← Back to feed

Moonshot AI launched Kimi K3.1 Pro on October 24, a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters and a 1-million-token context window, priced at $2 per million input tokens and $10 per million output — exactly half Claude Opus 5.5. Within hours of launch, the release drew scrutiny over allegations of undisclosed paid promotion that ran alongside the model’s genuine benchmark story.

The allegations

X user @elshayib_, a developer who had been publicly positive about Kimi products, published screenshots he described as a promotional brief sent to influencers before launch. The instructions required creators to submit content drafts for approval before posting. The most consequential line: “Please DO NOT put on the paid partnership label.”

Additional talking points told influencers to position K3.1 Pro as “literally Opus 5.5 open-source, but five times cheaper,” and to run scripted cost comparisons — pre-supplied with Kimi’s pricing — as though from their own testing. A separate note said there was “No need” to attach Kimi’s site link or tag its official account.

@elshayib_ says he declined. After he posted about the campaign, someone connected to it reportedly asked him to delete the comment. A Kimi official later privately apologized, he said on October 24.

The document has not been independently authenticated, and Moonshot has not issued a public statement addressing the specific allegations.

Why it matters

X’s policies require paid-partnership labels on qualifying posts. FTC guidance calls for clear disclosure of material connections between endorsers and brands, including payments, free access, or editorial approval before posting. Whether any specific post violated those standards depends on facts that remain unverified — but the alleged brief describes practices that would run afoul of both.

For enterprise teams evaluating K3.1 Pro, the practical concern is whether independent-seeming recommendations they encounter online reflect undisclosed incentives. K3.1 Pro’s weights are also not yet published (Moonshot promises them by November 4), making self-hosting claims unverifiable until then. Descriptions such as “open source” or “five times cheaper than Opus 5.5” need qualification: the model is closed-weight for now, and the actual API price gap to Opus 5.5 is 2x, not 5x.

The actual numbers

Independent evaluation from Artificial Analysis places K3.1 Pro at 55 on its Intelligence Index, fourth overall, behind Claude Opus 5.5 (58). On Terminal-Bench 4.0, AA measured 65.7% for K3.1 Pro versus Opus 5.5’s 66.4% — a 0.7-point gap under different effort settings. Moonshot’s own evaluation reports 67% on the same benchmark.

The jump from Kimi K3 is real. K3’s Terminal-Bench 4.0 score was 13%; K3.1 Pro at 65.7% represents a 52-point improvement. CyberGym-E2E rose from 58% to 95%. GDPval-AA ELO climbed from 1,534 to 1,846, placing it third behind Opus 5.5 (1,866) and Muse Wonder Full (1,857). SWE-Marathon scores 51.7% for K3.1 Pro versus 46.3% for Opus 5.5.

Moonshot also chose to deliberately limit the model’s offensive cyber capability ahead of open-weighting. The company removed cyber-related training data and reinforcement learning environments, reducing its internal cyber benchmark from 90.5% to 81%. The company stated it “aim[s] not to release open weights that would extend the absolute frontier of cyber capability” — a notable voluntary restriction, made as China circulated draft export controls on AI-driven malware technology.

K3.1 Pro Highspeed, running at 2-3x the standard speed at twice the price, is available now. Open weights are promised November 4.

The benchmarks make a legitimate case. The alleged marketing approach creates questions the model did not need.