Three model releases, and three different stories

The read

Three model releases in one window, and they do not tell the same story. Grok 4.6 rejoins the frontier at about a third of frontier pricing — 61 on Artificial Analysis, level with GPT-5.6 Sol and just behind Opus 5, on unchanged US$2/US$6 pricing. Gemini 3.7 Flash arrived three weeks after 3.6 at half the price, with large jumps on Google's own benchmark figures. And DeepSeek V4 Pro landed at 53, one point above its own Flash — against an expectation in this conversation that it would be the moment a fourth frontier lab arrived at a fraction of frontier prices. Two of those say capability keeps getting cheaper on a three-week clock. The third is the reminder that a scaling step can simply fail to arrive.

The financing argument had its sharpest exchange of the week. A Sydney Morning Herald piece on AI bonds prompted the case that depreciation is the premise of equipment finance rather than an argument against it — aircraft, rigs and rolling stock are financed against depreciating collateral as a matter of routine, and the real question is the schedule, not the direction. Set against that, Steve Eisman warns the boom now rests on the fortunes of two companies, while Anthropic's CFO has reportedly held early IPO meetings at a valuation some investors expect near US$2tn — their expectation, not the company's. And a Chinese maker has moved past Micron and Kioxia into third place in NAND shipments, a volume ranking rather than a revenue one, but pointing the same way memory supply has been pointing all month.

Below the models, the layer that matters keeps shifting. The harness, not the model, is becoming the product surface — DeepSeek has shipped its harness as a developer preview, and practitioner consensus keeps landing on memory, approvals, evals and tools rather than bespoke models. Geoffrey Litt's "understanding is the new bottleneck" repays reading carefully: he raises "we understand to verify" only to reject it, and argues the reason to understand the code is to participate in it. And open weights reached the big end of town while shrinking by 91 per cent — AT&T betting on open-weight AI per the Journal's headline, alongside Unsloth's own unbenchmarked compression claim.

Two Australian items bear directly on the white paper above. The Federal Court has found the Relevant Electronic Services Standard does not apply to X, removing from that platform the instrument the eSafety Commissioner has enforced since mid-2025. And a data centre operator declined to say whether siting a facility beside a Sydney Metro station is good for Sydney, while the UN puts data-centre electricity at a projected 945 terawatt-hours by 2030. Mark's point is the one the industry keeps declining to make for itself: we already work in data centres — all of us, every single day.

Four more worth your time. A materials benchmark where seven models found 500 candidate compounds and one usable recipe, which puts a number on the gap between generation and synthesis. A reported ChatGPT-assisted solve of an open problem in numerical linear algebra — second-hand, and the provenance is left visible for a reason. Espressif has put an agent runtime on a few-dollar microcontroller. An open-source finance terminal has shipped with AI screens built in. And a free Rust textbook on building an inference stack from scratch, out of an Australian lab, if you would rather learn the layer underneath all of this than read about it.

Read all 16 signals →