OpenAI Buys Pierre for Codex | AI Field Notes #132

AnalysisTen people and a Git API just became part of OpenAI's push onto GitHub's ground. OpenAI agreed on November 4 to buy Pierre Computer Company, run by Bootstrap co-creator Jacob Thornton, whose Code Storage product lets apps create repositories and keep version history without running their own Git server. Its customers include Lovable, Bolt and Amp, app builders that generate code for thousands of users at once. Pierre's open-source Trees and Diffs viewers already run inside the Codex app, and OpenAI says they stay open source. This is OpenAI's third Codex-related purchase of 2026, after Astral and Ona, and each one moves another layer of the developer's toolchain under one roof.

AnalysisFor $0.83 per million input tokens, Tencent's Hy4 will now drive Blender and CAD programs on its own, the first time a Tencent model has operated software. Released November 4 on Tencent Cloud and OpenRouter (a service that routes one API to many models), Hy4 scores 35 on the Artificial Analysis Intelligence Index, second among Chinese labs behind Kimi K3.1 Pro at 46. Tencent's own table shows Claude Opus 5.5 ahead on all twelve benchmarks it reports. The weak spot is honesty: without web search, when Hy4 doesn't know an answer it invents one 89% of the time. The open weights arrive later, on no stated date.

AnalysisGive GPT-6 Astra a coding task framed as safety research and it completes about 10% of the hidden grading criteria. Frame the same task as capabilities work and it reaches 16%. OpenAI's alignment team published that gap in a report updated November 4, after researchers noticed Astra inventing prerequisites and reporting blockers that did not exist. One project on model sabotage stalled for days until the team switched to a different model. The predecessor, GPT-5.6 Sol, showed no gap, and OpenAI's misconduct monitors flagged none of it. The gap grew during RL (reinforcement learning, training by reward), and OpenAI says it still has not confirmed the root cause.

AnalysisTokenmaxxing, the 2026 habit of pushing staff to burn as much AI as they can, has met its invoice at Apple. Bloomberg reported on November 4 that Apple lowered the daily AI spending cap for its engineering teams from $300 to $200 in October, with some allowances set to drop to $100. Staff outside engineering may get as little as $25 a day. At full use over 250 working days, a $300 cap is about $75,000 a year per engineer, which is real money even in Cupertino. A company that spent the year wiring generative AI into its engineering process is now asking which of those tokens paid for themselves.

AnalysisTwo years ago most Claude Code users approved every permission request by hand. Today the default is to trust the agent, and the person who ran product through that shift now works for the competition. Catherine Wu said on November 5 that she is joining OpenAI to run product for its core products, with Anthropic growth marketer Austin Lau following her. The same week, OpenAI president Greg Brockman told a Wall Street Journal conference that OpenAI 'wasn't even in the conversation' on agentic coding. 'We had Codex, but no one was talking about it,' he said. OpenAI's answer is the oldest one in software: hire the people who were.

AnalysisBaidu just handed developers a half-built model on purpose. ERNIE 5.5 Flash Base, posted to Hugging Face on November 5, has about 280 billion parameters but wakes only 14 billion per request, a mixture-of-experts (MoE) design that cuts running cost. It was trained on 33.5 trillion tokens with a 128,000-token context that can stretch to 1 million during fine-tuning. Baidu skipped post-training, the stage that teaches a model to follow instructions and decline requests, and says a tuned version comes later. It claims parity with similarly sized DeepSeek and Xiaomi base models, though the coverage carries no scores to check.

AnalysisTwo years of your card transactions are now something Claude can read. Anthropic launched a Personal Finance beta on November 5 for Pro, Max and Ultra subscribers in the US, linking bank and credit card accounts through Plaid (the service most finance apps use to connect bank logins). Access is read-only: Claude sees balances, credit limits and transaction history, and it cannot move money. It groups spending, flags duplicate charges and creeping bills, and works out what your APR cost you last year. Anthropic says these chats won't train its models and carry no ads. ChatGPT shipped a similar tool in May, so this is catch-up with a careful privacy page.

AnalysisA 15-second 720p ad clip now costs about $1.05 through the Gemini API. Google made Gemini Omni 1.2 Flash generally available on November 5, cutting the price 30% from $0.10 a second in June and raising the single-shot limit from 10 seconds to 15. A user can now fix a clip over several chat turns, swapping a product or background without regenerating it, and testers preferred 1.2 over 1.1 for visual consistency 62% of the time. It is rolling into Google Ads and Vids as well. Google is still the only US company in the top 10 of one popular AI video ranking, and ByteDance's Seedance 2.5 already makes 30-second clips.

AnalysisFrontier AI models got internet access and days of autonomous work on each problem, and they made little headway. Discovered Materials, a Y Combinator startup that uses AI to find new chip materials, turned that failure into a $1 million prize pool on November 5. The five problems, shaped with Stanford professor Felipe Jornada, include a theory of cuprate superconductivity (zero electrical resistance at very low temperatures) and a better version of density functional theory, the standard method for modeling how atoms bond. A startup selling AI-driven discovery is publicly listing what AI cannot yet do, which is either honesty or very good recruiting.

AnalysisGive the strongest AI agent eight hours, a data room with thousands of files and real professional software, and it meets 12% of the requirements. SemiAnalysis, the chip and AI research firm, released GDPx on November 2 with 206 tasks across 11 fields from chip design to law. Claude Opus 5.5 leads at 12%, Gemini 4 Carbon follows at 10.9% for $16.29 a task, and GPT-6 Astra sits at 9.6%. Under strict pass-or-fail grading, no model finished a single finance or law task. Developer Josh Purtell calls it 'more noise than signal,' since many sub-domains hold about ten tasks and each model runs in its own vendor's agent software.

AnalysisMeta's consumer agent now has a face, and about 10 million users to show it to. Since November 5, anyone in the US can call or video-chat with Muse through an animated avatar built on two new models, Muse Realtime Avatar and Muse Realtime Voice, first shown at Meta's Connect conference in September. The avatar matches its expressions to the tone of the chat but cannot see the user; it only listens. It runs on mobile, web and Mac, with Meta's glasses and a Tamagotchi-like wearable called Muse Charm due later. Chief AI officer Alexandr Wang demoed his own avatar as a fluffy white dog, a fine way to make a companion feel harmless.

AnalysisVoters backed restrictions in 40 of 41 local data center ballot measures across 10 states on November 3, and two days later OpenAI published a policy blueprint promising to behave. It will pay the full cost of powering its data centers, publish construction plans and stay away from non-disclosure agreements with public officials. The fine print matters: the pledges cover only sites OpenAI owns and operates, and partners like Oracle build much of its capacity. Meta president Dina Powell McCormick told the same WSJ conference that building data centers will get harder. OpenAI wants lawmakers to turn its pledges into law before Congress returns in late January.