Probably less than the headline token count implies. A 93% cache ratio, Flash-class economics, and a vertically integrated serving stack could turn a spectacular giveaway into a single-digit-million-dollar daily experiment. But the evidence currently points away from Google as the model’s maker.
OpenCode advertised capacity for 100 trillion Ox Alpha tokens a day. Taken literally, that is 1.16 billion tokens every second—more than three times the roughly 31.7 trillion-per-day rate implied by Google’s disclosed 22 billion first-party API tokens per minute.
But those numbers are not comparable measures of compute. One is a capacity claim, the other is observed API traffic, and neither discloses the mix of cached input, fresh prefill, visible output, or hidden reasoning. The mix is the model.
CLAIMED CAPACITY100T/dayOpenCode launch claim
OBSERVED TO DATE25TOpenCode model data
INPUT CACHE RATIO93%OpenCode model data
GOOGLE API RATE31.7T/dayderived from 22B/min
That estimate is not a claim about Google’s books. It is a transparent scenario: 100T tokens a day, 93% cached input, 5% fresh input, 2% output, and internal cost equal to 15% of the public Gemini 3.7 Flash price benchmark. Change any assumption below and the result updates immediately.
ADJUSTABLE SCENARIO · NOT COMPANY GUIDANCE
Build your own serving-cost case.
DAILY COST PROXY$2.73M$19.1M for seven days
FRESH INPUT5%derived remainder
BLENDED API PROXY$0.182per million tokens
RETAIL EQUIVALENT$18.2Mper day at published rates
SCENARIO COST$2.73Mper day at 15% of proxy
The calculator uses August 2026 Gemini 3.7 Flash standard prices—$0.75 fresh input, $0.075 cached input, and $3.75 output per million tokens—as a transparent benchmark. The internal-cost percentage is a user-controlled assumption, not a disclosed Google margin.
The 100T headline is mostly a cache story.
OpenCode’s live model page reports 25T recent tokens and a 93% input cache ratio. A coding agent repeatedly resends system instructions, tool schemas, repository context, and prior turns. If the serving layer can reuse the resulting key-value cache, those repeated tokens should not be treated like newly generated output. (OpenCode model data)
This is why dividing 100T by an assumed output speed produces misleading GPU counts. Prefill can be parallelized; cached prefixes can avoid most of that work; autoregressive output is serial and generally more expensive. Without the token ledger, “100T” measures traffic accounting, not physical inference.
Could Google make the economics work?
Yes. Google already describes Flash as its cost-performance workhorse and reports approximately 22B model API tokens per minute. Its current public Flash pricing discounts cached input by 90% relative to fresh input. Those facts do not prove ownership of Ox Alpha, but they demonstrate an existing serving system built for this scale. (Alphabet Q2 2026 remarks; Google model pricing)
01
Use fewer active parameters
A smaller Flash-style model or sparse mixture-of-experts route can preserve total capacity while reducing work per token.
02
Specialize for code
Post-training on internal coding trajectories can raise task success without paying frontier-generalist compute on every request.
03
Exploit repeated context
Repository prompts are unusually cacheable. Prefix reuse, prompt routing, and speculative decoding attack the cost that the token counter hides.
04
Own the stack
A provider that controls accelerators, compiler, model, scheduler, and data center can optimize across boundaries an API buyer cannot.
The Google theory is not the evidence-backed base case.
OpenRouter says only that an anonymous third party operates Ox Alpha. A public black-box investigation now reports a 44-for-44 tokenizer match to the GLM-5 generation plus compatible serving errors and reasoning controls. That makes Z.ai / GLM the stronger current attribution, while the exact checkpoint, quantization, and deployment variant remain unproven. (OpenRouter model listing; reproducible attribution evidence)
The scenario still matters. If a provider can expose 100T tokens of daily capacity for free—whether Google, Z.ai, or someone else—the market signal is the same: optimized inference can be deployed as customer acquisition, model evaluation, or competitive pressure before slower adopters have found enough valuable work to absorb it.
THE LINK TO THE GLUT THESISWhen cached and specialized tokens get cheap enough, supply can be marketed before demand is monetized.
What would change the estimate?
Highest-value missing facts
Unknown
Why it matters
Direction
Actual utilization
Capacity is not delivered traffic.
Lower usage reduces cost linearly.
Output + reasoning share
Decode tokens are the expensive portion.
More output raises cost sharply.
Active parameters
MoE and distillation change compute per new token.
Fewer active weights lower cost.
Cache accounting
Reported tokens may include reused prefixes.
More real reuse lowers physical work.
Source cutoff: August 24, 2026. The 100T/day figure is a provider capacity claim, not an audited utilization measure. Cost outputs are analytical scenarios, not estimates endorsed by Google, OpenCode, OpenRouter, or the anonymous provider.