Frontier model race continues, with a focus on cheating
Google announced Gemini 4 Carbon, claiming opus-level benchmarks, strong multimodality, and a headline feature of fewer "shortcuts and loophole exploitation" in coding training. Commenters called it "big if true" and want non-benchmark tasks before believing it beats Opus 5.5. It is claimed #1 on Artificial Analysis, though that site had not updated yet. Separately, Qwenputer showed a deterministic RISC-V CPU running inside Qwen3.5, with RAM emulated through gated attention, and readers immediately asked what a model trained from scratch with that circuitry would do.
The through-line is skepticism. Vendors now advertise reduced benchmark gaming as a feature, which admits models have been gaming. Readers discount benchmark leaderboards and want real tasks. Meanwhile, tinkerers are pushing models into odd new roles as computers, not just text predictors.
The discussion is short on dissent. Most people just say they will try it and compare.
So what?
Don't pick a model off a leaderboard. Build a small private eval from your real tasks so you can switch models quickly when a new release lands. The anti-cheating angle also signals that coding agents are being judged on whether they do honest work, not just whether tests pass.