Gemini 4 Argon undercuts GPT-6 Astra at $1.99 a task — the frontier fight shifts from smartest to cheapest

The claim is narrow and specific: on cost benchmarks, Gemini 4 Argon finishes a task for $1.99, while GPT-6 Astra needs $3.26. That is roughly a 39% discount for the same unit of work — and, at scale, the difference between an agent pipeline that pencils out and one that doesn't.

What makes the metric interesting is that it isn't a price list. Cost per task bundles two things: what the provider charges per token, and how many tokens the model burns to actually finish the job. A model can have expensive tokens and still win on cost by being terse, calling tools efficiently and not looping. So a per-task figure rewards operational discipline, not just an aggressive price sheet. It is closer to what a buyer actually pays than any benchmark score.

The honest caveat: we don't have, from this material, the task mix, the harness that ran it, or — critically — the success rate attached to each price. A cheaper model that fails more often is not cheaper; retries, human review and downstream errors are costs too, just ones that don't show up on the invoice. We've made this point before about raw benchmark numbers, and it applies with equal force here: a figure is uninterpretable without knowing what scaffolding produced it and who ran the measurement. Cost tables are a marketing surface like any other.

Our reading: treat the headline as a direction, not a verdict. The direction is real and it is the most underrated good news in AI — the price of a unit of machine reasoning keeps collapsing, and Google positioning on efficiency rather than on topping a leaderboard is a sign the market is maturing past the capability-only phase. That matters because cheap intelligence, not clever intelligence, is the mechanism that eventually puts diagnostic support, tutoring and legal help in the hands of people who could never pay professional rates. Abundance arrives as a price curve, quietly, long before it arrives as a headline.

Short term, be skeptical of the specific number and ask the question the comparison skips: at $1.99, how often does the task actually get done? Until vendors publish cost and success rate together, cost-per-task is a promising metric being used as a promotional one.