Thompson credits the harness for the agent jump that retired his bubble call

The date is the interesting part, and Thompson flags it himself. Anthropic released Opus 4.5 on November 24, 2025, to what he describes as relatively little fanfare ^[12]. The capability he is pointing at arrived later. At some point in December, he writes, Claude Code with Opus 4.5 "suddenly seemed to be able to do things that were never possible previously" ^[13]. Thompson credits the harness. He writes that a critical component of making agentic workloads work is the harness, the software that actually controls the model ^[16].

That distinction has a practical edge for anyone buying capacity. A harness is code in a repository, and what goes into the context window can be changed without a training run. OpenAI shipped GPT-5.2-Codex on December 18, twenty-four days after Opus 4.5, and Thompson says it was similarly capable ^[14][18]. By his account both were finishing tasks that took hours, and finishing them correctly ^[15].

Set against the earlier inflections, the spacing is uneven. Transformer-based LLMs were introduced in 2017 and ChatGPT launched in November 2022, roughly five years later ^[6][5][21]. o1 followed in September 2024, about 22 months after ChatGPT, and Opus 4.5 came about 14 months after o1 ^[9][19].

Each step changed who carries the error-handling. Thompson's complaint about the first ChatGPT releases was that they hallucinated, and that a user had to know what to use them for and verify the output ^[8]. On o1 he wrote in an Update at the time: "The big challenge for traditional LLMs is that they are path-dependent; while they can consider the puzzle as a whole, as soon as they commit to a particular guess they are locked in, and doomed to failure" ^[10]. Reasoning moved that self-check inside the model, and the harness moves the loop that runs the task outside it ^[9][16].

What the published excerpt supports is a claim about capability. It breaks off mid-sentence in the passage on Claude Code and Codex abstracting the user away from the model, and it does not include figures for capital spending, utilization or token volumes ^[20][17]. So a buyer who wants to cite this in a capex review is citing three model release dates and one writer's December. For the conclusion to transfer, the hours-long tasks would have to be ones that a buyer's own reviewers accept without redoing the work, and the tokens burned on them would have to be billed to a customer who renews. Thompson had previously argued that bubbles can be good ^[1]. Writing in March 2026, on the morning of Nvidia's GTC, he reached a different conclusion and hedged it in the same breath: "I don't think we're in a bubble (which, paradoxically, maybe is the truest evidence we are)" ^[2][3].