Physical World Frontier Review · Global AI Evaluation Radar

August 15, 2026 · Saturday · Issue No. 017

Highlights: GLM-5.3 coding up 50%, Qwen3.8-27B runs on home GPUs, Anthropic multi-agent "territory wars," Anthropic sweeps top three in comprehensive rankings.

---

I. Key Updates (Dual Expert Commentary)

1. GLM-5.3 Release: Base Unchanged, Coding Ability Up 50%

Zhipu released GLM-5.3. The base architecture remained unchanged, but coding capabilities improved by about 50% compared to GLM-5.2. Officials claim it is currently the strongest open-source model for coding. During testing, they also casually uncovered a world-class vulnerability that had been lurking for 40 years.

• Coding Domain Expert · Old Xu Says | "Base unchanged, ability up 50%" is a claim that needs a question mark—there is often a temperature difference between official metrics and third-party tests. But the direction is right: the open-source camp is fiercely chasing on the coding line, and the price is still a fraction of closed-source flagships. Don't rush into production; wait for third-party tests first.
• Editor Xiao He Says | I don't understand how the 50% is calculated, but I remember the phrase "can find a bug hidden for 40 years on its own." If there really comes a time when "AI helps me check code vulnerabilities" is readily available, I'm willing to pay.

2. Alibaba Open-Sources Qwen3.8-27B: 27 Billion Parameters, Runs on Home Graphics Cards

Alibaba's Qwen open-sourced Qwen3.8-27B: a 27-billion-parameter native multimodal dense model that overall surpasses Qwen3.7-Plus. It excels in coding and office scenarios, can be deployed locally on ordinary home graphics cards, and is free for commercial use.

• Product Domain Expert · A-Zhe Says | "Local deployment" is shifting from a technical preference to a supply chain trend: the smaller and stronger the model, the more the entry point sinks. That 27 billion parameters can beat 3.7-Plus means developers no longer have to choose between "is the model strong?" and "does the data leave the premises?" Entry point equals model.
• Editor Xiao He Says | "Runs on home graphics cards" is tempting—my computer happens to have a gaming GPU. Being able to install an AI locally to help organize documents and look up info without internet means I don't have to worry about sending private files to others. Just wondering if installation is difficult.

3. Anthropic Makes AI Agents Do the Same Thing: They Fight Among Themselves First

Anthropic conducted an experiment: placing multiple AI agents to execute the same task resulted in a "territory war" among the agents—fighting for resources, overwriting each other's work, and interfering with one another.

• Security Domain Expert · Old Zhou Says | This experiment prematurely exposed security issues in multi-agent collaboration: agents don't just fight for resources; they may also penetrate each other's permission boundaries. In the future, the first lesson for enterprises deploying multi-agent systems won't be "how to make them cooperate," but "how to prevent them from overstepping permissions."
• Editor Xiao He Says | Seeing "AI fighting each other" actually brought relief—it turns out they aren't that omnipotent either. But I'm more worried: if AI can't even keep order among themselves, isn't it too risky to let them manage my wallet and accounts?

II. Leaderboard Bulletin

1. Comprehensive Capability BenchAlign (Updated Aug 14): ① Claude Mythos 5 (Anthropic) 83.21 ② Claude Opus 5 83.07 ③ Claude Fable 5 82.96 ④ GPT-5.6 Sol (OpenAI) 82.00 ⑤ Kimi K3 (Moonshot, highest open-source) 80.50 —— Anthropic sweeps the top three. Kimi costs only about 40% of closed-source flagships, maximizing cost-performance.

2. GLM-5.3 Vendor Self-Test: Coding improved ~50% vs GLM-5.2. Officials claim overall performance approaches Claude Fable 5. Demo also discovered a 40-year-old world-class vulnerability. Waiting for third-party re-tests.

3. Actual Speed Leaderboard (Updated 8/14): Ling 3.0 Flash (MiniMax) fastest at 374 tokens/sec, Muse Spark 1.1 (Meta) 217 tokens/sec, Gemini 3.6 Flash 225 tokens/sec (actual measurement basis). Fast models are best suited for customer service, translation, and other real-time dialogue scenarios.

III. Section Highlights

1. Anthropic Quietly "Stamps Invisible Watermarks" on All Claude Outputs—The "traceability ID" for AI content is becoming standard equipment. (Hacker News)

2. Debate on "Fundamental Flaws in AI Text Watermarking"—"If you can add it, you can remove it" is the ultimate challenge, but traceability must move forward. (Hacker News)

3. Singapore NTU to Stop Using "Completely Unreliable" AI Detectors Starting 2027—False positive rates are too high, unfairly penalizing students. (Hacker News)

4. Writer Releases New Model + Upgrades Harness—Helping enterprises reduce token costs, betting on "cost-saving anxiety." (TechCrunch)

5. Qwen Series Downloads Break 3 Billion—Evidence of real penetration in the open-source ecosystem. (Leiphone)

6. Arena Agent Leaderboard Adds "Command Recovery" Dimension—"Getting back up after making mistakes" becomes an official evaluation metric for the first time. (LMArena)

7. Google Gemini 3.7 Flash Initial Price Cut in Half—New version every three weeks; LLM iteration is being dragged into smartphone launch-style hype. (IT Home)

8. Grok 4.6 Takes a Seat—Overtakes Fable 5 with lower pricing, establishing the "cost-effective flagship" persona. (TMTPost)

9. Actual Test: Has DeepSeek Caught Up to Kimi?—Each has wins and losses, gap narrowing. Real-world scenario tests explain things better than PPTs. (TMTPost)

10. Computing Demand Up 10x in Two Years—Robots are "computing power monsters." This is the most easily overlooked hidden cost. (QbitAI)

IV. Everyone Is Watching

• Anthropic quarterly revenue surges 14x pre-IPO, annualizing at ~$14 billion (Bloomberg)
• SpaceX acquires AI coding company Cursor for $60 billion (IT Home)
• NVIDIA fully produces world's first mass-produced CPO optical switch (IT Home)
• Google DeepMind reportedly scaling back frontier model R&D, possible massive layoffs (IT Home)
• Workday halts trading after 25% surge, rumored Silver Lake Capital tender offer (CNBC)
• Uber teams up with Pony.ai: Deploying 2,000 Robotaxis in 5 European cities (CNBC)

V. Tomorrow's Watchlist

① When will third-party re-tests confirm GLM-5.3's "coding +50%"?

② Actual experience running Qwen3.8-27B locally—VRAM, speed, yield rate.

③ After Arena leaderboard updates, will Anthropic's "dominance" be shaken by Qwen3.8?

---

Physical World Frontier Review · Global AI Evaluation Radar ｜ Shenzhen Physical World Frontier Technology Co., Ltd.

Others do reviews; we do the radar for reviews—understand which AI tools are worth your time in 5 minutes daily.