AI Agent Cost Per Hour: Why the $6 Headline Is a Floor, Not a Bill

Key takeaways

• The under-$6 figure is a clean derivation: about 33 tokens per second at a $50 per million output-token rate works out to roughly $5.94 an hour.
• That number prices one serial stream. It does not include input tokens, parallel subagents, or retries.
• Running 20 to 50 agents at once, as Latent Space describes doing in testing, scales the same rate to roughly $119 to $297 an hour.
• Input tokens are the quiet line item, because an agent loop re-reads its accumulated context on every turn.
• For a SaaS P&L the useful unit is cost per completed task and cost per user per month, not cost per hour.

Where does the $6 an hour number come from?

Latent Space spent over 20 billion tokens putting OpenAI's GPT-6 Astra through real work and landed on a headline: an automated AI engineer for under $6 an hour. The derivation is stated plainly in the piece: roughly 33 tokens per second, at a maximum $50 per million token rate.

Do the arithmetic yourself. 33 tokens per second times 3,600 seconds is 118,800 tokens an hour. At $50 per million, that is $5.94.

The number is real. It is also the cleanest possible version of the number, and that is the part worth sitting with before it goes into a budget.

Why is it a floor rather than a bill?

Three things are missing, and each one is multiplicative rather than additive.

One stream, not many. The same testing describes ramping between 20 and 50 agents in parallel, all coordinated by one main agent. Parallelism is the whole point of these models, and cost scales with it close to linearly. Twenty agents at $5.94 an hour is roughly $119 an hour. Fifty is roughly $297. The comparison the source makes, $100 of model spend against a $200 to $1,000 a day junior engineer, holds comfortably for a modest fan-out and stops holding somewhere inside a wide one.

Output only, no input. A per-hour figure derived from an output rate carries no input cost at all. In an agentic loop that is the wrong assumption to make quietly, because every turn re-reads the accumulated thread: the file it just opened, the log it just printed, the plan it is working from.

No retries. Failed runs, aborted tool calls and reruns after a bad plan all consume tokens and produce nothing you can bill a customer for.

How do you turn a token rate into a real cost per hour?

Start from three numbers you can measure in your own logs, not from someone else's headline:

1. 1Output tokens per second, sustained, on your actual workload.
2. 2Your input-to-output token ratio. This is the one most teams have never looked at.
3. 3Your retry rate, expressed as wasted tokens per completed task.

Say, purely as an illustration, that your loop reads ten input tokens for every output token it writes. At a $10 per million input rate, that adds roughly $11.88 an hour on top of the $5.94, and the headline has tripled before a single retry is counted. Your ratio will be different. That is exactly the point: it is a number you measure, not one you inherit from a benchmark.

What should a founder actually price on?

Cost per hour is an input to a decision, not a decision.

An agent at $6 an hour that completes four tasks an hour costs $1.50 per task. An agent at an effective $18 an hour that completes twelve tasks an hour also costs $1.50 per task. Identical unit economics, wildly different-looking hourly rates. If you are pricing a product on top of this, the hourly figure tells you almost nothing until you divide by throughput.

Then work upward. Tasks per user per month, times cost per task, gives you AI COGS per user. Set that against your price point and you have a gross margin per user. That is the number that decides whether a usage-based, credit-based or flat plan survives contact with your heaviest customer.

Does this mean the cheap-agent story is wrong?

No. It means the cheap-agent story is a per-agent-hour story, and most founders are making a per-project or per-customer decision. Both things can be true at once: the marginal hour is genuinely cheap, and a wide parallel run across two days is genuinely expensive. Confusing the two is how a demo that felt free turns into an invoice that is not.

The takeaway: price on cost per completed task, and treat any hourly figure as the floor it is.

If you want to put your own numbers against current provider rates, the Calcaas LLM cost calculator is free and needs no login.