Prime Intellect is releasing a self-improving Prime Agent that scores…

What was claimed

Prime Intellect is releasing a self-improving Prime Agent that scores 95.5% on ARC-AGI-3 (above human baseline) and the gains are not benchmark-specific

Our verdict

Needs caution

ARC-AGI-3 documentation and announcements state humans score 100% on the benchmark, so 95.5% would be below the published human baseline.

1 of 3 AI systems agree19 sources citedChecked Aug 6, 2026

Check your own claim

Paste any statement, headline, or AI answer — 3 independent AIs verify it in seconds, with sources.

Key findings

95.5% is above the human baseline on ARC-AGI-3.

Incorrect90%

1 AI checked

Prime Intellect is releasing a self-improving Prime Agent.

Verified90%

All 2 AIs agree

The gains are not benchmark-specific (i.e., improvements generalize beyond ARC-AGI-3).

Verified90%

1 of 2 AIs agree·ChatGPT: Can’t verify

Prime Agent’s 95.5% score on ARC-AGI-3 is above the human baseline

Verified95%

2 of 3 AIs agree·ChatGPT: Can’t verify

Verified against 19 sources

Detailed Analysis

All key factual elements of the statement match Prime Intellect’s own announcement and independent coverage. The only subtlety is that the 95.5% figure refers to ARC-AGI-3’s RHAE Best@1 metric, but that is the standard way scores are reported, and the claim is consistent with sources.

Why this verdict

All key factual elements of the statement match Prime Intellect’s own announcement and independent coverage.
The only subtlety is that the 95.5% figure refers to ARC-AGI-3’s RHAE Best@1 metric, but that is the standard way scores are reported, and the claim is consistent with sources.

Claims checked

The gains from Prime Agent are not benchmark-specific

Verified90%

Prime Intellect’s X post states that although Prime Agent scores 95.5% on ARC-AGI-3, “the gain is not benchmark-specific,” and other posts describe improvements across multiple models compared to proprietary harnesses.[2] [5]

Prime Intellect is releasing a self-improving Prime Agent

Verified95%

Prime Intellect’s blog explicitly introduces Prime Agent as “A self-improving RLM agent” and describes its release and capabilities.[1] [4] [5]

Prime Agent scores 95.5% on ARC-AGI-3

Verified95%

Prime Intellect’s technical blog states that Prime Agent with Opus 5 achieves 95.5% RHAE Best@1 on ARC-AGI-3, and this figure is repeated in their social posts and third-party coverage.[1] [2] [3]

What other AI think

Prime Intellect is releasing Prime Agent, a self-improving coding harness that scores 95.5% on ARC-AGI-3, surpassing the human-expert baseline, with gains that are not benchmark-specific. All major claims in the statement are directly supported by official sources from Prime Intellect and recent reporting.

Why this verdict

Prime Intellect is releasing Prime Agent, a self-improving coding harness that scores 95.5% on ARC-AGI-3, surpassing the human-expert baseline, with gains that are not benchmark-specific.
All major claims in the statement are directly supported by official sources from Prime Intellect and recent reporting.

Part of the statement matches public reporting about Prime Intellect releasing agent-training tools. However the specific numeric score and the claim it exceeds the human baseline lack independent confirmation and contradict ARC-AGI-3's published human baseline. The claim that gains are "not benchmark-specific" is unverified.

Why this verdict

Part of the statement matches public reporting about Prime Intellect releasing agent-training tools.
However the specific numeric score and the claim it exceeds the human baseline lack independent confirmation and contradict ARC-AGI-3's published human baseline.
The claim that gains are "not benchmark-specific" is unverified.

Claims checked

95.5% is above the human baseline on ARC-AGI-3.

Incorrect90%

ARC-AGI-3 documentation and announcements state humans score 100% on the benchmark, so 95.5% would be below the published human baseline.

This claim may mislead readers without added context.

The Prime Agent scores 95.5% on ARC-AGI-3.

Can’t verify45%

I could not find any authoritative report, benchmark entry, or ARC-AGI-3 result page attributing a 95.5% score to a "Prime Agent." No primary-source evidence confirms that exact numeric result.

The gains are not benchmark-specific (i.e., improvements generalize beyond ARC-AGI-3).

Can’t verify50%

There is no published evaluation showing the claimed improvements generalize across multiple independent benchmarks; broader generalization claims require evidence from diverse, peer-reviewed evaluations.

Share this result

Try DoubleChecks