Our new regular column where we wrap up the most exciting news in AI this week!
The week at a glance (TL;DR)
- OpenAI GPT-5: shipping with an automatic router; claims SOTA on math, coding, multimodal; massive user base to absorb it. openai.comBusiness Insider
- Anthropic Opus 4.1: incremental but meaningful coding/agentic gains; 74.5% SWE-bench Verified; now easier for U.S. agencies to procure. anthropic.com+1
- Open-weight surge: OpenAI’s gpt-oss models + NVIDIA’s RTX optimizations point to cheaper, local reasoning options. openai.comNVIDIA Blog
- Chip geopolitics: U.S. licenses Nvidia’s H20 exports to China; big implications for global AI capacity and costs. Reuters
- AI for ecology: DeepMind’s Perch widens its species coverage; open access for scientists. blog.google
OpenAI launches GPT-5, pushes “router” UX, touts big benchmark jumps
OpenAI rolled out GPT-5 on August 7, calling it “our best AI system yet.” The release emphasizes a real-time router that automatically picks between a fast model and a deeper-reasoning model, so users don’t have to. On benchmarks, OpenAI claims: 94.6% on AIME 2025 (no tools), 74.9% on SWE-bench Verified, 84.2% on MMMU, 46.2% on HealthBench Hard, and 88.4% on GPQA with GPT-5 Pro. openai.com
- Why it matters: removing the “model picker” could make advanced reasoning feel invisible—and mainstream.
- Real-world scale: OpenAI is already serving 700 million weekly ChatGPT users, and the new router aims to make answers “smarter” without extra user effort. Business Insider
It makes that decision for you.
OpenAI COO Brad Lightcap on the new openai router
Anthropic ships Claude Opus 4.1—and opens a federal procurement door
On Aug 5, Anthropic released Claude Opus 4.1, an upgrade targeting agentic tasks, real-world coding, and reasoning. Headline stat: 74.5% on SWE-bench Verified, with claimed gains in multi-file refactors and precise, minimal-change bug fixes. It’s available to paying users, via API, and on Bedrock and Vertex AI at the same price as Opus 4. anthropic.com
Two hours later in the same news cycle, Anthropic said U.S. federal agencies can now buy Claude via the GSA schedule, streamlining procurement with pre-negotiated terms. For a vendor hunting large public-sector rollouts, that’s a big unlock. anthropic.com
OpenAI goes “open-weight” with gpt-oss-120b and gpt-oss-20b; NVIDIA tunes for local RTX
Also Aug 5, OpenAI unveiled gpt-oss-120b and gpt-oss-20b—its first open-weight releases since GPT-2. They’re positioned as reasoning-forward text models, with download and fine-tune rights. NVIDIA published optimizations so these models run efficiently on GeForce RTX/RTX PRO PCs; the broader NVIDIA posts highlight performance pipelines from “cloud to PC.” Reuters also framed the models as optimized to run on laptops, capturing the on-device push. openai.comNVIDIA NewsroomNVIDIA BlogReuters
- Nuts and bolts: NVIDIA’s technical feed underscores accelerated inference and the larger infra picture (including an eye-catching “1.5M TPS inference” demo on GB200 NVL72—very much a data-center proof point). For builders, the headline is simple: more ways to run competent, inexpensive models locally and at the edge. NVIDIA Newsroom
Washington gives Nvidia a (conditional) green light back into China
The U.S. Commerce Department has started issuing licenses that let Nvidia export H20 AI chips to China—reversing an April ban last month and now unlocking actual shipments. Nvidia previously warned the curbs could slice $8B from its July-quarter sales; the company later reduced the expected hit by $1B via material reuse. Context: H20 generated $4.6B in Q1 revenue, and China was 12.5% of overall revenue. CEO Jensen Huang even met with President Trump this week as the situation evolved. Reuters
- Why it matters to AI: model training and inference economics ride on GPU supply. Any reopening—even limited—to the world’s second-largest economy changes demand/supply math for everyone building or serving AI at scale. Reuters
Google DeepMind’s Perch gets a conservation-scale upgrade
DeepMind/Google updated Perch, an open audio model used to analyze bioacoustic recordings and spot endangered species. The new version generalizes across habitats—from Hawaiian honeycreepers to coral reefs—and is available openly (Kaggle). It’s a solid example of frontier AI spilling into climate/biodiversity tooling, not just chatbots. blog.google
If you only watch two things next week: (1) how GPT-5’s router changes everyday UX (and retention), and (2) whether open-weight models plus local acceleration start pulling small teams off pricey proprietary stacks.