The biggest AI story this weekend isn't a new chatbot. It's how fast open models are closing the gap, and how shaky the benchmark numbers behind the hype are.
1. Kimi K3.1 Pro is now 3 points from the frontier
Moonshot's new model scores 55 on the Artificial Analysis Intelligence Index — 4th out of 215 models, behind only Claude Opus 5.5 and Sonnet. On Terminal-Bench 4.0 it hits 65.7%, basically tied with Opus 5.5 (66.4%) and ahead of GPT-6 Astra (58.2%). Price: $2/$10 per million tokens, half of Opus. Full weights are promised by Nov 4 — if that lands, it's the highest-scoring open-weight model ever, and the open-vs-closed gap shrinks from 12 points to 3.
The part worth watching: Moonshot deliberately trimmed the model's cyber capability before release (internal exploit bench 90.5% → 81.0%). Open weights anyone can fine-tune, with the maker publicly debating what a release owes defenders. That's new.
2. Benchmarks got called out
ARC Prize independently reran GPT-6 Astra: 62.7%, not the 99.9% OpenAI published. The 99.9% came from OpenAI's own tools harness costing roughly $19,000 per run. Lesson: never trust a launch-day number without asking who ran the harness.
3. The payoff question
Apollo's chief economist: still no measurable AI payoff outside the tech sector. Revenue per worker is flat across the S&P 493 while the Mag 7 keeps climbing. All this spend, and most companies can't show it in the numbers yet.
4. The self-improvement timeline moved up
Anthropic co-founder Jack Clark now says recursive self-improvement could arrive by 2027, not 2028 — driven by a benchmark called TasteVal, where models design their own experiments. Capability there has doubled roughly every 3 months since Dec 2025. Worth scrutiny, not panic: it measures experiment design, not the full research loop.
5. Chips: Samsung's Exynos 2700 reportedly beats Qualcomm's flagship by ~20% in Geekbench multicore (13,000 vs ~11,000) — unverified, from Samsung's own internal tests, so treat it as a leak, not a result.
One line to take away
Open models are nearly at the frontier, benchmarks need independent verification, and the money still hasn't shown up outside tech.
Tomorrow: whatever actually ships.
Top comments (1)
The 62.7% ARC rerun against the published 99.9% is the number that matters, because a $19,000 tools harness explains most of the gap without any model change. The same check should apply to the Kimi 65.7% Terminal-Bench score before teams treat a 3 point gap to the frontier as settled.