Log inSign up
user avatar
Ashrith
@devk67
cs/cyberscurity ai ml engineer
Princeton
Joined October 2020
44
Following
1
Followers
RepliesMedia
  • user avatar
    Ashrith
    @devk67
    Jul 15
    I love 5.6 Sol's willingness to get things done. 5.5 was good, but there would be plenty of back and forth. Now with 5.6 Sol, I spend 5-10 minutes on a plan, then Sol rips away for 10+ hours at times.
  • user avatar
    Ashrith
    @devk67
    Apr 20
    Cannot fucking believe that the "new" AI model does worse than the old. What is anthropic doing? Same AI Lab that created Mythos btw.
    user avatar
    Bridgebench
    @bridgebench
    Apr 20
    Claude Opus 4.7 just regressed hard on BridgeBench. Bullshit Benchmark tests if models push back on nonsense or just make things up. Claude Opus 4.6: 95.0. Rank 1. Claude Opus 4.7: 75.5. Rank 5. Claude Opus 4.7 accepts made up jargon 24% of the time. Claude Opus 4.6 accepts
    BridgeBench BS Benchmark leaderboard, updated April 20, 2026. Tests whether AI models push back on nonsensical premises or confidently invent answers across 100 tasks in finance, legal, medical, physics, and software. Rankings: 1. Claude Opus 4.6 — 95.0 score, 93% pushback, 3% accepted. 2. Claude Sonnet 4.6 — 91.5, 88%, 5%. 3. GPT-5.4 — 91.5, 87%, 4%. 4. Grok 4.20 — 82.5, 80%, 15%. 5. Claude Opus 4.7 — 75.5, 75% pushback, 24% accepted. 6. Grok 4.20 Reasoning — 74.0. 7. GPT-5.4 Mini — 78.5. 8. Kimi K2.5 — 65.5. 9. GLM 5V Turbo — 65.5. 10. Gemini 3.1 Pro — 66.5. Claude Opus 4.6 highlighted at rank 1, Claude Opus 4.7 highlighted at rank 5 with a red arrow showing the drop.