← Back to feed

Third Moment Research runs 10,000 AI agents simultaneously for sessions lasting more than 36 hours. The firm evaluates models across mathematical optimization, machine learning systems, and quantitative research, testing frontier releases at their highest available reasoning settings.

The operational stack that emerged from that testing tells a clear story about how serious agentic users are making model selection decisions in 2026: performance matters, but token cost is now the primary variable.

The Model Stack

Co-founder Jeffrey Yan puts the tradeoff plainly. “GPT-6.1 Sol is currently our most-used model given its token efficiency. It completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra.”

Astra earns its place for the tasks where raw compute output matters most. On a benchmark that asks agents to reduce the latency of a live scoreboard service while preserving its output — measuring speedup by dividing an expert-crafted solution’s runtime by the model’s solution on identical hardware — GPT-6 Astra delivered the biggest speedup of any model tested.

GPT-5.6 Terra fills a narrower role: targeted checks for specific security weaknesses in the firm’s simulation infrastructure. The pattern matches what several other frontier-scale users report — smaller, cheaper models handle well-scoped tasks while the expensive frontier models reserve their compute for genuinely hard problems.

Third Moment is also testing GPT-6 Luna in current runs.

Token Efficiency as a First-Class Metric

At 10,000 concurrent agents, the cost difference between models compounds fast. If Sol costs one-fifth of Astra per million tokens and delivers comparable output across most tasks, the savings on a 36-hour run are not marginal. They determine whether the scale of parallel exploration is economically viable at all.

Yan’s framing — “fewest tokens while delivering performance similar to” — is a description of task completion efficiency, not raw benchmark scores. Sol matched Astra on the DeepSWE v1.1 engineering benchmark, and for the class of quantitative research tasks Third Moment runs, Sol’s token-per-unit-output ratio makes it the better economic choice.

This is the tension that frontier model releases are increasingly navigating. Peak benchmark performance drives press coverage and Arena rankings. But the users running production agentic systems at scale — the ones spending real money on API calls — are optimizing for cost-normalized output across the distribution of tasks they actually run, not the tails that benchmarks oversample.

What Comes Next

Third Moment’s roadmap calls for expanding from a handful of parallel agent lines to hundreds of thousands of collaborating subagents capable of autonomously improving the firm’s own research pipelines. That ambition requires the infrastructure question to be solved at scale: reliable context handoff, coordination without token overhead, and the ability to spin up and terminate agent runs in response to live market signals.

In quantitative finance, the value of finding a durable signal justifies significant compute cost. The firms that get agentic infrastructure right early will run more experiments per unit time than competitors using human researchers and smaller model fleets. Third Moment is betting that 10,000 agents running for 36 hours is only the starting point.