Is 1.9 Billion Tokens in a Day Normal Yet?

ChatGPT Use cases and examples
, , ,
1

Is 1.9 Billion Tokens in a Day Normal Yet?

I ran 1.9 billion tokens in one day yesterday. 100% quota.

For context, I’ve also run about 23 billion tokens through DeepSeek over the last 30 days — roughly 733 million tokens a day on average.

Serious question for people building heavily with agents: is this normal now?

I’m not complaining. I’m genuinely curious what other people’s experience looks like once they’re running multiple agents, reviewers, repairs, and verification loops in parallel.

Yesterday I was working across several things at once:

  • pushing a native iOS memory/capture app through a major chunk of its working product and UI;
  • building the first Rust runtime foundation for a custom harness rebuild;
  • hardening and validating a DSH-forked governed multi-agent engineering harness;
  • extracting the orchestration/governance model from that harness into a portable specification for the rebuild;
  • and doing early product and architecture work on another consumer app concept.

There was also one pretty funny wrinkle: Astra goofed and didn’t read my governance until I was down to about 16% quota.

Before that, it had been doing a bunch of the fanout work as Astra High, which was obviously a much more expensive way to do it.

Once the governance was actually enforced and fanout moved to the intended cheaper model mix, that last 16% still took another 3–4 hours to burn.

That contrast was probably the most interesting part of the day.

It made the token number feel less like “wow, 1.9B” and more like a routing/governance question:

How much are people actually burning when the expensive model is reserved for the work that really needs it, versus accidentally letting it do all the breadth too?

Curious what others are seeing.

1.9B in a day: normal, excessive, or rookie numbers?

Edit: 1.6B DeepSeek usage yesterday after harness stability fixes - so that probably around the new daily average.

2

The most amount I’ve ever burn was 244m in a day

While being semi productive.

I don’t get what your asking your agents to do while being productive but hope it works cuz that seems like insane amount of tokens.

For reference I just got on Pro tho

and I spent like a weekend my hobby projects so my token usage has been on a steady increase since feb when I spent 21 mil for a week, now I spend on avg 7.3 bil tokens per week since I got 5.6 and for sure since I got on pro

My agent was burning too much of my token budget but the advice I got from alpha users was to remove most skills and gov since it could make that up on the get go cheap than it took to read it so I tried that and it actually improved the token efficiency by my estimate by a noticeable amount like 30%

But I never got the day one astra 3D renders other people were getting, a lot of people I’ve talked to say that the 3D rendering stuff in blender got lowered in quality? I’m not sure but I used my budget today to know for sure to know if its worth getting another ai subscription for just 3D stuff

But I’m a total rookie green-horn so don’t take what I say as the peak confidence

3

I’m trying to follow something I heard Sam Altman say in one of his many interviews. Paraphrasing, he was surprised that so many people were going after the low-hanging fruit instead of trying to reshape an entire industry.

That’s basically what I’m attempting, just for a very specific niche: membership organizations, associations, and publishers.

Fernain is the umbrella for that work: https://fernain.com

One thing that probably explains the token number better: my requests to the provider routinely carry six-figure context.

Harry is my custom harness, where most of my DeepSeek usage happens. I also use it as a rough stand-in for understanding my Codex usage, since there isn’t much telemetry exposed there unless we instrument and log it ourselves.

Harry keeps large shared context around across workers, and a very high percentage of the input is cache hits. So 1B+ aggregate tokens can accumulate surprisingly quickly even though the amount of genuinely new input on each request is much smaller.

So the burn is less “one giant coding session” and more multiple products, orchestration, governance, verification, and infrastructure all moving in parallel.

I’m still iterating on the methodology, though. I fully expect the way I’m doing this a year from now to look very different from today.