Every Tests GPT-6.2 Astra and Sol: Early-Year Unsolvable Tasks Now Saturated, Deep Work Still Lags
danshipper · x · 2026-11-18
Every's Vibe Check tested OpenAI's two new models, GPT-6.2 Astra and Sol, and found clear progress with a questionable direction.
- A batch of eval tasks that no model could handle well at the start of 2026 (theme-based training curricula, minute-by-minute show plans, executive exercises) are now basically saturated by the new models and Codex, per head of evals Mike Taylor.
- But that leaves the team "wanting more": models handle bounded tasks efficiently, while the judgment-heavy, follow-through work people most want help with still demands better reasoning.
- The author argues OpenAI is building models "for the masses" — fast and efficient with minimum viable effort — and, like Anthropic during the Opus 5 backlash, needs a course correction. Most users want to do work they're proud of, not sneak assignments through at the last minute.
Related event: Every's Vibe Check Tests GPT-6.2 Astra and Sol(2 posts)→
More from AGI Musings
- Reddit asks: Hinton's predictions from three years ago still hold — what do big labs think? — SuburbanDad18 · 2026-11-18
- 300 Researchers Convene on AI Consciousness in Berkeley as AP and Vatican Declare AI Non-Sentient — ShakeelHashim · 2026-11-18
- Casey Handmer Claims Claude 5.5 Beats All But Top 0.01% of Physicists at 10000x Speed — DanielWhiteson · 2026-11-18
- Davetroy's take: AI-driven vulnerability discovery is forcing mass upgrades — and outages — Dave_it_up · 2026-11-18
- Researcher ranks universities by publication quality, flags Chinese incentive variance — minhnguyen · 2026-11-18
- Anima Anandkumar's SC26 keynote: the next scaling law is AI that understands the physical world — AnimaAnandkumar · 2026-11-18