Every team building agentic products faces the same tradeoff: the models smart enough to run a task end to end are the ones that make the monthly bill flinch-worthy. Agnes 2.5 Pro Alpha ends that choice — the first Agnes AI text model on Artificial Analysis, the independent leaderboard, where its image and video models already rank top ten. Agnes AI is based in Singapore and trains its models in-house.
The model delivers frontier capability without the frontier bill, offering agentic performance at significantly lower cost.
On the Intelligence Index v4.1, which blends nine evaluations, it scored 39 — the standout in its cohort, ahead of Cohere's Command A+ (23) and Mistral's Devstral 2 (19).
It scores 67% on Terminal-Bench v2.1, running commands and working a task end to end, ahead of DeepSeek V4 Pro, and 88% on GPQA Diamond, search-resistant graduate-level science, near the leading 94%. Write the code, drive the machine, think through the problem: that's the whole job.
On GDPval-AA v2, it ran the average task for about three and a half cents. Ten thousand tasks a month is $342 — versus $12,300 on GPT-5.6 Sol and $37,000 on Claude Fable 5. Two things drive the gap: a low price ($0.45 per million input tokens, $0.90 output), and restraint — 26 turns and 38,000 tokens where some models burn three to four times as many.
"Most developers are priced out of the best AI — not for lack of skill, but because the meter never stops," said Bruce Yang, Founder of Agnes AI. "We built Agnes to change that, in public, so anyone can hold us to it."
Agnes 2.5 Pro Alpha is the earliest version to reach Artificial Analysis, not the finished product. It will improve over the coming months, each version measured independently and in public. That is AI parity: frontier capability for the developers, startups, and teams priced out of it. Free API access is where it starts.
About Agnes AI
Agnes AI builds full-modality foundation models across text, image, and video. Its mission is AI parity: frontier-grade AI accessible to developers, startups, and small teams.
