Evan Cloutier
· SemiAnalysis
· 21 августа 2026, 16:40
· ⏱ 11 мин чтения
| Читать в Substack ↗
Резюме
Open-source models are catching up to proprietary frontier models faster in each era, with the catch-up window roughly halving from the early scaling era to the agentic era, but the author argues this is less bearish for frontier labs than it looks because durable value is shifting to the full model-plus-harness product. Market implications: pure model-layer margins face pressure, while inference demand and integrated AI product players benefit.
•The author measures the open vs closed gap across three eras — early scaling, reasoning, and agentic — and finds open models take roughly half as long to catch up with each era.
•In Era 1, GPT-3.5 Turbo scored 75.7 vs Llama-2-70B's 39.9 on a normalized composite; Llama-3.1-405B closed the GPT-3.5 Turbo gap by July 2024, and DeepSeek V3 nearly matched GPT-4o (94.1 vs 95.5) by December 2024.
•In Era 2, the o1-era reasoning gap opened at only 12.1 points, largely because DeepSeek R1 existed, versus a 35.8-point gap at the start of Era 1; R1-0528 closed that gap in 8.5 months by May 2025.
•Anthropic's Opus 4.5 is called the unofficial start of the agentic era, and Claude Code has added more than $65B in ARR since its May 2025 general release.
•OpenAI and Anthropic released frontier models every 51 days on average during the agentic era, yet Kimi K2.6 surpassed Opus 4.5 in 4.8 months and GLM-5.2 cleared GPT-5.2 in 6 months.
•Fireworks is processing over 40T tokens per day, which the article says is 2x OpenAI API's volume at the end of March, highlighting exploding inference demand.
Meta's Llama models are repeatedly cited as major open-source milestones — Llama-2-70B first approached the frontier, Llama-3.1-405B closed the GPT-3.5 gap, and Llama-4 Maverick extended open-model mo
Meta's Llama models are repeatedly cited as major open-source milestones — Llama-2-70B first approached the frontier, Llama-3.1-405B closed the GPT-3.5 gap, and Llama-4 Maverick extended open-model momentum — validating Meta's open-weight AI strategy.
Risk: Meta does not directly monetize Llama in a meaningful way, so open-model leadership may not translate into near-term revenue.
The article highlights that Qwen2.5-72B landed within striking distance of GPT-4o at one-sixth the parameter count on 18T pre-training tokens, showing Alibaba's Qwen line is a credible frontier-adjace
The article highlights that Qwen2.5-72B landed within striking distance of GPT-4o at one-sixth the parameter count on 18T pre-training tokens, showing Alibaba's Qwen line is a credible frontier-adjacent open model.
Risk: Open-source model capability is not the same as Alibaba's cloud or commercial AI monetization, so the positive signal is indirect.
This newsletter, published August 21, 2026,
features Evan Cloutier
discussing META, BABA.
2 trade ideas extracted by AI with direction and confidence scoring.