# Capabilities — Autonomy monitor ## 18 of 49 models with a published task horizon have a public compute estimate. METR measures how long a task can be before a model stops finishing it half the time. The measure is the closest thing the field has to a scalar for autonomous capability, and it is routinely read against training compute — the quantity every compute-threshold regime in force actually gates on. That reading is available for **18** of the **49** models METR publishes. It is not available for any of the ten highest. Inside the range METR states its task suite measures reliably, the most capable model finishes tasks of 12.0 h; the most capable one with a public training-compute estimate finishes tasks of 3.4 h, a factor of 3.5 below it. One measurement sits above that range. METR's note on the dataset this view draws reads _“Measurements above 16 hrs are unreliable with our current task suite”_, and METR excludes such points from its own trend fit. Claude Mythos (Preview) is drawn here rather than dropped, at 17.4 h — which would put the frontier a factor of 5.1 above the compute record instead of 3.5. Both figures are on this page because the gap between them is the same fact the page is about, one instrument further out: the frontier has run past not only the public compute record but past the measurement. Unit Horizons are in minutes — METR's own publication unit — at 50% success against the same task performed by a human professional. A horizon of an hour is not “an hour of work”; it is the length at which the model still succeeds half the time. Vintage METR data through 2026-04-07, pulled 2026-05-09. The pull date is Scrutica's, not METR's last release — METR republishes irregularly, so this view refreshes on release rather than on a blind schedule. The published set is mixed-vintage and counted as such: 16 rows draw values from METR-Horizon-v1.1, 24 retain METR-Horizon-v1.0 values METR has not re-released, and for 9 the release cannot yet be established from the records Scrutica holds. Cite CSVJSONJSON+Prov ### Both axes, and the models that only have one Scatter plot of 49 models: METR task-horizon estimate on a logarithmic vertical axis against estimated training compute on a logarithmic horizontal axis. 18 models have a public training-compute estimate and are drawn as filled marks in the field with their published confidence interval. The remaining 31 have none and are drawn as hollow rings on a rail at the left, at their true horizon. A rule at 3.4 h marks where the public compute record stops — GPT-5 (high) is the highest horizon in the set that can be placed on the compute axis. A second rule at 16 hours marks the horizon above which METR states its own measurements are unreliable. The frontier of the horizon axis is held entirely by models on the rail.6 sec1.0 min10 min1.0 h10.0 hTask horizon10²¹10²²10²³10²⁴10²⁵10²⁶10²⁷Estimated training compute (FLOP)EU AI Act 10²⁵EO 14110 10²⁶rescinded 202516 h · METR calls measurement above here unreliablepublic compute record stops here — 3.4 hno figure31 of 49at or above 1.0 hClaude Mythos (Preview) (Anthropic) — task horizon 17.4 h (METR's published interval 8.5 h to 55.1 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis. Above the range METR states its current task suite measures reliably.Claude Opus 4.6 (Anthropic) — task horizon 12.0 h (METR's published interval 5.3 h to 60.6 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.Gemini 3.1 Pro (Google DeepMind) — task horizon 6.4 h (METR's published interval 3.9 h to 11.6 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-5.2 (high) (OpenAI) — task horizon 5.9 h (METR's published interval 3.3 h to 13.6 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-5.3 Codex (OpenAI) — task horizon 5.8 h (METR's published interval 3.2 h to 13.6 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-5.4 (OpenAI) — task horizon 5.7 h (METR's published interval 3.1 h to 12.8 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4.5 (Anthropic) — task horizon 4.9 h (METR's published interval 2.7 h to 10.4 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4.5 (16K) (Anthropic) — task horizon 4.8 h (METR's published interval 1.8 h to 20.4 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.Gemini 3 Pro (Google DeepMind) — task horizon 3.7 h (METR's published interval 2.3 h to 6.3 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-5.1 Codex Max (OpenAI) — task horizon 3.7 h (METR's published interval 2.2 h to 6.6 h; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Sonnet 4.5 (16K) (Anthropic) — task horizon 2.0 h (METR's published interval 56 min to 4.2 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.o3 (OpenAI) — task horizon 2.0 h (METR's published interval 1.2 h to 3.2 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4.1 (16K) (Anthropic) — task horizon 1.9 h (METR's published interval 58 min to 3.5 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4.1 (Anthropic) — task horizon 1.7 h (METR's published interval 59 min to 2.7 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4 (Anthropic) — task horizon 1.7 h (METR's published interval 60 min to 2.7 h; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.o3 (medium) (OpenAI) — task horizon 1.5 h (METR's published interval 46 min to 2.7 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Opus 4 (16K) (Anthropic) — task horizon 1.4 h (METR's published interval 45 min to 2.4 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.o4-mini (medium) (OpenAI) — task horizon 1.3 h (METR's published interval 35 min to 2.5 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.Claude Sonnet 4 (16K) (Anthropic) — task horizon 1.2 h (METR's published interval 38 min to 2.2 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.o1 (medium) (OpenAI) — task horizon 39 min (METR's published interval 21 min to 1.1 h; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.Gemini 2.5 Pro (Google DeepMind) — task horizon 39 min (METR's published interval 19 min to 1.2 h; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.Claude 3.5 Sonnet (Oct) (Anthropic) — task horizon 21 min (METR's published interval 10 min to 41 min; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.o1 Preview (OpenAI) — task horizon 20 min (METR's published interval 12 min to 33 min; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4o (Nov 2024) (OpenAI) — task horizon 9.2 min (METR's published interval 4.2 min to 18 min; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4o (Aug 2024) (OpenAI) — task horizon 7.0 min (METR's published interval 4.0 min to 13 min; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4 Turbo (Apr 2024) (OpenAI) — task horizon 6.6 min (METR's published interval 3.2 min to 12 min; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4 (Jan 2024) (OpenAI) — task horizon 5.4 min (METR's published interval 2.8 min to 9.8 min; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4 (Nov 2023) (OpenAI) — task horizon 4.0 min (METR's published interval 1.9 min to 8.4 min; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.Claude 3 Opus (Anthropic) — task horizon 4.0 min (METR's published interval 1.7 min to 8.8 min; no confidence level is published for it). Values from release not yet attributed. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-4 Turbo (OpenAI) — task horizon 3.7 min (METR's published interval 1.9 min to 6.7 min; no confidence level is published for it). Values from METR-Horizon-v1.1. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-3.5 Turbo (OpenAI) — task horizon 36 sec (METR's published interval 14 sec to 59 sec; no confidence level is published for it). Values from METR-Horizon-v1.0. No public training-compute estimate, so it cannot be placed on the compute axis.GPT-5 (high) (OpenAI) — task horizon 3.4 h (METR's published interval 1.9 h to 6.8 h; no confidence level is published for it). Values from METR-Horizon-v1.1. Estimated training compute 6.6×10²⁵ FLOP.GPT-5 (medium) (OpenAI) — task horizon 2.3 h (METR's published interval 1.1 h to 4.5 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 6.6×10²⁵ FLOP.Grok 4 (xAI) — task horizon 1.8 h (METR's published interval 48 min to 3.9 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 5.0×10²⁶ FLOP.Claude 3.7 Sonnet (Anthropic) — task horizon 1.0 h (METR's published interval 33 min to 1.7 h; no confidence level is published for it). Values from METR-Horizon-v1.1. Estimated training compute 3.4×10²⁵ FLOP.Claude 3.7 Sonnet (16K) (Anthropic) — task horizon 56 min (METR's published interval 29 min to 1.6 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 3.4×10²⁵ FLOP.Kimi K2 (Moonshot) — task horizon 54 min (METR's published interval 25 min to 1.6 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 4.2×10²⁴ FLOP.GPT-OSS 120B (OpenAI) — task horizon 42 min (METR's published interval 19 min to 1.4 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 4.9×10²⁴ FLOP.DeepSeek R1 (0528) (DeepSeek) — task horizon 31 min (METR's published interval 13 min to 1.1 h; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 4.0×10²⁴ FLOP.DeepSeek R1 (DeepSeek) — task horizon 27 min (METR's published interval 13 min to 51 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 3.5×10²⁴ FLOP.DeepSeek V3 (0324) (DeepSeek) — task horizon 23 min (METR's published interval 12 min to 41 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 3.3×10²⁴ FLOP.DeepSeek V3 (DeepSeek) — task horizon 18 min (METR's published interval 9.0 min to 34 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 3.3×10²⁴ FLOP.Claude 3.5 Sonnet (Jun) (Anthropic) — task horizon 11 min (METR's published interval 5.5 min to 22 min; no confidence level is published for it). Values from release not yet attributed. Estimated training compute 2.7×10²⁵ FLOP.GPT-4 (Mar 2023) (OpenAI) — task horizon 5.4 min (METR's published interval 2.5 min to 9.7 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 2.1×10²⁵ FLOP.Qwen 2.5 72B (Alibaba) — task horizon 5.2 min (METR's published interval 2.4 min to 10 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 7.8×10²⁴ FLOP.GPT-4 (Jun 2023) (OpenAI) — task horizon 4.0 min (METR's published interval 2.0 min to 7.9 min; no confidence level is published for it). Values from METR-Horizon-v1.1. Estimated training compute 2.1×10²⁵ FLOP.Qwen 2 72B (Alibaba) — task horizon 2.2 min (METR's published interval 51 sec to 4.8 min; no confidence level is published for it). Values from METR-Horizon-v1.0. Estimated training compute 3.0×10²⁴ FLOP.Davinci 002 (OpenAI) — task horizon 8 sec (METR's published interval 5 sec to 13 sec; no confidence level is published for it). Values from release not yet attributed. Estimated training compute 3.1×10²³ FLOP.GPT-2 XL (OpenAI) — task horizon 3 sec (METR's published interval 1 sec to 9 sec; no confidence level is published for it). Values from release not yet attributed. Estimated training compute 1.9×10²¹ FLOP.compute estimate on record (18)no compute estimate · drawn on the railMETR's published interval Both axes are logarithmic. Rail marks sit at their true horizon and are offset sideways only where they would otherwise print on top of one another, so all 31 can be counted; lateral position on the rail means nothing. Horizons are METR's 50%-success estimates in minutes, with METR's own interval as the whisker — METR publishes no confidence level for it, so none is stated here. _The compute axis is an estimate throughout: Epoch publishes training-FLOP figures as derived values, not as disclosures, and no lab publishes the number directly._ Sources: METR task-horizon dataset; Epoch AI training-run estimates. Horizon cut1.0 h **23** of 49 models complete tasks of this length at 50% reliability. **4** of those have a public training-compute estimate, so 17% of the models above this line can be placed on the compute axis. ### What the fitted relationship can and cannot carry Regressing log horizon on log training compute over the **18** models that have both numbers gives an R² of **0.66**: each decade of training compute buys a factor of 4.8 in task horizon, at a fitted slope of 0.685 ± 0.124 log-minutes per decade. That is a real relationship and a genuinely loose one — at a given compute level the residual spread is a factor of 3.8 either way, which is what you would expect when post-training, scaffolding and elicitation effort all move the measured value and none of them is in the fit. Two things weaken the fit further and both are visible in its own operands. It spans 5.4 decades of compute, from 1.9×10²¹ to 5.0×10²⁶ FLOP, and the plate above shows that span is not evenly populated. And the 18 points sit at only **14** distinct compute values, because configuration variants of one training run share an estimate: the fit treats them as independent observations, and they are not. The binding limitation is still the range. The fit is anchored on models topping out at 3.4 h, and the frontier — even inside METR's stated measurement range — is at 12.0 h. Every claim of the form “this much compute buys an agent that works unsupervised for N hours” at frontier N is extrapolation past the last observation, not interpolation within the data. That a tighter fit makes this worse rather than better is the point: a well-fitting model extrapolated past its last observation is the more seductive error. METR's own headline is a claim about _time_ rather than compute, measured across the whole set, and it is the better-founded of the two readings for exactly that reason. The dataset this view cites publishes it in its own header: the horizon doubles every **4.23 months** over models released from 2023 on (3.43–5.19 months), and every 6.17 months over the whole published record. Read the two together with care: METR fits the doubling time with the above-16-hour points excluded, and the frontier horizon on this page includes them, so the two quantities here are computed under different inclusion rules. Not drawnA regression line and prediction band over the covered subset appeared on the previous version of this view. They are stated here instead of drawn: a band rendered across a compute axis whose upper reaches contain no frontier observation invites exactly the extrapolation the paragraph above warns against. ### The models above an hour with nothing on the compute axis An hour is METR's legible cut-point. **23** models clear it; **19** of them have no public training-compute estimate — the set any compute-to-capability argument at the frontier has to reach over. The other 4 are the whole of the compute record above an hour. Compute on recordAll 19 rows: none published. Model Developer Task horizon METR interval METR release Claude Mythos (Preview) · above METR's stated measurement range Anthropic 17.4 h 8.5 h – 55.1 h METR-Horizon-v1.1 Claude Opus 4.6 Anthropic 12.0 h 5.3 h – 60.6 h METR-Horizon-v1.1 Gemini 3.1 Pro Google DeepMind 6.4 h 3.9 h – 11.6 h METR-Horizon-v1.1 GPT-5.2 (high) OpenAI 5.9 h 3.3 h – 13.6 h METR-Horizon-v1.1 GPT-5.3 Codex OpenAI 5.8 h 3.2 h – 13.6 h METR-Horizon-v1.1 GPT-5.4 OpenAI 5.7 h 3.1 h – 12.8 h METR-Horizon-v1.1 Claude Opus 4.5 Anthropic 4.9 h 2.7 h – 10.4 h METR-Horizon-v1.1 Claude Opus 4.5 (16K) Anthropic 4.8 h 1.8 h – 20.4 h METR-Horizon-v1.0 Gemini 3 Pro Google DeepMind 3.7 h 2.3 h – 6.3 h METR-Horizon-v1.1 GPT-5.1 Codex Max OpenAI 3.7 h 2.2 h – 6.6 h release not yet attributed Claude Sonnet 4.5 (16K) Anthropic 2.0 h 56 min – 4.2 h METR-Horizon-v1.0 o3 OpenAI 2.0 h 1.2 h – 3.2 h METR-Horizon-v1.1 Claude Opus 4.1 (16K) Anthropic 1.9 h 58 min – 3.5 h METR-Horizon-v1.0 Claude Opus 4.1 Anthropic 1.7 h 59 min – 2.7 h METR-Horizon-v1.1 Claude Opus 4 Anthropic 1.7 h 60 min – 2.7 h METR-Horizon-v1.1 o3 (medium) OpenAI 1.5 h 46 min – 2.7 h METR-Horizon-v1.0 Claude Opus 4 (16K) Anthropic 1.4 h 45 min – 2.4 h METR-Horizon-v1.0 o4-mini (medium) OpenAI 1.3 h 35 min – 2.5 h METR-Horizon-v1.0 Claude Sonnet 4 (16K) Anthropic 1.2 h 38 min – 2.2 h METR-Horizon-v1.0 ### What a real site could accumulate The compute axis above is an attribute of models. This is the same axis read as an attribute of places: how long a facility Scrutica tracks would need to run to accumulate a training budget the size of the largest run on the figure above that has a public estimate — GPT-5, 6.6×10²⁵ FLOP, Epoch AI. It is not the largest figure on that plate: Grok 4 sits at 5.0×10²⁶ FLOP. “Frontier scale” here means the frontier of the public compute record, which is the smaller thing and the page's whole subject. Scrutica publishes 4,257 facility records. **512** of them have either a disclosed GPU count or a nameplate power figure, which is what the estimator needs — so this table is a 12.0% slice of the published substrate, and the other 3,745 are absent from it for want of a disclosure rather than for want of capacity. Of the 512, **151** clear the reference budget in 90 days of continuous training; the 24 fastest are listed. Rows are facility records, not campuses: where a site is recorded as more than one facility, each appears on its own line with its own figures. Which evidence The estimation path is named per row because a hardware-derived budget (disclosed GPU count) and a power-derived one (nameplate MW) are different grades of evidence, and on the hardware path the accelerator is named too — 405 rows take that path, of which 405 disclose no accelerator model and are scored at the estimator's H100 SXM5 default. The estimator has a third path, from total investment, and it cannot apply here: Scrutica's substrate has no per-facility total-investment figure. That absence is itself a substrate-coverage fact, which is this page's subject. Facility Country Est. daily FLOP Days to reference budget Range across the estimator's assumptions [Stargate UAE (OpenAI/G42/Oracle)](/facilities/fac-stargate-uae) · estimated AE 8.8×10²⁵ power <1 <1–4 [Meta Louisiana Datacenter](/facilities/fac-epoch-gc-meta-louisiana-datacenter) · estimated US 4.0×10²⁵ power 2 <1–8 [AWS Project Rainier (New Carlisle, IN)](/facilities/fac-aws-new-carlisle) · estimated US 3.9×10²⁵ power 2 <1–8 [Meta Richland Parish (Hyperion)](/facilities/fac-meta-richland-parish) · estimated US 3.5×10²⁵ power 2 <1–9 [Crusoe Cheyenne Wyoming](/facilities/fac-epoch-gc-crusoe-cheyenne-wyoming) · estimated US 3.2×10²⁵ power 2 1–10 [DataVolt Neom 1.5 GW Phase 2](/facilities/fac-epoch-gc-datavolt-neom-1-5-gw-phase-2) · estimated SA 2.7×10²⁵ power 2 1–12 [IREN Sweetwater 1](/facilities/fac-epoch-gc-iren-sweetwater-1) · estimated US 2.5×10²⁵ power 3 1–13 [Oracle Vantage Data Centers Frontier](/facilities/fac-epoch-gc-oracle-vantage-data-centers-frontier) · estimated US 2.5×10²⁵ power 3 1–13 [Stargate Abilene (OpenAI/Oracle/SoftBank)](/facilities/fac-stargate-abilene) · estimated US 2.1×10²⁵ power 3 2–16 [OpenAI/Microsoft Mt Pleasant, Wisconsin Phase 2](/facilities/fac-epoch-gc-openai-microsoft-mt-pleasant-wisconsin-phase-2) · estimated US 2.0×10²⁵ hardware · H100 SXM5 assumed 3 2–9 [Crusoe Goodnight in Claude, Texas](/facilities/fac-epoch-gc-crusoe-goodnight-in-claude-texas) · estimated US 1.8×10²⁵ power 4 2–19 [1 Gig Data Center East Fishkill, NY](/facilities/gridstatus-nyiso-1738) US 1.8×10²⁵ power 4 2–19 [xAI Colossus 2 Memphis Phase 2](/facilities/fac-epoch-gc-xai-colossus-2-memphis-phase-2) · estimated US 1.6×10²⁵ hardware · H100 SXM5 assumed 4 3–12 [Colossus 2](/facilities/fac-epoch-xai-colossus-2) · estimated US 1.5×10²⁵ hardware · H100 SXM5 assumed 4 3–12 [Fluidstack France Gigawatt Campus](/facilities/fac-epoch-gc-fluidstack-france-gigawatt-campus) · estimated FR 1.5×10²⁵ hardware · H100 SXM5 assumed 5 3–13 [Meta Prometheus New Albany](/facilities/fac-epoch-gc-meta-prometheus-new-albany) · estimated US 1.5×10²⁵ hardware · H100 SXM5 assumed 5 3–13 [IREN Childress](/facilities/fac-epoch-gc-iren-childress) · estimated US 1.3×10²⁵ power 5 2–25 [Reliance Industries Supercomputer](/facilities/fac-epoch-gc-reliance-industries-supercomputer) · estimated IN 1.3×10²⁵ hardware · H100 SXM5 assumed 5 4–14 [Project Rainier](/facilities/fac-epoch-gc-project-rainier) · estimated US 1.2×10²⁵ hardware · H100 SXM5 assumed 6 4–16 [Microsoft Fairwater Atlanta](/facilities/fac-epoch-microsoft-fairwater-atlanta) · estimated US 1.1×10²⁵ power 6 3–29 [Meta Prometheus](/facilities/fac-epoch-meta-prometheus) · estimated US 1.1×10²⁵ power 6 3–30 [IREN Sweetwater 2](/facilities/fac-epoch-gc-iren-sweetwater-2) · estimated US 1.1×10²⁵ power 6 3–31 [Micron Fab 2](/facilities/gridstatus-nyiso-1627) US 10²⁵ power 6 3–32 [TeraWulf Lake Mariner Campus](/facilities/fac-epoch-fluidstack-lake-mariner) · estimated US 9.0×10²⁴ power 7 4–37 Lower bound“Days to” assumes 24/7 single-model training at full facility capacity, which no real frontier run achieves once shared inference, failures, restarts and evaluation runs are accounted for. The range beside it is the estimator's parameter-sensitivity band — interconnect efficiency and model-FLOP utilisation across their documented ranges — and is not a confidence interval: there is no distributional model behind it. Derivations: [Methodology](/methodology). ### What a compute gate is doing on this axis A cumulative-FLOP threshold is a proxy. It is administrable — a provider knows its own training budget, and the figure is auditable after the fact — which is a real virtue and the reason the instrument exists in the form it does. What it is not is a measurement of the thing being regulated. The horizontal axis above is the proxy; the vertical axis is closer to the concern. Two facts on this page bear on how tightly the proxy binds. The fit between them is real but loose (R² 0.66 over 18 observations at 14 distinct compute values), and it is unobserved exactly where the policy questions now are. A reader should take neither as an argument against compute thresholds — the alternative instruments have their own problems, and [the capability side is floor-bound in a way that is arguably worse](/capabilities/elicitation-gap). It is an argument for knowing which of the two quantities a given claim actually rests on.