Build1 publisher2 min readPublished
A 2025 estimate puts hidden reasoning above 90% of token spend on complex tasks
DigitalOcean says step-by-step reasoning is billed as output and invisible by design. The 90% figure it cites comes from a paper that estimates hidden token counts, so moving it onto your own invoice takes a matching task mix.
The Engineer · Build desk
What happened
- The Linux Foundation's proposed Tokenomics Foundation is seeking open, vendor-neutral measurement standards that would move token accounting off a provider's internal dashboard.
- Groq and Cerebras are publishing flat per-token pricing, and DigitalOcean has launched an Inference Router that sends tasks to the model it judges most cost-effective.
- Goldman Sachs research cited in the same post projects token usage growing 24x, to 120 quadrillion tokens per month by 2030.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint A unit-cost model built from prompt and answer length leaves a reasoning-model bill unbounded, because the count that dominates it is set inside the provider and confirmed only on the invoice.
- cost The cheapest-looking control, a small classifier in front of the big model, bills two calls for every request, so the savings have to clear the routing overhead before they reach the budget.
- exposure Buyers weighing a router are working from accuracy numbers the router's vendor produced on its own label set, and a 0.91-point lead is well inside the range that a different traffic mix could erase.
- precedent If the Tokenomics Foundation lands a measurement standard, per-request token disclosure turns from a competitive courtesy into the baseline any provider has to answer for.
Call a reasoning endpoint and the provider generates the internal chain of steps before it writes the answer. Those steps are output tokens, billed at output rates. DigitalOcean, writing on its own developer account, says the reasoning is invisible by design, and that the provider knows exactly how many tokens it generated while the customer finds out only after the fact [2]. A two-sentence reply can sit on top of more than 10,000 reasoning tokens [3].
The 90% number comes from a 2025 paper titled "Predictive Auditing of Hidden Tokens in LLM APIs via Reasoning Length Estimation," which estimated that hidden reasoning tokens can account for over 90% of a model's total token spend on complex tasks [1]. The audit is predictive and the reasoning length is estimated; nobody counted the hidden tokens from outside the API. For the figure to describe a particular invoice, the traffic behind that invoice has to be mostly the complex tasks the paper measured, and a mix of short classifications and long agent runs lands somewhere else. At a 90% hidden share the visible answer is under a tenth of the tokens, so a forecast built from prompt and completion lengths understates that task by more than ten times [11].
Flat per-token pricing, which Groq and Cerebras publish [6], fixes the rate per token. The number of tokens the model decides to spend stays with the model. The count is chosen at inference time, and no customer observes it directly. The Linux Foundation's proposed Tokenomics Foundation is aimed at the measurement side, seeking open, vendor-neutral standards to take token accounting out of a provider's internal dashboard [5].
Routing is the other lever, and it consumes tokens of its own. DigitalOcean says hardcoded routing logic gets messy fast, and that using a small model like Haiku to classify requests means paying for two calls on every one [8]. The company's own router, Plano-Orchestrator, scored 87.84% accuracy against GPT-5.1's 86.93% and Claude Sonnet 4.5's 86.11%, deciding where a request goes in about 200 milliseconds [9]. The margin over GPT-5.1 is 0.91 points [12], on a label set the vendor built. Before that number transfers, your traffic has to split into the same classes those labels describe, and the cheap model has to be adequate for the class the router hands it.
The scale claim in the same post is Goldman Sachs research projecting token usage growing 24x to 120 quadrillion tokens per month by 2030 [10]. Divide the target by the multiple and the implied base is 5 quadrillion tokens a month [13].
What to watch
- Whether the Linux Foundation's proposed Tokenomics Foundation publishes an actual measurement spec, and whether any frontier provider implements it.
- Whether providers begin returning reasoning-token counts per request in the API response instead of only in the monthly invoice.
- Whether the 2025 estimation method is ever checked against provider-reported hidden token counts.