Your bill depends on the shape of each request. So does our routing.
TokenIQ is an OpenAI- and Anthropic-compatible gateway in front of two contracted suppliers. Neither is cheapest across the board: on our highest-volume offer one is cheaper on input while the other is cheaper on output, so the correct route depends on each request's input:output ratio, its cache-hit rate and the hour of the day. We price every request on both rate cards and send it to the cheaper one.
You call one endpoint with one key and one offer name. You never pick a supplier, and you never see one.
Portfolio maximum spread
83.8%
Widest gap between our two contracted quotes on a single price line, peak sheet included
Decisions changed by the clock
29.9%
Share of routing decisions that flip because one route publishes a peak sheet
Offers with two live routes
85.7%
Six of seven offers are dual-sourced; one is single-sourced and labelled as such
A 42-turn agent session
$2.3103
Billed off-peak at the standard multiplier, against $4.69 if pinned to the wrong route at peak
The thesis in one session
The same session, four ways
A real coding-agent session: 42 turns, 11,600,000 input tokens of which 88% came from prompt cache, and 265,000 output tokens. Total input:output is 43.8 : 1.
Landed supplier cost in CNY for one 42-turn deepseek-v4-pro session. Route A off-peak ¥9.6310, route B ¥10.1638, route A on its peak sheet ¥19.5317. TokenIQ least-cost pays ¥9.6310 off-peak and ¥10.1638 at peak, because the router changes route when the sheet changes.
| Group | Pinned route A | Pinned route B | TokenIQ least-cost |
|---|---|---|---|
| Off-peak hours | ¥9.63 | ¥10.16 | ¥9.63 |
| Peak hours | ¥19.53 | ¥10.16 | ¥10.16 |
Off-peak
The router picks route A and beats a pinned route B by 5.24%.
At peak
Route A's sheet roughly doubles. The router flips to route B mid-session, with no visible change to the agent, and beats a pinned route A by 47.96%.
What you are billed
¥15.4096 off-peak — $2.3103 — with the saving stated on the request, reconstructible from the stored token counts and the sheet in force at the time.
A short chat turn, same offer
2,400 fresh input and 900 output tokens — 2.67 : 1, no cache. Now route A wins by 8.23 %. The same offer, the same hour, the opposite decision. That is why the router is a service and not a config file.
Real rate cards
Where the spread actually lives
Every figure below is computed from the committed price catalogue, in CNY per million tokens as each supplier quoted it. Split leadership is real: one route leads our economy family on all three lines and our highest-volume offer on input; the other leads everywhere else.
| Offer | Line | Route A ¥/M | Route B ¥/M | Spread | Cheaper |
|---|---|---|---|---|---|
| deepseek-v4.1-flash | input | 0.83 | 0.89 | 7.2% | route A |
| output | 3.32 | 3.58 | 7.2% | route A | |
| cache read | 0.0150 | 0.0179 | 16.2% | route A | |
| deepseek-v4-pro | input | 3.88 | 5.06 | 23.4% | route A |
| output | 11.34 | 10.13 | 10.7% | route B | |
| cache read | 0.12 | 0.0422 | 64.8% | route B | |
| glm-5.3 | input | 7.79 | 5.82 | 25.4% | route B |
| output | 25.68 | 20.35 | 20.7% | route B | |
| cache read | 1.66 | 1.45 | 12.4% | route B | |
| glm-5.3-flash | input | 1.02 | — | — | single-sourced |
| output | 3.13 | — | — | single-sourced | |
| cache read | 0.31 | — | — | single-sourced | |
| kimi-k3 | input | 20.76 | 13.72 | 33.9% | route B |
| output | 102.61 | 68.58 | 33.2% | route B | |
| cache read | 2.06 | 1.37 | 33.4% | route B | |
| qwen3.7-plus | input | 2.32 | 1.27 | 45.4% | route B |
| output | 8.76 | 5.06 | 42.2% | route B | |
| cache read | 0.38 | 0.25 | 33.4% | route B | |
| qwen3.8-max | input | 12.93 | 10.76 | 16.8% | route B |
| output | 38.32 | 32.28 | 15.8% | route B | |
| cache read | 1.41 | 1.35 | 4.6% | route B |
Maximum spread per offer
The discount available on the widest price line, once one route's peak sheet is included. The portfolio maximum is 83.8%.
- deepseek-v4-pro83.8%
- deepseek-v4.1-flash47.4%
- qwen3.7-plus45.4%
- kimi-k333.9%
- glm-5.325.4%
- qwen3.8-max16.8%
The widest price spread between the two contracted routes for each offer, as a percentage of the dearer quote. The portfolio maximum is 83.8% on deepseek-v4-pro cache-read tokens against the peak sheet.
| Item | Value |
|---|---|
| deepseek-v4-pro | 83.8% |
| deepseek-v4.1-flash | 47.4% |
| qwen3.7-plus | 45.4% |
| kimi-k3 | 33.9% |
| glm-5.3 | 25.4% |
| qwen3.8-max | 16.8% |
What the spread is not
A spread is not a discount. Against the best single supplier chosen with hindsight, per-request routing is worth 0.3–4.2 % on realistic traffic mixes. Anyone quoting more is comparing against a deliberately bad baseline.
What routing is worth is 29–31 % of insurance: that is the cost of choosing the wrong supplier for a month, and you cannot know which is wrong in advance. One of our two quotes expires in days.
The clock is the largest single lever. One route publishes a peak sheet at roughly 2× its off-peak rate on our highest-volume family, and 29.9 % of routing decisions change because of it. A gateway that ignores the clock overpays about 2× for half of every working day.
The honest comparison
Why a passthrough marketplace cannot do this
A marketplace forwards your request to a provider and takes a fee on the provider's list price. It has nothing to arbitrage, because the cost it faces is the price you pay. We own two rate cards, so landed cost per request is something we can act on.
| Dimension | OpenRouter — a passthrough marketplace | TokenIQ |
|---|---|---|
| Supplier relationship | Marketplace listing at the provider’s own list price | Two contracted, audited suppliers on negotiated sheets, with prepay discounts at ¥100k and ¥500k |
| Routing objectivewedge | Availability and your stated provider preference | Landed cost per request, subject to the offer’s quality and latency target |
| Cost modelwedge | None per request — the price is the provider’s | Input, output, cache-read and cache-write lines priced separately on both peak and off-peak sheets |
| Ratio awarenesswedge | None. A model has one price | The winner flips at a cache-dependent input:output threshold, computed per request |
| Time of daywedge | Not modelled | One route’s peak sheet runs ~2× off-peak; the clock changed 29.9 % of decisions in simulation |
| Token reductionwedge | Not offered — a fee earner has no reason to shrink the transaction | Semantic cache, prompt compression and a reasoning switch cut the tokens before billing |
| Supply verificationwedge | Trust the provider | Model-substitution detection on every response, with the evidence stored |
| Spend controlwedge | A credit balance | Per-key budget, RPM limit, offer allow-list, month-end forecast, anomaly alert and a hard cap |
| Supplier exposure | Explicit and customer-visible | Never exposed. The offer is the product |
| Catalogue breadth | Hundreds of models — genuinely better here | Seven curated offers. We sell an outcome tier, not a model zoo |
OpenRouter is a good product with a wide catalogue, fast onboarding and a real community. We do not win on breadth and do not try to. We win because owning two rate cards makes landed cost per request something we can act on, and a fee on someone else's list price is not.
Project your own bill
A month of your traffic, priced both ways
Enter your monthly token volume and the shape of your traffic. The projection prices that shape on both contracted rate cards, applies the peak sheet to the share of traffic you set, and takes the cheaper route per sheet — the same arithmetic the router runs per request.
Projected TokenIQ spend
$604.18
per month, at the standard 1.6× retail multiplier
Saving vs the wrong single supplier
$258.02
29.9% lower than pinning the dearer route for the month
Saving vs the best single supplier
$17.95
2.9% — the honest figure, against a supplier chosen correctly with hindsight
Monthly supplier cost by strategy
Landed supplier cost for 3,028,000,000 tokens per month on deepseek-v4-pro at 43.80:1 input to output and 88% cache hit rate, comparing a pinned route A, a pinned route B and TokenIQ per-request least-cost routing.
| Group | Pinned route A | Pinned route B | TokenIQ least-cost |
|---|---|---|---|
| This workload | $539 | $389 | $378 |
| Strategy | Supplier cost | Your bill at 1.6× | Difference |
|---|---|---|---|
| Pinned route A, clock-blind | $538.87 | $862.19 | 29.9% dearer |
| Pinned route B | $388.83 | $622.12 | 2.9% dearer |
| TokenIQ, routed per request | $377.61 | $604.18 | baseline |
Projection only. It prices your stated shape on both contracted rate cards, applies the peak sheet to the share of traffic you set, and takes the cheaper route per sheet — the same arithmetic the router runs per request. Your real bill is metered from the tokens a supplier actually billed, not from this estimate, and cache hits are not billed at all.
What we sell
Seven offers, published pricing
An offer is an outcome tier, not a model name. Each maps to one or two contracted supplier routes that the router chooses between per request.
| Offer | Tier | Context | Input $/M | Output $/M | Cache read $/M | Input ¥/M | Output ¥/M | Routes |
|---|---|---|---|---|---|---|---|---|
| deepseek-v4.1-flashDeepSeek V4.1 Flash | economy | 131,072 | 0.1991 | 0.7964 | 0.00360 | 1.3280 | 5.3120 | 2 |
| deepseek-v4-proDeepSeek V4 Pro | balanced | 131,072 | 0.9307 | 2.4295 | 0.01012 | 6.2080 | 16.2048 | 2 |
| glm-5.3GLM-5.3 | premium | 131,072 | 1.3949 | 4.8823 | 0.34874 | 9.3043 | 32.5649 | 2 |
| glm-5.3-flashGLM-5.3 Flash | economy | 131,072 | 0.2447 | 0.7508 | 0.07436 | 1.6320 | 5.0080 | 1 |
| kimi-k3Kimi K3 | premium | 262,144 | 3.2900 | 16.4498 | 0.32900 | 21.9440 | 109.7200 | 2 |
| qwen3.7-plusQwen3.7 Plus | balanced | 262,144 | 0.3037 | 1.2148 | 0.06074 | 2.0256 | 8.1024 | 2 |
| qwen3.8-maxQwen3.8 Max | premium | 262,144 | 2.5813 | 7.7440 | 0.32267 | 17.2176 | 51.6528 | 2 |
Retail is 1.6× the lowest contracted off-peak rate for that line, converted at 1 USD = 6.67 CNY. The floor is taken per line rather than per route, because an input-heavy call is routed to the cheaper input rate and an output-heavy one to the cheaper output rate. Cache reads are priced but cache hits are not billed at all. Every offer supports tool calls, streaming and a reasoning on/off switch.
A month of mixed traffic
What this is worth across a catalogue
Modelled monthly volumes for a mid-size coding-agent customer with 45% of traffic inside the peak window.
Pinned to one route, clock-blind
¥12,701.83
Landed supplier cost for the month. Picking wrong is the expensive outcome.
Pinned to the other route
¥9,125.25
Better here — but you could not have known that in advance, and one quote expires in days.
TokenIQ, routed per request
¥8,993.99
29.19% under the dearer pin and 1.44% under the better one.
Aggregate arithmetic understates routing, because averaging a month of requests averages away the per-request flips. A simulation over 200,000 synthetic requests with realistic distributions of size, ratio and cache-hit rate puts least-cost 30.90 % under the dearer pin and 1.13 % under the better one across the full catalogue, rising to 4.24 % in a coding-agent-only mix — which is exactly where this traffic concentrates.
Ten minutes, two lines
Developer quickstart
Point your existing SDK at TokenIQ. Change the base URL, change the key, change the model string to an offer. Nothing else moves.
# 1. Create an organisation and its owner. The response carries a session JWT.
curl -s https://api.yunshurouter.com/auth/register \
-H 'content-type: application/json' \
-d '{"org_name":"Acme","email":"owner@acme.test","password":"a-long-enough-pass"}'
# 2. Fund the organisation and mint a key. The plaintext is shown exactly once.
curl -s https://api.yunshurouter.com/billing/topup \
-H "authorization: Bearer $TOKENIQ_JWT" \
-H 'content-type: application/json' \
-d '{"amount":"50"}'
curl -s https://api.yunshurouter.com/keys \
-H "authorization: Bearer $TOKENIQ_JWT" \
-H 'content-type: application/json' \
-d '{"name":"agent-local","rpm_limit":60,"monthly_budget":"25"}'
# 3. Call the gateway. Same request body as OpenAI; the model is a TokenIQ offer.
curl -s https://api.yunshurouter.com/v1/chat/completions \
-H "authorization: Bearer $TOKENIQ_API_KEY" \
-H 'content-type: application/json' \
-d '{
"model": "deepseek-v4-pro",
"messages": [{"role":"user","content":"Summarise this diff in one line."}],
"max_tokens": 256
}'Four wedges, not one: per-request routing, token reduction, verified supply and spend control.
Routing is the visible one. Token reduction cuts the bill rather than shopping it: cache hits are not billed at all. Verified supply means every response is checked against the model we asked for. Spend control means a key has a budget, an RPM limit and an allow-list, so a runaway agent returns an error instead of an invoice.
Prices settle in CNY and are billed in your own currency at 1 USD = 6.67 CNY. The rate used for a request is stored with it.