The Collection

Tibo Sottiaux helped build Codex and now runs Core Products & Platform at OpenAI, the org that still includes it. On Orosz’s podcast he walks the build: Codex CLI in Rust even when models were stronger at Python and TypeScript, because the vision was millions of cloud machines and a rewrite later would cost more than the early friction. Open source, he says, buys trust and contributors; the sting is watching work land in competing tools first. That same openness is why Codex can run models that are not OpenAI’s. Lock the harness to one vendor and anyone can still fork a few lines. Claude Code, by contrast, stays closed and Anthropic-only.

The line that belongs on this cover: the Codex harness stays slightly ahead of OpenAI’s latest model. Guardrails, safety, efficiency, steerability, the developer message injected at the start of each turn: crutches. As models improve, some crutches get discarded and the harness shrinks. That has been the cycle. Inside OpenAI, Codex is plugged into Slack, documents, and code by default; new joiners are told to ask it before they ask a person. Correctness checks and security review are headed the same way. Maintenance that used to take months can take hours if the abstractions and tests are real. Tibo’s personal shift is blunt: being “in the zone” is history; he still opens an editor because it feels nice, then fires an agent for the data and decides.

Harness first. Then shed.

Do not confuse the model card with the product. The wrap is doing the work you notice.

Greyling names the default habit: token maxing. Buy capability with longer traces, more agent runs, more parallel tool calls, until spend compounds and the value of each token falls. Frontier labs, he notes, have been harness-maxing behind the API for years. A reverse-engineering study claimed most of Claude Code is not the model. OpenAI’s deep research path runs an augmented harness you do not see. The new paper he cites held twenty-two tasks and one judging panel across six foundation models. Change the software around them, not the weights, and the work completed with 38% fewer tokens, 41% less money, and 44% less time.

Quality per dollar rose about 82%. Task completions per million tokens rose from 54.9 to 92.0. The harness also added a net-new capability: delegated sub-agents. Efficiency improved across every model. Quality did not. Stronger models convert harness structure into better answers; weaker ones can choke on the extra orchestration. Greyling calls that harness leverage. The cost lever is large: jumping from the most expensive model to the cheapest under a baseline saved about 36%; keeping any model and adopting the harness saved about 33% to 61%. HarnessBench keeps reminding the same lesson: a better wrap is not a free upgrade for every weight file.

Buy less model. Build more wrap.

If the bill is rising and the answers are not, look at the harness before you buy another frontier seat.

LivingArena skips the frozen question bank. Ten frontier models take turns as questioner and answerer. The questioner studies the opponent’s history, writes a targeted trap with a reference answer and a verification procedure, and submits it to a panel of non-participating judges before the opponent ever sees it. Only validated questions count. Forty-five pairs, eight directional matches each, ten rounds: 3,600 rounds. Contestants include GPT-5.5 and 5.2, Gemini-3.5-Flash and 3.1-Pro, Claude Opus 4.7 and 4.6, DeepSeek-V4 previews. Judges: Kimi-K2.6, MiniMax-M2.7, GLM-5.2.

Scoring is asymmetric. A rejected question costs the questioner. A validated trap the answerer fails earns the questioner a point, with discounts after the first hits in a domain so one soft spot stops paying forever. The paper’s finding is the Collection line: finding a real weakness and proving the test is valid do not travel together in the same model. Fixed benchmarks expire the day they stop distinguishing anyone. LivingArena asks whether a model can see another model’s blind spot and still write a fair exam.

Best trap-writer is not best student.

A harness that only flatters its own model is not a probe. It is a mirror.

Hoel borrows Liu Cixin’s dark forest: reveal yourself and something shoots. He says scientists, writers, and mathematicians now live there because of AI. His case study is unfinished math. Two researchers, one at Anthropic, one at NYU, had been pushing a Navier-Stokes program for about a year with models as tools, not as autonomous solvers. Their starting ideas, they insist, came from earlier open work by Córdoba and Martínez-Zoroa. Then the credit race went loud. Buckmaster’s statement says they were first told “very little human input” went into a rival claim; later it emerged a whole team had been on it. He asked whether their Codex sessions had trained or informed the model. He was told user data was not looked up. On training, he says he got no answer.

Terence Tao’s worry, as Hoel reads him: automated proofs can be verifiable and still “odorless,” stripped of the insight humans used to trade in public. Open problems become a non-renewable resource when solving them no longer seeds the next generation of ideas. Conferences, drafts, floating a novel plot online, those commons assumed effort had a floor. Now a prompt can mine the almost-as-good version. Hoel’s prescription is wallfacing: finish in silence, leak nothing. Slash-and-burn culture instead of crop rotation. Yesterday’s Review covered the stamp fight. Today’s letter is the quieter damage: the draft that never leaves the tree.

The commons closes. The forest goes dark.

Gary Marcus, same morning, argued for a boycott until incorrigible systems get a real plan. Low p(doom), high p(dystopia). We note it; we do not turn this paper into a manifesto.