ChatGPT and Codex Shared Limits | AI Field Notes #129

AnalysisA coding agent that waits for a prompt already looks dated to DeepSeek. Its Harness team published dsh-superagent on GitHub around November 1 under the MIT license (free to use and modify, including commercially). The agent watches one repository's GitHub threads, and optionally Feishu chats, opens a work item when an authorized person asks or a team-written rule allows it, then does the job in its own session and replies with a pull request. DeepSeek's stated goal is "the maximum value with the minimum human attention." The repo drew 13,500 stars and calls itself a research prototype that may contain bugs. No benchmark yet shows it beating Claude Code or Codex.

AnalysisA risk factor rarely reads like a research note. Anthropic's revised IPO filing, submitted on November 2, says internal estimates suggest future models may "imminently" meet its Responsible Scaling Policy threshold for "dramatically accelerated progress attributable to AI," the line it draws around recursive self-improvement (AI that helps build better versions of itself). No current model has crossed it, the company says. The more awkward admission is that Anthropic "has in the past not identified" such jumps until after they happened. It adds that it does not expect to let Claude build successive frontier models without meaningful human supervision. Investors now get to price that sentence.

AnalysisOne meter will soon run both products. From December, eligible ChatGPT and Codex usage on Plus, Pro and Business plans will draw from a single token-based allowance, according to an OpenAI help-center page updated on October 31. Plain Instant chats stay exempt; the change targets compute-heavy work such as agentic coding runs and long reasoning. Settings will show what is left and when it resets, and anyone who hits the wall can wait, upgrade or buy credits. OpenAI has not published the actual token numbers, so nobody can yet say whether a long Codex refactor now costs a week of research questions.

AnalysisPost-training, the step where a general model gets tuned for one job, has mostly needed machine-learning engineers. Trajectory wants a lawyer to do it. The 18-person startup, founded about a year ago by three Stanford classmates from Google DeepMind and Apple, raised a $45 million Series A at a $300 million valuation led by Sequoia, with Nvidia joining, TechCrunch reported on November 2. Its now-public platform starts from an open-source model, collects feedback from real use and retrains the model on it. Customers already include Harvey, Rogo, Decagon and Mercor. Experts grade, the system learns, and the engineers mostly step back.

AnalysisMost companies blame restructuring. SUSE, the German Linux company, put AI in writing. CEO Dirk-Peter van Leeuwen told staff on November 2 that roughly 10% of about 1,900 employees will go because "automation and AI are fundamentally shifting how software companies operate" and "the requirement for some existing roles has greatly diminished." Affected staff hear by November 4. The cuts follow a voluntary exit program in September that too few people took, The Register reported. The memo adds that AI is "advancing exponentially, forcing us to radically reimagine our work," a hard sentence to argue with across a redundancy-meeting table.

AnalysisA prompt box now does what a drag-and-drop template used to sell. Squarespace confirmed on November 2 that it is cutting about 19% of its workforce, roughly 300 people, to "simplify how we operate" and "sharpen our priorities." The statement never mentions AI. The market around it does: vibe-coding builders (tools that generate a working site from a typed request) have upended web design, Lovable raised $400 million at a $13.3 billion valuation in August, and Wix bought Base44 for $80 million. Squarespace's own answer, an AI-native platform called Forma, still sits behind a waitlist. The cuts arrived before the product did.

AnalysisA food-delivery company now serves one of the largest models on the market. Meituan released LongCat-2.5 on November 2 with 1.6 trillion parameters in total and about 48 billion active per token, thanks to a mixture-of-experts design (a model that wakes only part of itself for each request to cut cost). It takes images and text in one stream and was trained on long task sequences for agents that keep working on whole software projects. It is live in Meituan's chat app and API, with GitHub and Hugging Face releases promised. Meituan claims parity with Claude Opus, but no independent benchmark existed at launch.

AnalysisA trillion-dollar valuation still needs someone to write the check. OpenAI has lowered the goal for its new round from $30 billion to $20 billion, the Financial Times reported on October 30, with the first close expected below that and about $11 billion secured so far, according to the Wall Street Journal. UAE-based investors lead, and a retail tranche for wealthy individuals may fill the gap. The implied valuation slid as well, from $1.4 trillion in late-September reports to about $1.25 trillion. TechCrunch ties the cooling to a tech-stock selloff and Wall Street fatigue with AI infrastructure spending. OpenAI did not comment.

AnalysisThird-quarter revenue grew about 1,127% from a year earlier and still looks small next to the bill. Anthropic's amended S-1, filed November 2, shows roughly $17 billion in quarterly revenue and $33.24 billion for the first nine months, CNBC reported. The quarter's net loss was $23.29 billion, mostly from fair-value changes on convertible notes. The figure that matters for anyone building on Claude is $895.4 billion in purchase and lease commitments, "substantially all" of it for compute and infrastructure across Amazon, Google, Microsoft, AMD and SpaceX-hosted Nvidia capacity. That is about 27 times what the company booked in nine months.

AnalysisThe coding assistant on your laptop is on theirs too. Microsoft's threat team reported on October 29 that four North Korean groups likely used AI development platforms such as Cursor to write, debug and port malware, and that since late October two groups have used ARTEX, an open-source agentic penetration-testing framework (software that probes networks for weaknesses on its own). One group produced more than 100 tailored PDF lures in August and September, dozens within minutes. Another, Jasper Sleet, uses generative AI to run fake remote IT workers across several personas. Microsoft rates adoption as high and human-directed: the AI speeds up old playbooks rather than writing new ones.

AnalysisThe company built on defense contracts now sells more to American businesses than to Washington. Palantir reported on November 2 that US commercial revenue reached $984 million in the third quarter, up 148%, edging past US government revenue of $953 million for the first time. Total revenue hit about $2.31 billion, and full-year guidance rose to roughly $8.53 billion. Its commercial pitch is an AI platform that wires language models into a company's own data and approval chains. CEO Alex Karp called the run unmatched "in the history of software AI, ever." Shares still slipped about 3% after hours.

AnalysisEven a preannounced raise spooked the market. Oracle filed on November 2 to sell $10 billion of common stock and $10 billion of senior notes to fund its AI cloud buildout, and the shares fell more than 6%, CNBC reported, after a roughly 33% slide this year. Oracle already carried $125.34 billion of debt as of August 31, and S&P cut it to BBB-, one notch above junk, in July. The company says capital spending peaks this fiscal year and next, and it plans no more than another $20 billion of AI financing through fiscal 2030. Customers rent the capacity. Shareholders now pay for it.