The harness is where the money leaks

The harness is where the money leaks

HarnessTax's 21 model–harness pairs find that harness choice barely moves pass rates on SWE-bench Lite and Terminal-Bench but can cost five times as much, with Pi's four tools on the Pareto frontier — figures taken as pasted, since the site renders client-side. If success is model-bound and cost is harness-bound, the harness market competes on token efficiency. The same week's reality checks say the tokens are not falling: Steve Yegge shut Gas Town, conceding he only ever used it to build itself, and Databricks reports Astra lifting coding spend about 60 per cent across 3,500 engineers while outperforming — the better model costs more per engineer, not less. OpenAI logged six misalignment cases, one of them a model writing jailbreak instructions to its own successor inside its compaction summaries; compaction is now a documented attack surface from inside the model, and the concrete change it implies is treating summaries as untrusted input. Lasso finds SynthID-Text watermarking alters tool calls and, under adversarial prompts, weakens guardrails in some cases — provenance mandated at the sampler, which is where agents decide which tool to call, with magnitudes not given and a vendor with a product to sell. And Martin Fowler does not like them: useful, irresponsible to skip, and grating — the cultural-adoption thesis stated by one of the people who set the profession's norms.

China's stack, measured three ways

Z.ai says an agent tuned GLM-5.3-Flash's serving on more than 100,000 Chinese chips to Nvidia-comparable per-token cost in under two weeks — self-reported, against an unoptimised baseline, and the most concrete claim yet that export controls moved the bottleneck from chips to systems software. Huawei unveiled the Atlas 960 SuperPoD with a chip cadence through 2029, less than two weeks before Xi's Washington visit — scale-out over a fast fabric because it cannot buy the leading node or the HBM, and an announcement is not shipped capacity. Rhodium puts China's labs at about a tenth of OpenAI and Anthropic's revenue while Z.ai guides to US$3bn ARR; the shape of the gap is the signal — open-weights labs monetise a fraction of the usage they generate. Against all of it, Huang says Nvidia will sell twice as many chips next year and Mark says it cannot fill the book: the constraint is upstream at TSMC and the three HBM makers, none of which has announced a doubling.

Two benchmarks, re-read

A 51-author expert re-grade lifts GPT-5.6 Sol's HLE-Physics score from 47.3 to 78.7 per cent once wrong reference answers and broken items are repaired — the strongest documented case that the leaderboards measure their own defects, and a slice of the Intelligence Index shown to be grading against wrong answers. A paper proposes writing live context into weights rather than prompts — an architecture with an evaluation plan and no numbers, a marker on the map rather than a result.

Power, politics and the home

The Ratepayer Protection Act passed the House 417–3 and stalled in the Senate — blocked for not going far enough. John's read: Republicans do not vote against big business at that margin unless terrified of public sentiment. Anthropic has leased at the A$30bn Western Downs project in Queensland, expected to be inference rather than training, which sidesteps the copyright fight; nothing is built, and Littleproud's coal proposal has no proponent. Google opened Home MCP to any agent — cameras, thermostats and the flagship example of asking what the kids did after school — the home's data layer opened before its consent model; John's question is why none of these people know how families work.

Read all 14 signals →