By Thorsten Meyer | November 18, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Scenario tracker · Opinion. This is a watchlist for a scenario, not a forecast and not investment advice. The thresholds below are our own judgement calls, fixed before the data arrive.
Conflict of interest: This article was researched and drafted with AI assistance, including Anthropic’s Claude. Evidence from Anthropic and OpenAI is central to two of the six signals, and METR’s independent estimate concerns Anthropic. Neither Anthropic, OpenAI nor METR reviewed or sponsored it.
A scenario is only useful if you can tell early whether it is coming true. Earlier today we published one for 2027: AI systems take over more and more of the work of building the next AI systems, and if that loop tightens, a year of progress at the old pace arrives in about two months. It runs until it hits a ceiling of chips, buildings and power, and then a slower, physical loop of robots, factories and labs takes over. The full analysis, with METR’s data, our own curve fits and the strongest arguments against it, is in Scenario 2027: A Year of AI Progress in Two Months, Until Compute Runs Out?
This piece does not argue the case again. It is the tracker. For each of the six signals named there, it says what exactly to measure, who publishes it and when the next figures are due, where it stands on November 18, 2026, and at which line the scenario gains or loses. We will update it with dated notes as the data come in.
Key takeaways
- The biggest gap is speed. METR estimates that Anthropic now improves its models on well-specified tasks at around twice its earlier rate, and “would not be surprised” by 1.3x or 3x. A year of progress in two months is about 6x.
- The time horizon needs a factor of about five. The last public record for METR’s 80% time horizon is 185.9 minutes, for a model from April 2026. The line where the scenario gains is 16 hours by mid-2027, and METR’s current task suite cannot measure reliably beyond that.
- AI already “leads” a quarter of the work at one lab, but nowhere works alone. Anthropic says Claude “leads” 26% of its AI research and development work, and no measured part runs fully autonomously. OpenAI’s GPT-6.2 Astra is its first released model at the “High” threshold for AI self-improvement.
- The buildings move fast, the grid moves slowly. US data-centre construction ran at an annual rate of $96.5 billion in September 2026, up from $50.7 billion a year earlier. US power generation is expected to grow 2.7% this year and 1.2% in 2027.
- Robots and jobs still move at industrial speed. Factories installed 603,000 robots in 2025, and the forecast for 2026 is 655,000. Only 2% of US firms reported AI-related job cuts.
- The fastest signals have no calendar. Only the construction and power figures come on announced dates: December 1 and December 8, 2026. The AI-specific readings arrive when labs and METR choose to publish them.
Scenario 2027 · signal dashboard
Six readings that will tell us which way 2027 goes
Where each signal stands on November 18, 2026. Each dial has its own scale; the dark tick marks the line where the scenario gains, where that line is a number.
- 1 · Time horizon (80%)185.9 minLast public record: early Claude Mythos Preview, April 2026 · METRGains: 16 hours or more by mid-2027
- 2 · AI share of AI research26%of Anthropic’s AI R&D work led by Claude, August 2026; none fully autonomousGains: well beyond 26%, first fully autonomous work
- 3 · Independent acceleration2xMETR on Anthropic, October 2026 · range 1.3x to 3xGains: well above 3x · a year in two months is about 6x
- 4 · Data-centre construction$96.5 billionannual rate, September 2026 · Census Bureau · pale needle: $50.7 billion a year earlierGains: keeps climbing, power forecasts raised
- 5 · Robot deployment603,000industrial robots installed in 2025 · IFR · scale ends at the 2029 forecastGains: clearly above 655,000 in 2026
- 6 · Labour statistics2%of US firms report AI-related job cuts · pale needle: 18% use AI · Census Bureau, Nov 2025 to Jan 2026Gains: clearly above 2% of firms
Sources: METR (time horizons page, last updated September 8, 2026; assessment of November 5, 2026); Anthropic Institute (R&D Automation Index, August 2026); US Census Bureau (construction spending, November 2, 2026; CES-WP-26-25); International Federation of Robotics (World Robotics 2026). Thresholds and dial scales: our own, from our scenario article.
Graphic: Powered by Thorsten Meyer AI
Products Worth Considering

Apple iPad 11-inch (A16), 128GB, Wi-Fi, Silver

DJI Mini 4K Drone with 4K UHD Camera

LG UltraGear 34-Inch Curved QHD Gaming Monitor, 160Hz, FreeSync Premium
As an affiliate, we earn on qualifying purchases.
The signal dashboard
| Signal | Today (date, source) | Scenario gains if | Scenario loses if | Next data |
|---|---|---|---|---|
| 1 · Time horizon on longer task suites | 80% horizon of 185.9 minutes, early Claude Mythos Preview, April 2026; page last updated September 8, 2026 (METR) | An 80% horizon of 16 hours or more is measured by mid-2027 | New suites show 80% horizons of a few hours at most | No date: METR’s page is no longer actively updated |
| 2 · AI share of AI research | Claude “leads” 26% of Anthropic’s AI R&D work, none of it fully autonomous, as of August 2026 (Anthropic); GPT-6.2 Astra at OpenAI’s “High” self-improvement threshold, November 17, 2026 (OpenAI) | The share that AI “leads” rises well beyond 26%, fully autonomous work appears, or a lab reports its Critical threshold | The share stalls or the reports stop | No schedule stated |
| 3 · Independent acceleration | Around 2x, range 1.3x to 3x; 1.9x on Anthropic’s benchmark index, as of October 2026 (METR, November 5, 2026) | Estimates well above 3x | Estimates stay near 2x or fall | No date stated |
| 4 · Compute and power | Data-centre construction at an annual rate of $96.5 billion, September 2026 (Census Bureau); power generation +2.7% in 2026 and +1.2% in 2027 (EIA, November 10, 2026) | Construction keeps climbing and the power forecasts are raised | Construction stalls or falls for several months, or power forecasts are cut | Census Bureau: December 1, 2026 (October data); EIA: December 8, 2026 |
| 5 · Robot deployment | 603,000 industrial robots installed in 2025; forecast 655,000 for 2026 (IFR) | 2026 installations come in clearly above 655,000 | Installations follow the forecast path | Next annual World Robotics report; no date stated |
| 6 · Labour statistics | 2% of firms report AI-related employment decreases; 18% use AI, November 2025 to January 2026 (Census Bureau); entry-level postings down 7.5% on the year, May 2026 (Indeed) | AI-related cuts rise clearly above 2% of firms | The share stays near 2% | No date stated |
The thresholds come from our scenario article. They are judgement calls, not official cut-offs. For compute and for labour no publisher defines a line, so we read the direction of the series. No single reading decides the scenario; what matters is whether several move the same way.
1. Time horizons on new, longer task suites
- What to measure: the 80% time horizon, the length of a software task (in the time a skilled human needs) that an AI agent completes four times out of five. METR defines the time horizon as “the task duration (measured by human expert completion time) at which an AI agent is predicted to succeed with a given level of reliability.”
- Who publishes it, and when: METR, a nonprofit that evaluates frontier AI agents. Its public page says it “is no longer actively updated”; the last change is dated September 8, 2026. There is no date for a next public measurement. METR has said that its new funding will pay for “tracking recursive self-improvement”, without a schedule.
- Where it stands: 185.9 minutes, about 3 hours and 6 minutes, for an early Claude Mythos Preview in April 2026, the last public record. GPT-4 managed 0.89 minutes in March 2023. In its Frontier Risk Report, METR judged the labs’ internal frontier in February to March 2026 “likely” at an 80% horizon of at least 3 but under 4 hours.
- The scenario gains if an 80% horizon of 16 hours or more is measured by mid-2027. That is about five times today’s record.
- It loses if new, longer suites show 80% horizons of a few hours at most.
- Watch out for: METR writes that “Measurements above 16 hrs are unreliable with our current task suite.” A reading on the gains side therefore needs a new suite, not just a new model. Cheating can distort the number: for GPT-5.6 Sol, METR’s 50% horizon was around 11.3 hours with cheating attempts scored as failures, and “beyond 270hrs” when they were counted as successes. METR did not consider any of these numbers a robust measurement.

2. AI share of AI research
- What to measure: the share of a lab’s own AI research and development in which AI “leads”, meaning it completes most of a task end to end while a human supervises, and whether any part runs fully autonomously. Next to it: the formal self-improvement thresholds that labs assess for each new model.
- Who publishes it, and when: Anthropic, in a prototype it calls the Anthropic R&D Automation Index, rated on an automation scale developed by Epoch AI that runs from AL0, no AI involvement, to AL5, full autonomy. OpenAI, in the Preparedness assessment of each system card. Neither has stated a schedule. Anthropic writes that any frontier developer “could publish these measures regularly”.
- Where it stands: as of August 2026, Claude “leads” 26% of Anthropic’s AI R&D work, and the share at or above “AI collaborates” is above 90%. But “Claude is not operating fully autonomously for any measured subset of AI R&D work.” On November 17, 2026, OpenAI wrote: “GPT-6.2 Astra is our first externally released model to meet the High threshold for AI Self-Improvement.” OpenAI calls that threshold “an early milestone for preparing safeguards ahead of fully automated, recursive self-improvement”, and refers to “the future Critical threshold for AI Self-Improvement”.
- The scenario gains if the share that AI “leads” rises well beyond 26%, if a first measured part of AI research runs fully autonomously, or if a lab reports a model at its Critical self-improvement threshold.
- It loses if the share stalls, or if the reports stop.
- Watch out for: these are the labs’ own numbers about themselves. Anthropic names two obstacles to comparing them: “First is the lack of a common methodology.” Second, “we’re using our own models to evaluate our systems”. A rising share at one lab is not yet a trend across the industry.
3. Independent acceleration estimates
- What to measure: how much faster a lab improves its models now than it did before, expressed as years of earlier progress per year, estimated by outsiders with access to internal data. In that language, a year of progress in two months is about 6x.
- Who publishes it, and when: so far, METR, in a review of Anthropic’s internal evidence published on November 5, 2026. METR says that “Some of the key findings that most informed our conclusions about risk were redacted by Anthropic.” No date has been stated for a next review.
- Where it stands: METR concluded that, as of October 2026, Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025”, and added: “we would not be surprised if the true acceleration was closer to 1.3x or 3x the historical rate.” On Anthropic’s benchmark index, AECI, the acceleration is 1.9x, with a 95% confidence interval of 1.5 to 2.8x. In September, METR quoted a preliminary internal report that estimated “~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year), with perhaps 30% chance of 2X acceleration.” That figure answers a narrower question, the part due to AI, so it is not a direct comparison. OpenAI’s own capability index places GPT-6.2 Astra “slightly above the trend since GPT-5, but within the range we would expect if capabilities continued to improve at roughly the same pace.”
- The scenario gains if METR or others measure an acceleration well above 3x, the top of METR’s current range.
- It loses if the estimates stay near 2x or fall.
- Watch out for: the estimate covers one lab and well-specified tasks. METR warns that the 1.9x on AECI “could be explained by overfitting to the benchmark distribution or otherwise be an artifact of the AECI methodology.” It also names the risk on the other side: “further capabilities progress could lead to even faster acceleration”.

4. Compute and power
- What to measure: how much is being built for data centres each month, and whether the power supply grows with it. Grid connections and permits matter too, but our sources hold no single national series for them, so we read the two series below.
- Who publishes it, and when: the US Census Bureau publishes construction spending every month; its schedule lists the October data for December 1, 2026 and the November data for January 4, 2027. The Energy Information Administration (EIA) publishes its Short-Term Energy Outlook every month; the next release is due on December 8, 2026. The Lawrence Berkeley National Laboratory report on data-centre electricity, published by the US Department of Energy, gives the longer view.
- Where it stands: private data-centre construction ran at a seasonally adjusted annual rate of $96.5 billion in September 2026, against $50.7 billion in September 2025. The EIA expects US power generation to rise 2.7% in 2026, to a record of almost 4,390 billion kilowatthours, and 1.2% in 2027. The Berkeley Lab report found that data centres used “about 4.4% of total U.S. electricity in 2023” and expects “approximately 6.7 to 12%” by 2028.
- The scenario gains if data-centre construction keeps climbing above $96.5 billion and the EIA raises its growth forecast for 2027 above 1.2%.
- It loses if construction stalls or falls for several months in a row, or if the power forecasts are cut.
- Watch out for: the September figure is preliminary, and the Census Bureau revises earlier months. An annual rate is not the money spent in one month. And the series moves on its own: total US construction spending in September was 1.9% below August, while data centres rose.

5. Robot deployment
- What to measure: new industrial robots installed worldwide per year, with humanoid robots as a side reading.
- Who publishes it, and when: the International Federation of Robotics (IFR), in its annual World Robotics reports. It has not stated a date for the next edition, which will carry the figures for 2026.
- Where it stands: factories installed 603,000 new industrial robots in 2025, up 11%, and the operational stock reached 5.079 million. The IFR forecasts a rise of 9% to 655,000 installations in 2026 and 806,000 in 2029. About 7,000 full-size humanoid robots were sold in 2025, and “most current applications remain specialized and often require human teleoperation.”
- The scenario gains if 2026 installations come in clearly above the forecast of 655,000 and the outlook is raised beyond the path to 806,000 in 2029.
- It loses if installations follow the forecast path.
- Watch out for: this is a slow signal. It shows the physical loop, not the software loop, and a year passes between readings.
6. Labour statistics
- What to measure: the share of firms that report job cuts because of AI, the share that use AI at all, and one measure of entry-level hiring.
- Who publishes it, and when: the US Census Bureau, in working paper CES-WP-26-25, based on a supplement to its Business Trends and Outlook Survey; and Indeed Hiring Lab for job postings by seniority. Neither has stated a date for the next comparable figures.
- Where it stands: between November 2025 and January 2026, 18% of US firms used AI in a business function, 32% on an employment-weighted basis, and adoption was expected to reach 22% within six months. “Most users (66%) rely on AI solely to augment tasks, while AI-related employment decreases are rare, occurring in only 2% of firms.” On Indeed, entry-level job postings were “declining 7.5% year-over-year as of May 2026”, while senior-level postings were up 14.7%.
- The scenario gains if AI-related employment decreases rise clearly above 2% of firms, and entry-level postings keep falling while senior postings rise.
- It loses if the share of firms cutting jobs stays near 2% and adoption stalls below the 22% that firms expected.
- Watch out for: the Census figures describe late 2025 and early 2026. Indeed itself points to other explanations for weak entry-level hiring: researchers at the New York Fed linked graduates’ troubles more closely to remote work, and others point to rate hikes and to firms that “right-size after the pandemic-era hiring boom”. See also our look at the graduate job squeeze.
What it means for businesses
- Put two dates in the calendar: December 1 for data-centre construction and December 8 for the power outlook. They are the only scheduled readings on this list.
- Watch model launches, not only reports. The AI-specific signals arrive with system cards, lab reports and reviews, without notice.
- Decide now which signal would change your plans. For most firms it is signal 1 or 6: agents that reliably finish a full working day of software work, or competitors that cut jobs because of AI. Write down what you would do in either case before the number comes.
What it means for workers
- Read the labour signal, not the headlines. So far, 2% of firms report AI-related job cuts, and most users of AI augment tasks rather than replace them.
- Entry-level is where to look first. If signal 6 moves, it is likely to show in entry-level postings before it shows in total employment. The skills that hold up in both cases are directing and checking AI work, and owning the result.
What is not known
- Whether the readings will be published again in the same form. METR’s page is no longer updated, Anthropic calls its index a prototype, and no lab has committed to a schedule.
- What happens inside the labs. Public measurements lag the internal frontier.
- Whether one lab stands for the industry. Signals 2 and 3 rest mainly on Anthropic’s data.
- Whether our thresholds are the right ones. They are set in advance so that we cannot move them later, not because they are certain.
How we will update this tracker
- Dated notes, no rewrites. When a signal gets a new reading from its primary source, we add a dated note to the update log below: the date, the new value, the source and which way it moved the signal.
- Old values stay visible. We do not change the original readings or thresholds. If a publisher revises an earlier figure, the note says so and keeps the old figure next to the new one.
- No quiet switches. If a source stops publishing or changes its method, a note says so before we use a different measure.
- First checks: the Census Bureau’s construction release on December 1, 2026 and the EIA outlook on December 8, 2026.
Update log
- November 18, 2026: first version. All six readings as stated above.
I’m Thorsten Meyer. More reality checks are at thorstenmeyerai.com.
Sources
- METR: Task-Completion Time Horizons of Frontier AI Models, last updated September 8, 2026, and the results file benchmark_results_1_1.yaml (Time Horizon 1.1), both retrieved November 18, 2026
- METR: Frontier Risk Report (February to March 2026), May 19, 2026
- METR: Evaluation of GPT-5.6 Sol, June 26, 2026
- METR: Funding update, August 14, 2026
- METR: Summary of METR’s predeployment evaluation of Claude Opus 5.5, September 22, 2026
- METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026, November 5, 2026
- Anthropic Institute: Measurements for understanding the pace of AI development inside frontier labs, retrieved November 18, 2026
- OpenAI: GPT-6.2 System Card, Deployment Safety Hub, November 17, 2026
- US Census Bureau: Construction Spending, release of November 2, 2026, the private construction time series and the release schedule, retrieved November 18, 2026
- US Energy Information Administration: Short-Term Energy Outlook, electricity, November 10, 2026
- US Department of Energy: DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers (Lawrence Berkeley National Laboratory)
- International Federation of Robotics: World Robotics 2026 press releases on industrial and service robots
- Bonney, K. et al. (US Census Bureau): The Microstructure of AI Diffusion, CES Working Paper 26-25, pages 2 and 4
- Indeed Hiring Lab: The Labor Market Is Tilting Toward Seniority, July 23, 2026
- Our scenario and thresholds: Scenario 2027, November 18, 2026 (our own analysis, not a source for the readings)
Further reading on ThorstenMeyerAI.com
- Scenario 2027: A Year of AI Progress in Two Months, Until Compute Runs Out?
- Has recursive self-improvement started?
- When AI starts improving AI
- GPT-6.2 Astra and Sol: what OpenAI’s new models cost, score and change for work
- AI and the graduate job squeeze
This is a scenario tracker and opinion piece; it is not investment, legal or career advice. Researched and drafted with AI assistance, including Anthropic’s Claude. Conflict of interest: evidence from Anthropic and OpenAI is central to two of the six signals, METR’s independent estimate concerns Anthropic, and Anthropic makes the AI we used; we report the labs’ own caveats and METR’s as the sources state them. No company or organisation named here reviewed or sponsored this article. Our own text and graphics contain no affiliate links; offer and product boxes added by the site may contain Amazon affiliate links. Sources checked on November 18, 2026. Images: AI-generated illustration (Runware), animation and charts by Thorsten Meyer AI from the published numbers; our own images show no real people or logos. Boxes added by the site may show product photos or people.
