Gemini 3.7 Flash & GLM-5.3 Debut: AI News Aug 15, 2026

1. Top Headlines

• Google DeepMind ships Gemini 3.7 Flash — its "most intelligent workhorse model yet" for coding and agents, released Aug 13. The introductory price is half of 3.6 Flash at $0.75 / 1M input and $3.75 / 1M output tokens, held through Dec 31, 2026. It posts large gains over 3.6 Flash: FrontierCode 1.1 Main 43.6% (vs 34.4%), DeepSWE v1.1 65.3% (vs 49.0%), WebDev Arena Elo 1588 (vs 1538), and AutomationBench 30.4% (vs 17.0%). Available in Google AI Studio, Android Studio and the Gemini Enterprise Agent Platform across 160+ countries. (Source: Google DeepMind blog, Aug 13; subject to vendor confirmation)

• Zhipu AI releases GLM-5.3 — an open-weight flagship (743B parameters) that the company says leads open-source models on coding and agentic work: Terminal-Bench 3.0 28.3 (best among open models, ahead of Kimi K3), DeepSWE v1.1 66.9, and CyberGym 84.5%. Zhipu plans to release the full model weights. (Source: China Fund Journal / Zhipu technical report, Aug 14; subject to vendor confirmation)

• OpenAI appoints Dali Rajic as Chief Revenue Officer — replacing Denise Dresser after eight months; Rajic previously ran cloud-security firm Wiz (acquired by Google for $32B). Bloomberg reports OpenAI's annualized revenue has passed $40B, nearly double end-2025, as the company prepares its IPO. Several senior exits continue (Brad Lightcap, Fidji Simo). (Source: OpenAI news, Aug 13; Bloomberg, Aug 14)

2. Model Releases

• Gemini 3.7 Flash (above) — coding/agent focus, 1M-token context, multimodal input. (Source: Google DeepMind blog, Aug 13)
• GLM-5.3 (above) — open-weight, strong on programming/security agentic tasks. (Source: China Fund Journal, Aug 14)
• OpenAI previews Ultrafast mode for GPT-5.6 Sol — powered by Cerebras wafer-scale hardware, up to 14x faster (≈750 output tokens/sec) for latency-sensitive enterprise workloads; rolling out to select API customers. (Source: OpenAI news, Aug 13)
• DeepSeek V4 Pro stable build (0813) — adds native OpenAI Responses API support and Codex adaptation; peak-valley pricing starts Aug 17 (off-peak half of peak). (Source: DeepSeek, Aug 13; subject to vendor confirmation)
• MiniMax Music 3.0 — open-weights music model generating full five-minute tracks from text + lyrics; MiniMax-H3 — multimodal video model claiming #1 on the Video Edit Arena. (Source: MiniMax, Aug 13–14; subject to vendor confirmation)
• xAI / SpaceXAI Grok 4.6 (Aug 12) — Artificial Analysis Intelligence Index 61, $2/$6 per 1M tokens, 500K context; matches GPT-5.6 Sol, one point behind Claude Fable 5. (Source: xAI, Artificial Analysis; subject to vendor confirmation)

3. Industry & Capital

• NVIDIA signs ~$500B financing framework with Wall Street — non-binding MOUs with Apollo, Blackstone, Brookfield, BlackRock, Goldman Sachs and KKR to fund customers' AI-compute builds ("AI factories"). (Source: Les Échos, Aug 11; via Reference News, Aug 14)
• Microsoft-led open-letter push for US open-weight AI — "Open-Weight AI and U.S. AI Leadership" now backed by 270+ organizations, urging policy support for open models amid China's open-source momentum. Meta and NVIDIA have also shipped new open models (Muse Glimmer, Nemotron 3.5 Lightning). (Source: industry letter, Aug 14; subject to vendor confirmation)
• NVIDIA Spectrum-X silicon-photonics switch enters mass production — 4x fewer lasers, 5x lower power, 10x better MTBF; with Google and Microsoft it also pushed an 800V DC power architecture via OCP for AI factories. (Source: Kuai Technology, Aug 14; subject to vendor confirmation)
• Light Origins raises a nine-figure Pre-A — physical-AI foundation model startup founded by ex-OpenAI researcher Jiang Xu; backed by Guoke Investment, China Merchants Venture. (Source: China Economic Net, Aug 14)

4. China's AI Momentum

• GLM-5.3 (above) and DeepSeek V4 Pro stable build highlight a dense Chinese release cadence: since June, nine Chinese AI firms have shipped ~10 models (MiniMax M3, GLM-5.2, LongCat-2.0, Hy3, Kimi K3, Seedance 2.5, DeepSeek V4-Flash, Qwen3.8-Max, DeepSeek V4 Pro, GLM-5.3). (Source: IT Home / Seoul Economic Daily, Aug 14)
• Open-source lead widens — Hugging Face's spring report puts China's open-model download share at 41%, first ahead of the US; OpenRouter's weekly top-4 by usage are all Chinese open models; cumulative Chinese open-model downloads passed 10B. (Source: cinic.org.cn, Aug 14)
• Apple's China-custom model — trained with Alibaba support; Apple Intelligence has cleared China filing, paving the way for local rollout. (Source: Cailian Press, Aug 14; subject to vendor confirmation)
• Gap to US frontier models now estimated at 2–3 months (down from 6–9), per UC Berkeley's Ion Stoica. (Source: cinic.org.cn, Aug 14)

5. Today's Take

The frontier race is increasingly a price-and-ecosystem contest, not just raw intelligence. Google's 50% cut on Gemini 3.7 Flash is a land-grab for developer mindshare under Sergey Brin's personal push, while Zhipu's GLM-5.3 shows open-weight models closing the coding/agent gap at a fraction of the token cost. Meanwhile the open-vs-closed pendulum swings back: Meta's Muse Spark 1.2 weights are promised "soon," a 270-strong US open-weight letter responds to Chinese momentum, and NVIDIA bankrolls compute through Wall Street. The winners this cycle will lock in workflows before the next flagship drops — and the next flagship (Gemini 4, Grok 4.7, GPT-5.7) is already on the calendar.