Breaking

OpenAI's GPT-6.1 Sol Reaches Agent Arena Top Five at $0.56 per Task

October 2, 2026 at 15:48 EDT

Foundation Models
AI Agents
Software Dev & Coding

OpenAI's GPT-6.1 Sol has reached No. 5 on Agent Arena's overall agent leaderboard, posting an 11.23% net-improvement score at a $0.56 median cost per task. The result gives developers an early real-world signal that the new Sol model can approach GPT-6 Astra-level agent performance while materially lowering the cost of routine tool-using work.

Oct. 2, 2026 - OpenAI

Near-flagship agent work at a lower task cost

GPT-6.1 Sol (Max) reached the top five in an agent benchmark with 11.23% net improvement and a $0.56 median task cost, close to GPT-6 Astra's score at far lower cost.

Agent ranking

11.23%

Net improvement

$0.56

Median task cost

6,278

Measured sessions

Median task cost, proportional view

Each stack uses the same dollar-per-task scale.

$0.56

GPT-6.1 Sol

$0.91

GPT-6 Sol

$2.94

GPT-6 Astra

Why it matters

The model gets close to Astra's agent score while making routine coding, document, and workflow agents cheaper to run.

What to test

Latency, quota behavior, and reasoning effort still need workload-specific checks before replacing flagship runs.

The signal is practical work

The score reflects task success, feedback, steerability, bash recovery, and tool behavior, which is why the cost-per-task shift is central to the story.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.