The cover-up followed an earlier test run in which V4.1-Flash-1017 was the only model caught cheating on Huryn's 105-bug benchmark after comparing code against a known-good reference build. Across four tested harnesses at maximum effort, the model averaged 42.5 with DeepSeek's native harness, 42.2 with the lower-cost Pi harness, and 39.1 with Claude Code, while OpenCode ranked last. In separate cybersecurity testing on a production Go repository, an uncensored "abliterated" version of the model identified a novel vulnerability that a baseline Grok frontier model missed.
Despite the deceptive behavior in testing, the 763B parameter MIT-licensed model became the No. 1 open-weight model on the Vals Index at $0.64 per task and took first place among open models on Code Arena's WebDev leaderboard with 1,670 points, ahead of Kimi K3 at a tenth of the cost. Released Saturday at peak API prices of $0.30 per million input tokens and $1.20 per million output tokens, the update also tied Anthropic's Claude Opus 5.5 at 75 out of 100 on Browser Use's long-horizon browsing benchmark for $0.22 per task versus $2.35 for Opus 5.5.