On 10 September 2026 the Anthropic Institute published Economic Scenarios for Transformative AI (Korinek, Jones, Sacher, Cotter, and McCrory; Working Paper No. 2026-02). The same day it shipped an interactive explorer. The extreme row is the one in the headlines: GDP growth at 15.4 percent a year, and labour’s share of income at 45.2 percent.
Cipher Projects is an Australian engineering studio. We ship production agents with named jobs, evals, and a halt on writes. We do not forecast the US macro path, and we do not sell universal basic income.
Quick answer
The extreme scenario is a bigger pie and a falling labour share: 2030 GDP 32.4 percent above the no-AI path, labour’s slice 60 percent to 45.2 percent. An Australian SME’s move in the next 12 months is not to wait for UBI. It is to point agents at named work with evals. A copilot without a named job is a faster intern. Ten production workflows with a pass/fail test capture some of the 15 percent path. We build those workflows.
Best for: operators who read “15 percent growth and mass unemployment” and need a company action, not a policy seminar. Honest limit: the model is the United States. Australia will not reprint these levels. Cipher is not a macro shop. We specify work, put evals on it, and price stamp plus agent plus connector. We do not certify you, and we do not sell a singularity dividend.
Last updated: 12 September 2026. Plain-English neighbours live on the cluster glossary. If a term cannot survive that page, it does not belong in a board pack.
What did Anthropic model for GDP and jobs by 2030?
Three US paths to 2030 against a no-AI baseline, in 2025 prices. Modest adds almost nothing. Substantial doubles the usual growth rate. Extreme puts 2030 growth at 15.4 percent a year and cuts cognitive employment by 21.5 percent.
The paper maps AI capability and adoption onto GDP, wages, the labour share, and unemployment. Cognitive work is management, professional, sales, and office jobs. Other occupations are not directly exposed. The authors attach no probabilities. Table 3 is the 2030 snapshot. GDP dollar levels are from the explorer.
| 2030 (vs no-AI path) | Modest | Substantial | Extreme |
|---|---|---|---|
| GDP above no-AI | +1.6% ($34.1T) | +8.3% ($36.3T) | +32.4% ($44.4T) |
| GDP growth, percent a year | 2.4 | 5.4 | 15.4 |
| Labour share of income | 59.4% | 56.1% | 45.2% |
| Cognitive employment since mid-2026 | −0.5% | −3.9% | −21.5% |
| Unemployment, cognitive workers | 2.9% | 4.5% | 17.9% |
| Unemployment, all workers | 3.9% | 4.6% | 11.9% |
| Cognitive wage vs no-AI | +0.4% | −0.3% | −11.5% |
| Other wages vs no-AI | +1.1% | +5.9% | +33.6% |
| Capital income vs no-AI | +3.1% | +18.9% | +81.4% |
The no-AI path grows at 2.0 percent a year. Labour starts at 60 percent of income. In the extreme row, total labour income in 2030 is almost exactly what it would have been without AI: the share falls by about a quarter while GDP rises by about a third. All of the extra output accrues as capital income.
The explorer’s colour on capability: modest looks like the internet. Substantial can do about half of knowledge work by 2030, but most instances still run without AI. Extreme is more productive than humans at the vast majority of knowledge-work tasks and does nearly all of them on its own. A Morning Consult survey of US adults, fielded in August, maps the median respondent onto outcomes near substantial: GDP roughly 8 percent higher, cognitive employment down about 4 percent.
Euronews summarised the same table the next day. Use the paper for the numbers. Use the news piece only as a secondary walkthrough.
Is the 15% path a forecast?
No. The authors write that the scenarios are not predictions and that they attach no probabilities to them.
The model is a coordinate system. You pick capability, adoption, autonomy, productivity, and how hard it is to switch jobs. The explorer returns GDP, wages, and unemployment. Extreme sits beyond the published external forecasts the paper names (Acemoglu, OECD, Goldman Sachs, McKinsey, Penn Wharton). It is a thought experiment with a table, not a 2030 commitment.
The model also omits policy, business cycles, demand from the data-centre buildout, and catastrophic risk. If you need the civilisation-risk thread, that is a different page: what p(doom) means. This page is the income split and the company move.
What should an Australian SME do in the next 12 months?
Point agents at named work you can evaluate. Do not wait for UBI, and do not treat a seat of Copilot as the 15 percent path.
The paper’s mechanism is blunt. AI raises output and reassigns cognitive tasks to capital. Labour’s slice shrinks unless humans specify new useful work. That is a national-accounts sentence. The company version is the cutover we already use when we price production agents: if the job is still a chat, you bought a faster intern. If you can name the workflow, the eval, the connector, and the halt, you captured some of the gain.
A Claude Project or a ChatGPT workflow can prove the logic. It does not give you SSO, an audit trail, least-privilege connectors, or a quote that survives agent three. The cutover is written in Claude workflow vs production agent: keep iterating in the vendor UI until the tool list is real; then harden onto a platform once. We quote stamp, per agent, and per connector. A typical two-agent job starts from AUD $40,000 on /pricing/.
If your product is a cheap tool sold to other builders on search, the squeeze is already visible. That is a channel shift, not a collapse: Levelsio vs Greg Isenberg. The Millennium Prize week is the same lesson at a larger scale: the hard part is specifying the question, not hiring more genius. Read AI claimed a Millennium Prize.
Which ten production workflows can you name with evals?
Ten named jobs, each with a pass/fail test and a write that a human can stop. If you cannot fill the last two columns, the row is still a demo.
Cipher’s first-hand rule from quoting agent work: the fastest way to waste a copilot seat is to leave the job unnamed. The fastest way to blow a cheap agent quote is to bury two connectors inside one blob. Separate the workflow, the eval, and the halt.
| # | Workflow | Eval that must pass | Human halt |
|---|---|---|---|
| 1 | Invoice to PO match | Match rate vs the last 100 paid invoices; miss-rate on a variance over $50 | Payment release |
| 2 | Overdue AR chase draft | Human accept rate on 30 drafts from the ageing report | Send to the customer |
| 3 | Support classify and draft | Wrong-queue rate on one labelled week; every draft cites a KB row | Refund or account delete |
| 4 | RFQ first draft from CRM | Missing-field rate vs your quote checklist on 20 past jobs | Customer send |
| 5 | Job card from inbound email | Required fields complete vs a staff sample of 50 emails | Dispatch / schedule |
| 6 | Leave request vs award rules | False-approve count on a 20-case fixture | Payroll write |
| 7 | Privacy request intake | Clock starts the same business day; no missed deadline on a fixture | Reply to the requester |
| 8 | Reorder suggestion | Stockout and overstock vs last quarter’s 90-day velocity | Purchase-order send |
| 9 | Meeting notes to CRM actions | Action recall vs human notes on 15 calls | External email |
| 10 | Incident first pass | False-rollback on 10 labelled incidents | Production mutate |
Pick the ten that match your firm. Drop a row that has no system of record. A row without an eval is a prompt. A row without a halt is a write you cannot defend. We will not invent the list on a scope call. Bring the names.
How does today’s GDPNow compare with 15% growth?
It does not. The Atlanta Fed’s GDPNow nowcast for 2026:Q3 was 4.4 percent (annualised) on 10 September 2026. The extreme scenario’s 15.4 percent is a 2030 year-rate on one modelled path, not this quarter’s print.
GDPNow is a running estimate of what the Bureau of Economic Analysis will later publish. It is not an official Atlanta Fed forecast. The model applies no judgemental adjustment. The next scheduled update after 10 September is 16 September 2026. Check the live figure on the GDPNow page before you quote it.
The official second estimate for 2026:Q2 is 1.5 percent (annualised), published by the BEA. That is the recent history. 4.4 percent is a nowcast. 15.4 percent is a scenario. Do not stack them as a trend line.
Does a bigger pie mean a bigger paycheque?
No. In the extreme row the pie is 32.4 percent larger and labour’s share falls from 60 percent to 45.2 percent. Cognitive wages sit 11.5 percent below the no-AI path. Cognitive unemployment is 17.9 percent.
Average wages still rise (+9.7 percent vs no-AI) because other occupations jump +33.6 percent. The gain is not in the office. Capital income is 81.4 percent above the no-AI path. The paper is explicit: transfers could in principle leave everyone better off, because the economy’s gain is larger than the cognitive wage-bill loss. That sentence is policy. It is not a product Cipher sells.
For a firm, the local version of “labour share” is whether the extra output shows up as margin you own or as a seat licence you rent. A copilot that drafts and a human that pastes is rented speed. A named workflow with an eval and a halt is work you can keep.
Should a company wait for UBI?
No. UBI, a “singularity dividend,” or universal high income are political responses if the extreme path arrives. They are not a 12-month operating plan, and Cipher does not sell them.
Anthropic’s own policy writing treats income replacement as a later-tier question for governments, not as a line item on an SME invoice. Your board cannot legislate a dividend. It can name ten workflows, put evals on them, and decide which writes need a person. If the country later taxes capital and pays a transfer, that does not undo a year of unnamed busywork.
If you want the work built, that is applied AI engineering. Bring the list.
FAQ
Anthropic modelled 15% GDP growth and a falling labour share. What should an Australian SME do in the next 12 months? Name ten production workflows, put a pass/fail eval on each, and put a human halt on every write to a system of record. Do not wait for UBI. Do not treat a copilot seat as the 15 percent path. Cipher builds those workflows; we do not sell a dividend.
Are the Anthropic scenarios forecasts? No. Korinek et al. (2026) say the scenarios are not predictions and attach no probabilities. Extreme is the upper coordinate, not a base case. The median survey respondent lands near substantial.
What is labour share, and why does it fall when GDP rises? Labour share is the fraction of national income paid to workers. In the extreme row it falls from 60 percent to 45.2 percent because AI reassigns cognitive tasks to capital. Total labour income stays about flat versus the no-AI path. The extra GDP is capital income (+81.4 percent).
Does a bigger pie mean a bigger paycheque for office staff? Not on the extreme path. Cognitive wages are 11.5 percent below the no-AI line. Cognitive unemployment is 17.9 percent. Other occupations see the wage gain.
Should we budget for UBI or a singularity dividend? No. That is a government question if disruption reaches those levels. A company budget for the next 12 months is named work, evals, connectors, and a halt. Cipher will not invoice you for UBI.
How is a Claude workflow different from a production agent? The vendor UI proves the logic. Production needs your identity, isolation, connectors, audit, cost control, and a quote that survives the third agent. Full cutover: Claude workflow vs production agent. Commercial lines: how we price production AI agents.
Who should specify the ten workflows? The operators who already know the jobs. Cipher will help you turn a named list into stamp, agents, and connectors. We will not invent the list from a slogan. Australian-led; delivery across Australia and Vietnam.
Sources
- Anthropic Institute, Economic Scenario Explorer: canonical product page; GDP dollar levels at 2025 prices ($34.1T / $36.3T / $44.4T); capability colour for the three paths; labour-versus-capital split.
- Korinek, Jones, Sacher, Cotter, and McCrory, “Economic Scenarios for Transformative AI,” Anthropic Institute Working Paper No. 2026-02, September 2026: Table 3 numbers used above; “not predictions”; no probabilities; labour income identity in the extreme row.
- Euronews, 11 September 2026: Quirino Mealha. Secondary walkthrough of the same three scenarios.
- Atlanta Fed GDPNow: 4.4 percent nowcast for 2026:Q3, updated 10 September 2026; next update 16 September 2026. Not an official Atlanta Fed forecast.
- BEA, GDP (Second Estimate), 2026:Q2: real GDP +1.5 percent annualised.
Related: Buzzword soup glossary · How we price production AI agents · Claude workflow vs production agent · Is AI killing indie hackers? · AI claimed a Millennium Prize · What p(doom) means