Latent Space burned 20 billion tokens on GPT-6 Astra and measured the running cost at under $6 an hour

Latent Space published a hands-on report on GPT-6 Astra on 3 September, after early access and more than 20 billion tokens of use. The headline figure is a running cost: the newsletter measured Astra at 33 tokens per second against a maximum rate of $50 per million tokens, which works out to under $6 an hour of continuous work.

That arithmetic is the useful part, because it is a rate rather than a total. A price per million tokens tells you nothing about what an hour of an agent costs until you know how fast it produces them.

The cost depends on how many agents it starts

Latent Space is direct about the limit of its own number. It writes that running Astra at the Ultra setting will burn through far more than $6 per hour, because the model is effective at working in parallel. In practice the team says it was often running between 20 and 50 agents at once, all managed by one main Astra agent that started and stopped waves of work and watched its own runs.

So the per-hour figure describes one worker. The bill describes a fleet, and the model decides how large the fleet gets.

The comparison the authors draw

Latent Space frames the work as the kind a junior AI engineer does: watching runs, looking at data, finding problems, fixing them and running again. Against that, it puts hiring someone at $200 to $1,000 a day next to spending about $100 on GPT-6 over two days for the same babysitting.

That comparison is the authors' own, drawn from their own use rather than from a controlled study, and the tasks they chose are theirs.

The benchmark scores in the piece come from OpenAI's launch. Latent Space reports that Astra saturates the hardest versions of FrontierMath at 97.6% and ARC-AGI-3 at 99.9%, and attributes both to OpenAI's own materials rather than to its own testing. It separately points to Artificial Analysis as having independently confirmed that Astra uses fewer tokens than Sol and Fable for the same work, which is the finding that makes the hourly rate hold up.

One stated caveat sits on the latency number: the 33 tokens per second was measured during preview, and the authors note their conclusion assumes that speed survives into general availability.

What this means for a team building software

The number worth taking is the shape of the cost, not the dollar figure. When one agent can decide to become fifty, spending stops being a function of the task and becomes a function of concurrency. A team that budgets by estimating tokens per ticket will be wrong by whatever factor the orchestrator chose that afternoon.

The practical control is a ceiling on how many subagents a run may hold open, set outside the model rather than requested in a prompt. That is a scheduling problem with a known answer, and it is cheaper to solve before the first large invoice than after it.

It is also worth separating what was measured here from what was asserted. The throughput, the token spend and the parallel agent counts come from a team that ran the thing for days. The benchmark percentages come from the company selling it.

Source

GPT-6 Astra: an automated AI Engineer you can hire for <$6 an hour, Latent Space, 3 September 2026.