Four stories crossed the wire this morning, and every one of them moved the same number. DeepSeek scored 5.5 intelligence points higher at an unchanged price; radiologists earn in an hour what generalist data workers make in a week; a federal order walled a model family out of US government; Canterbury students beat world-class humans at their own puzzle. One variable: the price of skilled work.
🔍 THE BOTTOM LINE
The AI economy is holding one price steady so everything else can move around it: the price of skilled judgement. Models get cheaper per task, but every layer requiring trained human verification — expert data, agent oversight, the judgement separating a demo from a reliable machine — is re-priced upward, and the rehiring data shows the market enforcing exactly that.
The curve that moves everything else
This morning’s pricing update is the pattern in one number. On 17 October, DeepSeek swapped V4.1-Flash-1017 into the same endpoint, same 552B backbone, same MIT licence, same off-peak price of $0.15 per million input tokens — and Artificial Analysis scored the new checkpoint at 45, up from 39.5 for the September weights. Intelligence per dollar keeps climbing; the dollar barely moves. An index built to separate capability from cost has spent the year showing the same gradient: task cost trending toward the floor.
The falling task-cost curve is why the DHS order matters beyond a security story. On Monday the US filed FASCSA order 2026-21447 removing DeepSeek products from every federal civilian system — signed 26 June, four months before it went public — and, critically, reaching third-party tools built on DeepSeek’s open weights, per our own coverage this morning. If you cannot beat a falling marginal cost, you regulate who is allowed to buy it. The deepseek file reads as one strategy — drive the unit price of capability toward zero, sell it everywhere — and the order is the countermove, a template other governments can copy.
The curve that pushes back
While raw capability gets cheaper, the human layer that certifies it is bid up with equal discipline. Bloomberg’s 16 October feature, covered this morning, reports recruiters advertising US$120 an hour for software engineers, US$170 for condensed-matter physicists and US$400 for civil-rights attorneys — while Mechanical Turk shut down after 21 years and Mercor raises at a reported $20 billion valuation. The rehiring correction says the same from the demand side: Robert Half counts 32 percent of managers who cut roles citing AI later rehiring, Forrester predicts half of AI-attributed layoffs reversed, and GFT found 84 percent of 945 CIOs had cancelled an AI pilot on legacy-system limits. Companies ran the substitution experiment in production — and, as the correction piece puts it, discovered “which parts of a role AI can actually carry,” then paid to rebuild the rest.
The premium has to be earned
What the robot story adds are the premium’s boundary conditions. Four UC mechatronics students spent a year not learning to solve a 4×4 cube faster — an open-source solver handles that — but building a machine that beats both the 45-second robot record and the roughly 15-second human one at 9.806 seconds, as covered this morning. The primary context: their two-phase solver traces to Herbert Kociemba’s open-source repository — 762 stars, standing since 2017. The solving brain costs nothing; the record was earned in the colour-sensing, the jamming fixes, the rigidity — ninety percent of the year, by the students’ account, spent on everything the algorithm does not know about the physical world.
That ratio is the argument in miniature: when intelligence is cheap, the premium migrates to what is expensive to build and verify. And the record is pending Guinness confirmation — verification itself remains the scarce step even here.
What it means for NZ
New Zealand holds the edge in exactly this layer, with a window that will not stay open long. SEEK’s September report puts AI skills in 4.2 percent of NZ job ads — nearly double Australia’s rate — and as the IMF analysis argued, the augmentation band is where NZ work already sits — a theme continued in yesterday’s pairing: the countries building rungs into the judgement layer are doing it before the entry tier thins out. The UC record shows what NZ can do: the full stack built on a Christchurch bench, capability already here, under-funded and under-visible. As agencies adopt cheap open-weight models with no NZ-equivalent procurement rules: measure AI spend in unit costs, not subscriptions, and invest the savings in the people who can check the output.
❓ FAQ
If AI is getting cheaper, why are AI-adjacent wages rising? The two sides are priced differently: execution cost is falling fast — 5.5 intelligence points added at an unchanged price — while verification bids upward, at US$120 to US$400 an hour per Bloomberg.
Are AI layoffs being reversed? Partly. Robert Half found 32 percent of employers who cut roles citing AI later rehired, and Forrester predicts half of AI-attributed layoffs will be reversed by 2027 — reduced hiring in some functions, increased hiring in judgement-heavy ones.
Why did the US government ban DeepSeek from federal systems? A Federal Acquisition Supply Chain Security Act order — signed 26 June 2026, published this week — excludes DeepSeek products from federal civilian systems, citing data-transfer risk to China, and reaches third-party tools built on its open-weight models. Civilian agencies only.
What does the Christchurch robot record say about the AI economy? When the intelligence layer is cheap, the premium sits elsewhere. The UC robot ran a free, decade-old solver; its record came from a year of engineering on sensing, mechanics and reliability — the expensive-to-verify layer.
What should NZ workers and agencies actually do? Workers: the premium attaches to depth in an existing profession, not a pivot into “AI.” Agencies: track unit-cost pricing, and note allied procurement regimes are starting to wall off model suppliers by regulation.
🔍 THE BOTTOM LINE
The direction over the next 12–24 months: capability consolidates into cheaper units while the premium concentrates on the people who can verify, oversee and integrate it. DeepSeek drives execution toward zero, Mercor bids up judgement, Amazon rebuilds alumni pipelines for oversight. The organisations still fighting the correction are pricing AI as headcount substitution — buying the cheapest layer twice, once as a tool and once to undo it.
📰 Sources
- Artificial Analysis intelligence index v4.3.2 (site pricing update, 20 October 2026)
- Bloomberg — AI’s $400-an-Hour Gig Economy (16 October 2026)
- Robert Half / Forrester / GFT-Wakefield surveys (2026)
- US Federal Register public inspection 2026-21447
- University of Canterbury robot record attempt (16 October 2026)
- SEEK NZ Employment Report (September 2026)
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
