attachment published on 1 October has been replaced with the final version. The analysis above was written on 16 September from data for 2 to 15 September, and its judgements were not rewritten at finalisation; the tables under each section give the full-month closing figures. The delay is recorded in the corrections log.

All metric definitions are in Definitions; raw data:2026-09.json、price-index.json、data_v2.json. Figures are recalculated daily during the month, finalized at month end and marked "final"; the analysis text is updated together with the final version.

Cite this page

Data may be freely cited and republished with attribution and a link. We take no money from measured channels, sell no rankings and add no referral parameters, so you can cite us as an independent third-party measurement.

Media, research institutions and financial institutions needing raw data or custom definitions: write to hello@sinanlab.com, see Press & researchers。