GDPx, DeepSeek's Super-Agent, and More
The most useful thing I read this weekend was a benchmark. SemiAnalysis released GDPx on Sunday night, pitched as a successor to OpenAI's GDPval for the agent era. It has 206 tasks across 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. Each domain was built with specialist partners, mostly RL environment startups, and every task drops the model into a realistic environment instead of handing it all the context in the prompt. One finance task has the agent work through a deal room of about 4,000 files, and the hardware tasks run in live sessions of tools like Siemens NX and SolidWorks.
The best model scores just 12%. Claude Opus 5.5 leads, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. Cost varies a lot: Opus averages about $51 per task, Carbon about $16, and DeepSeek 4.1 Flash under a dollar at 5.4%. SemiAnalysis also points out that safety classifiers pull down Opus and Astra in cyber, bio, and alignment, and that the finance and legal tasks are so hard that nearly every model scores zero on binary grading. Ethan Mollick's take is fair: it's a valuable benchmark, but it needs human baselines the way GDPval had them. (SemiAnalysis thread, GDPx leaderboard, SemiAnalysis write-up, Mollick's take)
What Else Happened
DeepSeek open-sourced a prototype "super-agent" for one GitHub repo. The dsh-superagent prototype, posted Saturday under MIT, listens to GitHub and optionally Feishu, opens a work item when an allowed person asks, works in its own session and directory, and replies in the original thread, for example with a pull request. What I like is the stated principle: "take as little human attention as possible, and give as much value as possible." It acts only when asked or when a team-written rule allows it, and it speaks only when someone needs to act. DeepSeek is clear that this is research code, not an official product, and not identical to its internal version. (GitHub repo, philosophy doc)
ChatGPT and Codex will share one token budget starting in December. OpenAI's updated help page says Plus, Pro, and Business plans will draw from a single allowance across ChatGPT and Codex, measured in tokens, as the agentic features of Work move into the main ChatGPT experience. Instant chats won't count. There's no exact date and no allowance sizes yet, so if you lean on Codex heavily, long agent sessions in ChatGPT will start eating into your coding budget. (OpenAI Help Center, The Decoder)
Google, OpenAI, Microsoft, and Anthropic will honor one robots.txt opt-out for AI training. On Friday, the four companies backed an IETF draft by Anthropic's Kevin Kelley. Sites add a group for the __usage_ai-training token instead of naming every company's crawler. It only covers training: it doesn't block crawling, doesn't affect search, and can't undo past scraping. Anthropic has started supporting it, Microsoft starts February 1, 2027, Google says early 2027, and OpenAI says "the coming months." For now, keep your per-crawler rules too. (Engadget)
OpenCode shipped a stable release of Hex, local voice dictation. It runs locally on Mac, Windows, and Linux, and the pitch is that voice is the fastest way to talk to whatever agent you use. DHH added it as a dictation option in Omarchy the same day. (OpenCode on X, Hex, DHH's Omarchy PR)
Guillermo Rauch made the case for owning your own intelligence. His argument is that local models need to keep spreading, and since the biggest models are out of reach locally for now, you should keep your options open: own your harness, data, connections, and runtime, and embrace open source. DHH offered a concrete example the next morning, running Nemotron 3.5 Bolt at 90 tokens per second on Nvidia's N1X to diagnose a sleep issue while offline. (Rauch's post, DHH's post)
The Gemini 4 Argon debate got loud. ThePrimeagen said "Google cooked with Argon 4," and Theo pushed back hard, saying it hallucinates badly in real codebases. GDPx happens to land in between, with Argon at 7.3%, behind Carbon and the top Anthropic and OpenAI models but at a lower cost per task. (ThePrimeagen, Theo's reply)
Launches and Demos
- T3 Code side chats – T3 Code added
/btwand side chats that work with any model or harness, and followed up with managed policies so IT teams can configure it across an organization. - GitHub Game Off 2026 – GitHub's month-long game jam is back for its 14th year, and the theme is SHIFT. AI-assisted development is allowed, but you'll be asked to disclose how you used it. Submissions close December 1.
- Thorsten Ball's one-prompt rebuild – he gave Amp one loose prompt to move his two 10-year-old book sites to a modern stack. It worked for 52 minutes, rebuilt both on Astro, found testimonials on Hacker News and Goodreads, and fixed mobile.
What I'm Reading
- shadcn on harness UIs – "There are two UIs that every AI harness will eventually build. Slack is already happening. Linear is next."
- swyx on low effort plus a critique loop – for the throughput-constrained, a low reasoning-effort setting with a critique loop beats maxing out the reasoning effort
- Ethan Mollick on GPT-6 Pro – not fast or pleasant, but it gives an "ask one question of the great oracle" feel for hard one-shot problems
That's it for today. 👋