The One-Sentence Takeaway
Token collective procurement's price advantage is not an unverifiable low price but the result of four engineering layers: discount coefficients, cache-based billing, model tiering, and channel price comparison. When evaluating a procurement platform, you should compare its effective cost, not its nominal unit price.
First, the Real Price Structure: Discount Coefficient × Cache Pricing
Using official list prices and platform discount coefficients as an example (per million tokens):
| Model tier | Official output price | Platform discount coefficient | Platform output price |
|---|---|---|---|
| Flagship (GPT-5.6 Sol-class) | $30 | 0.1 (10% of list) | $3 |
| Value (GPT-5.6 Terra-class) | $12 | 0.1 (10% of list) | $1.2 |
| Batch (DeepSeek V4 Flash-class) | $1.2 | ~Official price | ≈$1.2 |
On the cache side, the portion that hits the cache is billed separately at the cached input price (cached_input), which can be as low as roughly 10% of the official full price. Nominal discounts solve the "unit price"; cache-based billing solves "repeated content" — the two are orthogonal cost-reduction dimensions, and combined, the effective cost is far lower than either dimension alone.
The Four Layers of Effective Cost
effective cost = nominal unit price × volume
× discount coefficient (supplier negotiation + usage scale)
− cache savings (cache hit rate × cache price gap)
− tiering savings (right-sizing high-spec tasks to matched models)
− price-comparison savings (real-time channel quality monitoring → dispatch to best channel)
Layer 1: Discount Coefficient (Supplier Negotiation + Usage Scale)
Platforms gain bargaining power through first-party supplier aggregation and quality distribution channels, then set discount coefficients per vendor (e.g., 10% of list for the OpenAI family, 20% for Anthropic, 20% for Google). Discounts float with cumulative usage — the larger the scale, the more room for negotiation.
Layer 2: Cache-Based Billing (The Most Underrated Cost Saver)
Coding agents' requests have a highly repetitive prefix structure (system prompt, persona settings, context scaffolding). Once they hit the prefix cache, they are billed at the cached_input rate. Quantitative example: official output $12/M, cache read price $1.2/M, monthly usage 80M tokens, 60% hit rate:
- Without cache-based billing: 80M × $12 = $960
- With cache-based billing (60% hit): 32M × $12 + 48M × $1.2 = $441.6
- The cache layer alone cuts costs by roughly 54%.
Layer 3: Model Tiering (Match Specs to Task Complexity)
Flagship-model capability is reserved for difficult reasoning; routine and batch tasks are routed to models matched to the spec. Quantitative example: monthly usage 100M tokens, official output price flagship $30/M, mid-tier $12/M, batch $1.2/M:
- All flagship: 100M × $30 = $3,000
- Tiered (30% flagship + 40% mid-tier + 30% batch): 30×30 + 40×12 + 30×1.2 = $1,416
- Model tiering alone cuts costs by roughly 53%, and the quality loss is negligible — because high-spec resources are only allocated to high-value tasks.
Layer 4: Channel Price Comparison (Real-Time Dispatch to the Best Channel)
The platform's routing engine monitors each channel's latency (TTFT), stability, and price in real time, dispatching requests to the channel with the best current price-performance ratio; when a single upstream provider raises prices or fails, it switches automatically. This layer addresses the cost and availability risk of vendor lock-in.
Which "Bargains" You Should Not Touch
A low price whose nominal unit price sits significantly below upstream cost and cannot be attributed to any of the engineering layers above usually corresponds to shared quota, abnormal channels, or missing cache capability:
- Shared quota: abnormal concurrency patterns trigger upstream risk control and mass account bans;
- Abnormal channels: quota of unknown origin, with no guarantee of availability;
- Fake low prices that lack cache-based billing: the nominal unit price is low, but with no cached_input billing, repeated content is billed at full price — so the effective cost is actually higher.
The screening formula:
explainable low price = discount coefficient (≤50%) + cache billing (≥30–50% hit rate)
+ model tiering (≥40–50%) + channel price comparison (bounded)
unexplainable low price = none of the above four factors can be attributed,
and the nominal unit price is more than 30% below official → exclude
Common Questions
Q: What determines the cache hit rate?
The stability of the request prefix. By keeping the system prompt, persona settings, and few-shot examples fixed as a stable prefix, you can raise the hit rate. The gateway side automatically recognizes and bills at the cached_input rate.
Q: How do discount coefficients relate to official pricing?
Discount coefficients are determined by supplier negotiation and usage scale (e.g., 10% of list for the OpenAI family) and float with cumulative usage; a healthy price structure is always explainable and auditable.
Q: Does model tiering affect output quality?
It depends on whether the tiering is sensible. Grade by task complexity and value rather than randomly downgrading; high-value tasks are still routed to flagship models.
Q: Can you get cache pricing with a direct official connection?
Yes, but you must precisely control the prefix structure and configure caching yourself, and you would lack tiered routing and channel price comparison. The value of a collective procurement platform is in automating cache-based billing, tiering, and price comparison.
Summary
Token collective procurement's price advantage = the four-layer engineering of discount coefficient + cache-based billing + model tiering + channel price comparison. The cache and tiering layers combined typically cut effective costs by more than 50%. Evaluate platforms by effective cost, and be wary of low prices that cannot be attributed. Sign up for TeamoRouter to verify your effective cost with GPT at 10% of list, Claude at 20% of list, and cached_input cache billing.
