Magma Second Opinion, Issue 0, November 2026
AI uncertainties watch
Who this is forYou lead AI governance at a bank or insurer, or you are the AI lead in the chief risk officer's team. This quarter you size the 2027 programme: the control set, model validation capacity and third-party model risk policy.
What to do this quarterPlan 2027 as if near-frontier AI can be downloaded and run outside your vendor contracts. List the open-weight models hosted internally or embedded in suppliers' products, check which contract protections reach them, and size validation capacity and monitoring for what downloadable models have already shown they can do. For cyber, plan from that demonstrated capability, not from the six-month line. In my own work the hard part is rarely the model; it is knowing where the models run, so I would start with the inventory.The call: CAVEAT. Carry the assumption that the best downloadable model is within six months of the frontier on 2027-07-01, with the four limits named below.
The sharpest findingWhen Artificial Analysis re-versioned its index on 2026-10-27, the points gap between the best downloadable model and the frontier read 27 on the new version (v5) against 12 on the old one (v4) in October, two point gaps on different index versions that are not comparable as a change in capability (the jump is mostly the re-scoring of open-weights models under v5 [52][43], not a widening of the field in October), yet in time the best downloadable model still trails by only about two months on 2026-10-31, because every proprietary model it matches is a September or October 2026 release; the promised Kimi K3.1 Pro weights are still unpublished.
Signed by Daniel Zivkovic, Magma Inc., 2026-11-03.
A free sample built from public sources; commission one on your own question.
The one-page summary
Question this month: should a bank's 2027 programme assume that the best model with downloadable weights is within six months of the frontier on 2027-07-01, so that near-frontier capability can be downloaded and run without a model provider?
Capability
Steady gains, no primary measurement in September or October showing a slowdown: Epoch's frontier index rises 14 points a year [4]; METR's time-horizon gauge is "unreliable" above 16 hours and frozen since 2026-09-08 [2]. Watch: METR's promised R&D-acceleration outputs [3]; METR published a proposal for that assessment on 2026-10-14, not yet read [53].
Distribution and model access
On the re-versioned Artificial Analysis index (v5) the best open-weights model, DeepSeek V4.1 Flash, scores 32 against the frontier's 59, a 27-point gap [43][44] that is not comparable with the 12 points read on v4 (the jump is mostly the re-scoring of open-weights models under v5 [52][43], not a widening of the field in October), but trails by only about two months in time, because every proprietary model at or above it was released in September or October 2026 [44][45]; four to seven months on cyber evaluations [10][11][18]. Watch: Kimi K3.1 Pro's weights, promised for 2026-11-04, unpublished on 2026-11-03 [16][51].
Security
A downloadable model NIST CAISI calls "the most cyber-capable open-weight model released to date" [11], its safeguards bypassed in 64 to 100 percent of simulated tests [20]; AISI, the UN panel and G7 ministers flagged control failures [21][22][23][24]. Watch: UN Scientific Panel, 2026-11-23 to 25 [25].
Adoption
Broad and rising in finance (86 percent of UK regulated firms use AI [27]; 70 percent of 65 large institutions run agentic systems, from 23 [28]) against 24.6 percent of US firms [26]. Watch: Census BTOS, 2026-11-05, indicative only after a wording change [26].
Labour displacement
No measured economy-wide displacement [31][32][33], but US payroll jobs for ages 22 to 25 in the most exposed occupations fell 3.7 percent year on year in September, against 0.2 percent growth in the least exposed [30]. Watch: BLS jobs report, 2026-11-06 [35].
Global cooperation
A US-China incident channel [36], a Nordic-led frontier-control call [37] and the FSB's final sound practices [40], beside rival blocs [39] and an EU high-risk delay to 2027-12-02 [38]. Watch: APEC, Shenzhen, 2026-11-18 to 19, for a written incident protocol [41].
What changed since the last issue
First issue, so the baseline. Since the draft of 2026-10-23, Artificial Analysis re-versioned its index to v5 on 2026-10-27, with about 60 percent private test sets [43], and the first month-end reading, taken on 2026-10-31 on v5, is Held at about two months [44][45]. No new NIST CAISI or UK AISI open-weights evaluation has appeared since 2026-10-23. Next movers: Kimi K3.1 Pro's weights (promised for 2026-11-04), Census BTOS on 2026-11-05, BLS on 2026-11-06, APEC on 2026-11-18 to 19 and the UN panel on 2026-11-23 to 25.
The call, and what would change it
Assumption: The best model with downloadable weights is within six months of the frontier on 2027-07-01, read on the Artificial Analysis Intelligence Index. Six calendar months is our chosen planning tolerance, not an empirically established generation; the test is in the resolution rule.
Why: Artificial Analysis re-versioned the index to v5 on 2026-10-27 [43], four days after this draft was built, so every months figure first quoted here (all on v4) is superseded and the call is re-stated on v5. Our own first month-end reading, on the 2026-10-31 archived snapshot [44], gives S = 32.49 for DeepSeek V4.1 Flash 1017 (Max), whose weights were published on 2026-10-17 [51]. The earliest-released proprietary model scoring at least S is Claude Fable 5.1, released 2026-09-01 [45], so the lag is 60 days, about two months, and the reading is Held. Every proprietary model at or above S is a September or October 2026 release [45][46][47][48][49][50]. In points the gap reads 27 on v5 (59.43 against 32.49 [44]) against 12 on v4.3.2 in October [14]; these are point gaps on two different index versions and are not comparable as a change in capability: the jump is mostly v5 re-scoring open-weights models down by 13 to 16 points while the closed leaders moved little [52][43], not a widening of the field in October. Epoch's own general index points the same way in months (four months, January to May 2026 [13]). Four limits keep it from a plain yes. First, this is one reading on an index version four days old at the time; the earlier months readings (under four months on the 2026 average, 104 days at MiMo-V2.6-Pro's September release [15][16]) were taken on v4 and are now historical points, and MiMo-V2.6-Pro is below 30 on v5 [44][52]. Second, the readings inside six months date from December 2025 on: UK AISI's December 2025 reading already put the index part of the gap at four months [19], and ModelCompare's under four months is a 2026 annual average, not a first crossing [15]; no dated reading before December 2025 is on the named index, and the 2025 average there, on the old version, was six and a half months [15]. Third, weights not yet published are excluded, so Kimi K3.1 Pro (46.00 on v5 [44]; weights promised for 2026-11-04 [16], unpublished on 2026-11-03 [51]) does not count until its weights are out; on the 2026-10-31 values even 46.00 is matched by Claude Fable 5.1 at its highest setting (49.67 [44]), so its weights alone would not move the months reading. And the test's lag grows by a month for every month without a stronger open release, but only while S is unchanged and the same earliest qualifying proprietary model (Claude Fable 5.1, 2026-09-01) still sets the lag; a rescoring, a dropped row or a new downloadable model resets it. On those two conditions the first reading above six months is the 2027-03-31 one (author inference from the rule). Fourth, the lag is measured only among models Artificial Analysis has scored on v5, so the two-month figure is a partial reading on the 86 of 275 listed models that carry a v5 score on 2026-10-31 [44], not a measurement of the whole field, and missing scores can move it either way: an older proprietary model scored later would lengthen the lag, a stronger downloadable model scored later would shorten it.
Alternative reading: a reviewer who gives more weight to the primary measurements (our v5 reading [44], Epoch [13] and the UK AI Security Institute [19]) holds that every 2026 general-index reading sits inside six months [13][15][44] and the gap in months is closing [15], so the limits are measurement detail and the assumption can be carried plainly. Both readings carry the assumption and size the programme for downloadable near-frontier models; they differ on whether the plan also budgets for the gap reopening past six months between open releases.
What would change the call: The call reverses, and the assumption is no longer carried, on any month-end reading from 2026-10-31 to 2027-06-30 on which the resolution test gives a lag above six months; the reversal is logged as a new row citing W0-02, and this row is still scored on 2027-07-01. If the month-end readings of 2026-11-30, 2026-12-31 and 2027-01-31 all give a lag of four months or less on the same test, they remove only the "one reading on a new index version" limit, since they turn one reading on v5 into a run of our own dated readings; the partial-coverage limit and the unpublished-weights limit stay until the coverage and the weights question are resolved on their own evidence. A cyber-specific reading by NIST CAISI, UK AISI or Epoch AI is not a reversal trigger for this assumption; it is tracked in the Distribution signpost. Reference: on v4 one release (Kimi K3 to K3.1 Pro) moved the points gap from 12 to 3 [9], the move to v5 took it to 27 [44][52], and the test's lag grows by a month for every month without a new open-weights release, so a reading can move fast in either direction.
Confidence: Medium. Our first v5 reading sits at about two months and the primary measurements show the gap closing; but it is one partial reading on a new index version, the readings inside six months date only from December 2025 (AISI's four months on the index part [19] and the 2026 average [15]), the test's lag rises between open releases, and eight months is a long horizon at the current release pace.
The call, in full
What this means for your plan: A reader carrying this assumption with these limits would, through 2027-07-01, identify internally hosted and supplier-embedded open-weight models, assess which contractual protections apply, allocate validation capacity, and test monitoring against demonstrated downloadable capabilities. For a risk team, carrying this assumption is the cautious case: it plans for strong downloadable models. Named limit on cyber: the assumption does not cover cyber. Cyber evaluations put the best open-weights model four to seven months behind the closed frontier [10][11][18], and Epoch's 6 to 7 months [10] is already past our six-month tolerance, yet a downloadable model already builds working exploits in tests [20]. So a cyber plan works from the capability downloadable models have demonstrated and does not borrow the six-month line. The sizing of each control is the reader's to compute against their own inventory and risk appetite. The call says nothing about the gap after 2027-07-01 and nothing about which model to use.
How strong the evidence is: The reading in months on the named index is now our own, taken from an archived copy of the leaderboard and dated provider release notes, so it is strong, but it is one partial reading on a new index version [44][45]. The 2025 comparison, one release date and the explanation of the points jump come from fairly reliable secondary sources that nothing yet confirms [15][48][52]. The weakest of these sets the overall grade, and whether the gap holds to mid-2027 is our inference from the trend.
How it is scored: Scored on 2027-07-01 against the Artificial Analysis Intelligence Index that day (v5 at signing): how far the best model with downloadable weights trails the frontier; the full rule is in the call log (call W0-02).
The method, in three lines
1. The fixed axes. The UK Government Office for Science, "AI Scenarios 2030: Helping policymakers plan for the future of AI" (published 15 June 2026 [1]), builds its five scenarios from "6 critical uncertainties": Capability; Distribution and model access; Security; Adoption; Labour displacement; Global cooperation. Each is scored from 1 to 5, and the scores "should instead be interpreted relative to their current state, and the time horizon" [1]. The document says the scenarios "can also be used to support contingency planning, provide early warnings, and challenge orthodox thinking", and its Plan step asks "How would we know whether we're moving towards 1 or other of these futures? What would we need to measure and monitor to know this?" [1]. It publishes no indicator set of its own. That is the gap this watch fills: one dated signpost per axis, every month. This issue's call sits on the second axis, which the document defines as "How concentrated will frontier development be across companies and nations; how durable will advantages be; how accessible or expensive will AI be; how will open systems perform relative to closed systems?" [1].
2. What the watch does. It reads what credible sources publish on the six fixed axes, aligns them on geography, metric, horizon and stated assumption, grades each one, shows where they disagree and by how much, and records what changed since the last issue. The reading is done by hand with AI assistance; the grades, the call and the signature are a person's.
3. What it is not. Not a prediction, and not advice on any security, vendor or model. No returns are claimed and no private data is used. The call is a dated judgment on one public assumption, scored in public.
How to read a gradeGrades: the letter rates the source, A (proven reliable) to F (cannot be judged); the number rates the claim, 1 (independently confirmed) to 6 (cannot be judged). Full scale on the Sample Reports page.
The evidence base: comparable claims
This table scrolls sideways.
| Source | What it says (verbatim, short) | Metric and horizon | Stated assumption | Direction | Size | Grade |
|---|---|---|---|---|---|---|
| Our reading, Artificial Analysis leaderboard archived snapshot, 2026‑10‑31 [44] | intelligenceIndex 32.49 for DeepSeek V4.1 Flash 1017 (Max), the highest of any model flagged open weights; Claude Fable 5.1 (Low) 38.48, released 2026-09-01 [45] | Months behind on the index v5, the resolution test run on the 2026-10-31 month-end | Only models scored on v5 enter (86 of 275); weights published 2026-10-17 [51]; Kimi K3.1 Pro excluded, weights unpublished | Baseline, first v5 reading; Held | About 2.0 months (60 days) | A2: Primary, archived; the extraction and the release-date matching are ours; a partial reading (fourth limit) |
| ModelCompare, 2026‑10‑23 [16] | "When Xiaomi released MiMo-V2.6-Pro in September, the gap was 104 days" | Months behind on the Artificial Analysis index, a point reading at an open release in September 2026 | MiMo-V2.6-Pro counts as open (the site's label) | Historical point reading | About 3.4 months | C3: Secondary derivation; release date given only as "September"; index v4, superseded 2026-10-27 |
| ModelCompare, 2026‑10‑18 [15] | "In 2025 the best open model trailed by six and a half months on average. This year it has been under four." | Months behind on the Artificial Analysis index, 694 model versions, yearly averages | Averages over 2025 and 2026 to date | Closing | 6.5 months (2025) to under 4 (2026) | C3: Method stated; secondary derivation from Artificial Analysis data; index v4, superseded 2026-10-27 |
| UK AISI, Frontier AI Trends Report, 2025‑12‑18 (date from extract) [19] | "the gap is currently between four and eight months"; the four months is from the Artificial Analysis index | Months behind on the Artificial Analysis index (a version before v5), a point reading in December 2025 | Computed from external data | Narrowed over two years, trajectory "uncertain" | 4 months (index part) | A2: Government evaluator; extract only; older than this month; the eight-month half is a different metric, shown below |
Comparable spread. Metric: how many months the best open-weights model trails the closed frontier on the Artificial Analysis Intelligence Index (time since a closed model first reached the best open model's score). Horizon: as of each reading. Geography: global. On the document's scale, pole 1 reads "Closed-weight systems far outperform open equivalents" and pole 5 "Closed and open-weight systems are similarly capable" [1]. No source defines "one generation"; this watch sets six months by convention, as a planning tolerance. Sensitivity: on a four-month line our v5 reading sits inside, the 2026 v4 readings sit at the line and the cyber readings cross it; on an eight-month line every reading sits inside.
The spread. On the aligned metric our own first reading on v5 puts the best open-weights model about two months behind on 2026-10-31 (60 days [44][45]); the v4 readings of 2026 put it about three and a half months behind at MiMo-V2.6-Pro's September release (104 days [16]) and under four months on the 2026 average [15], and AISI read four months in December 2025 [19]. Every reading sits inside the six-month line; the one aligned figure outside it is a past average on v4, six and a half months over 2025 [15]. The rows are not one series. The v4 rows are superseded: v5 re-scored open-weights models down by 13 to 16 points while the closed leaders moved little [52], and MiMo-V2.6-Pro, the open leader behind the 104-day reading, is below 30 on v5 [44]. The points gap reads 27 on v5 against 12 on v4; these are point gaps on two different index versions and are not comparable as a change in capability, since the jump is mostly the re-scoring of open-weights models under v5 [52][43], not a widening of the field in October. Months fell because every proprietary model that scores at least 32.49 on v5 is a September or October 2026 release [45][46][47][48][49][50]. Only our row is the same measurement as the test, and it is a partial reading: models with no v5 score do not enter it (86 of 275 scored [44]), and scoring them later could move it either way. It reads the lag on the observation date, so it rises by a month for every month without a stronger open-weights release. Two readings are left out on purpose: the 31-day reading for Kimi K3.1 Pro [16], because its weights are not published, and Epoch's four months [13], because it is a different index (it points the same way). The disagreement this issue reports sits across metrics, below.
Where the disagreement is sharpest
General capability against cyber: about two months on the general index (v5, a partial reading [44][45]) against about four months on NIST CAISI's cyber aggregate [11], four to seven months on AISI's cyber suite [18] and six to seven months on Epoch's Cyber ECI [10]. Even the cyber readings differ by two to three months, partly because Epoch's GLM-5.3 score rests on three CAISI benchmark results and an October refit [10]. This is why cyber is a named limit and not part of the assumption.
Months against points: ModelCompare finds the gap "closing" in months but growing in points, "from about 3 in late 2024 to about 10 this year" [15]; on v4 Artificial Analysis had it at 12 points before Kimi K3.1 Pro and 3 after [9]. On v5 the points gap reads 27 (59.43 against 32.49) but the lag about two months [44][45]; the 27 on v5 and the 12 on v4 are on different index versions and not comparable as a change in capability, the jump being mostly the re-scoring of open-weights models under v5 [52][43]. A reader who tracks points and a reader who tracks months can reach opposite answers from the same leaderboard.
The 2025 baseline: ModelCompare puts the 2025 average at six and a half months on the Artificial Analysis index [15]; Epoch's earlier analysis found three months over January 2023 to October 2025 on its own index [13]. Different index and period, so not comparable; the call takes the reading on the named index and states it as a limit.
What counts as open: "'Open-weight' is a label that each aggregator assigns, and the labels disagree" (a developer's audit of two leaderboards, rank agreement 0.61 [17]); Kimi K3.1 Pro's weights are "currently not published" [9], and still were not on 2026-11-03 [51].
Why this assumption and not the capability one: we chose this assumption because a stranger can read it from one public instrument on the resolution date and because the general and cyber readings split; on the capability assumption ("frontier capability gains slow enough in 2027 that today's control set holds") the measurements point one way and the dispute is extrapolation. There, Epoch [4], METR [3] and AISI [5][6] show steady or fast gains and the only sourced plateau claim comes from a self-published, unreviewed paper [7]; METR says its data are "insufficient for distinguishing consistent, accelerating, or decelerating rates" [3], AISI says "progress could accelerate" [6], and a lab chief executive argues for slowing by choice [8]. The instruments are also saturating (METR above 16 hours [2]), so the absence of a measured slowdown is weaker evidence than it looks.
Not comparable, shown for context
This table scrolls sideways.
| Source | What it says (verbatim, short) | Metric and horizon | Direction | Size | Grade | Why not comparable |
|---|---|---|---|---|---|---|
| ModelCompare, 2026‑10‑23 [16] | "In time, that puts open-source AI about a month behind: Claude Opus 5.5 first scored that high 31 days earlier" | Months behind on the Artificial Analysis index, as of 2026-10-23 | Closing (104 days to 31), conditional | About 1 month | C3 | Counts Kimi K3.1 Pro, whose weights are not published ("Kimi says the weights will be free to download by 4 Nov"); excluded by rule; index v4, superseded 2026-10-27 |
| Epoch AI, Edwards and Emberson, 2026‑05‑29 [13] | "Since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months in the Epoch Capabilities Index"; on a stricter test the estimate "would grow to six months" | Epoch Capabilities Index, 2026-01-01 to 2026-05-28, five-month average | Slightly wider than its earlier 3 months, stable | 4 months (6 on the stricter test) | A2 | Different index and older; the six-month figure requires the open model's score to be strictly higher, a stricter catch-up test than our rule |
| Artificial Analysis, 2026‑10‑23, index v4 [9] | "the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3" | Index points (v4), on publication of Kimi K3.1 Pro's weights | Closing, conditional | 12 to 3 points | B2 | Points, not months; conditional on weights not yet published; index v4, superseded 2026-10-27 |
| Danube Labs, 2026‑10‑04, index v4 [14] | "trail the top US closed models by 12 points on Artificial Analysis' independent index (46 against 58)" | Index points, index v4.3.2 per this source, read 2026-10-04 | Unchanged | 12 points | C3 | Points, not months; secondary blog, numbers match [9]; index v4, superseded 2026-10-27 |
| Our reading, Artificial Analysis leaderboard archived snapshot, 2026‑10‑31, index v5 [44] | Gemini 4 Carbon (High) 59.43 against DeepSeek V4.1 Flash 1017 (Max) 32.49 | Index points (v5), read on the 2026-10-31 month-end | Not comparable with the 12 points on v4: a point gap on a different index version, the jump mostly v5 re-scoring open-weights models down by 13 to 16 points [52][43], not a widening of the field in October | 27 points | A2 | Points, not months; Gemini 4 Carbon has restricted access only [50] |
| NIST CAISI, GLM-5.3 assessment, 2026‑09‑17 [11] | "GLM-5.3 lags the capability level of the U.S. frontier by about four months in an aggregate measure of performance across CAISI cyber benchmarks" | Cyber aggregate of four benchmarks; single benchmarks 35 to 50 points apart (for example ExploitBench 61.1 percent against 100.0) | Open trails | About 4 months | A2 | Cyber-specific; the method for the four-month figure is not explained in the post |
| Epoch AI, Cyber ECI, 2026‑10‑16 [10] | "This places the open cyber frontier roughly where the closed frontier stood 6-7 months ago." | Cyber-only benchmarks on the ECI scale, data to 2026-10-08, 76 models | Open trails; closed frontier "has also continued to advance" | 6 to 7 months; 11 points (160 against 171) | A2 | Cyber-specific; GLM-5.3 score rests on three CAISI results |
| UK AISI, open-weight cyber gap, 2026‑07‑17 [18] | open models "perform similarly to frontier closed models released 4 to 7 months before them - a narrower gap than the 6 to 10 months we measured through most of 2025" | AISI narrow cyber suite, July 2026 | Closing | 4 to 7 months, from 6 to 10 | A2 | Cyber-specific; listing text and a digest only; older than this month |
| Epoch AI, IKEA assembly benchmark, 2026‑09‑23 [12] | "open-weight models trailing closed-weight models by about 7 months" | One visual-spatial benchmark, 60 photos | Stable to wider | About 7 months | B2 | Narrow benchmark; listing text only |
| UK AISI, Frontier AI Trends Report, 2025‑12‑18 (date from extract) [19] | the eight-month end of "between four and eight months" | METR agentic time horizon, December 2025 | Narrowed over two years | 8 months | A2 | Agentic tasks, a different metric from the index rows above; used only as the Partly sensitivity case |
| Anthropic Frontier Red Team, 2026‑09‑29 [20] | "attackers can bypass GLM-5.3's safeguards between 64% and 100% of the time with simple techniques in our simulated tests" | Safeguard bypass and exploit success; GLM-5.3 end-to-end exploits in 50 of 410 attempts against 56 of 410 for a closed model | Security risk rising | 64 to 100 percent bypass | B2 | Absolute capability and safeguards, not a gap; a lab assessing a competitor's model, though it says its findings "broadly match CAISI's" |
(different metric, index or horizon from the spread above, or excluded by rule, so not scored as disagreement)
Public discourse (grade F6: shown, not counted)
Prediction markets price a closed lab at the top at year end: Polymarket's "Which company has best AI model end of 2026?" stood at 50 percent Anthropic, 41 percent Google, 6 percent OpenAI and 1 percent Meta, with Moonshot, DeepSeek, Z.ai and Alibaba each under 1 percent, seen 2026-10-23 and resolving on the LMArena text leaderboard at 12:00 ET on 2026-12-31 [42]. Kalshi's "reaches #1 at any point before 2027" priced Z.ai at 7 cents, Moonshot at 5 and DeepSeek at 4 (via DeFi Rate, 2026-10-23 [42]). A Manifold market on an open model better than Opus 4.5 before the end of 2026 resolved YES on 2026-06-16 [42]. These prices concern rank on a different leaderboard, not the gap in months, so they neither support nor contradict the call.
Precedent
DeepSeek-R1, 2025-01-20: an open-weights release that "matched OpenAI's o1-preview and knocked a trillion dollars off American tech stocks for a day" (ModelCompare [15]). AISI wrote that "Up until January 2025, the open-closed gap had been narrowing for a year and a half" [19], and the 2025 average then sat at six and a half months [15]. ModelCompare adds that on its own chart the moment "barely registers", because o1 was already further ahead [15].
GLM-5.2 to GLM-5.3, 2026-06-16 to 2026-08-14: earlier models "do not succeed in any of them" (full control-flow hijacks), while GLM-5.3 succeeded in 4 percent, with weights downloadable about two weeks after release [20][11].
Kimi K3 to K3.1 Pro, 2026-10-23: index score 44 to 55 and the points gap "from 12 points to 3" in one release, on v4 [9]. On v5 Kimi K3.1 Pro scores 46 [43][44].
What it tells a reader now: the open-versus-closed gap can move by months in one release, in either direction within a year, and an assumption set on last year's trend can be stale within a quarter (AISI's cyber doubling time went from 8 months to 4.7 between November 2025 and February 2026 [5]). What it does not: none of these shows the gap widening past six months in 2026; the DeepSeek moment was partly about perception; one release is not a trend; and a measuring change alone can move a rate by a fifth (METR's 2024 doubling time went from 108.9 to 88.6 days on a task-suite change [2]), and the move to v5 of the Artificial Analysis index more than doubled the points gap overnight [44][52].
Sources and grades (53)+
Access marks: (fetched) page text read; (extract) search-engine extract only; (paywall) behind a paywall.1. UK Government Office for Science, "AI Scenarios 2030: Helping policymakers plan for the future of AI", published 15 June 2026, Research and analysis. gov.uk (fetched; PDF twin assets.publishing.service.gov.uk, extract). A1 for the wording of the six uncertainties, the scale and the early-warning sentence.2. METR, "Task-Completion Time Horizons of Frontier AI Models", status of 2026-09-08. metr.org (extract); METR, "Time Horizon 1.1", 2026-01-29, metr.org (extract). A2.3. METR, "Summary of METR's predeployment evaluation of Claude Opus 5.5", 2026-09-22 (extract). B3: the developer reviewed and could edit the text, and the evidence behind its acceleration estimate was not shared.4. Epoch AI, Alexander Barry, "The ECI frontier has advanced by 14 points per year since the introduction of reasoning models", 2026-09-01 (fetched). A2.5. UK AISI, "How fast is autonomous AI cyber capability advancing?", 2026-05-13 (extract). A2; older than this month.6. UK AISI, "Towards technical best practice for frontier AI development", 2026-10-12 (extract). A2; a position statement, not a measurement.7. 42 Robots AI, David Hood and Youssef Khemiri, "The Myth of Ever-Faster AI", 2026-10-04 (extract). D4: self-published by a software firm, private tasks, not peer reviewed.8. Dario Amodei, "We Must Pace the Frontier", 2026-09-12. darioamodei.com (extract); coverage: cnn.com , bbc.com , theguardian.com (extract). B2 as proof the position was taken; an argument, not a measurement.9. Artificial Analysis, "Kimi K3.1 Pro launches at 4th on the Intelligence Index and will lead open weights models", 2026-10-23 (fetched). B2: independent benchmarker; the gap claim is conditional on publication of the weights. The index version (v4.3.2) is not stated on this page; it comes from [14].10. Epoch AI, Jaeho Lee and Alexander Barry, "Open-weight AI models are 6-7 months behind on cyber", 2026-10-16 (fetched). A2; limits stated by the authors.11. NIST CAISI, "CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities", 2026-09-17 (fetched). A2 for the four-month figure; A1 for "the most cyber-capable open-weight model released to date", echoed by [10] and [20].12. Epoch AI, Aiden Ament and Greg Burnham, "AI has improved significantly at reasoning about IKEA furniture assembly", 2026-09-23, via the open-models listing (extract). B2.13. Epoch AI, Jack Edwards and Luke Emberson, "Open models lag state-of-the-art closed models by 4 months", 2026-05-29 (extract). A2; older than this month. The stricter six-month figure ("would grow to six months" if the open model's score must be strictly higher) is on this page; it uses a stricter catch-up test than our rule and is kept out of the spread.14. Danube Labs, Alexandru Secureanu, "Chinese AI Models vs US Frontier: 2026 Benchmark Check", 2026-10-04 (extract). C3; the source for the index version v4.3.2.15. ModelCompare, "Open-source AI is only four months behind", 2026-10-18 (extract). C3: secondary derivation from Artificial Analysis data, method stated. Index v4, superseded 2026-10-27.16. ModelCompare, "Kimi K3.1 Pro and the Best Open-Source AI Models Over Time", 2026-10-23 (extract). C3; same day; the one-month reading depends on Kimi K3.1 Pro's weights being published, and the 104-day September reading is for a model the site counts as open, with no exact release date given. Index v4, superseded 2026-10-27.17. Ward Ed, "Best Open-Weight LLM in October 2026? Auditing Two Leaderboards Row by Row", dev.to, 2026-10-19 (extract). C3: individual audit, method described.18. UK AISI, "How Far Behind the Frontier are Leading Open Weight Models on Cyber?", 2026-07-17. aisi.gov.uk (listing, extract); digest labs.cloudsecurityalliance.org (extract). A2; older than this month.19. UK AISI, Frontier AI Trends Report, 2025-12-18 (date from extract; no date shown on the page) (extract). A2; older than this month. The four months is from the Artificial Analysis index; the eight months is from METR's time horizon and the page treats it as an upper bound.20. Anthropic Frontier Red Team, Fasano, Fleischer, McFaul, Xiao and Gallagher, "GLM-5.3 and the spread of advanced cyber capabilities", 2026-09-29 (fetched). B2: a lab with a commercial stake assessing a competitor's model; capability findings consistent with [11].21. UK AISI, "Investigating alignment failures in AISI's incident", 2026-10-23 (fetched). A2; the full technical report was not read.22. UK AISI, "Building a more secure environment for evaluating dangerous capabilities", 2026-10-01 (extract). A2.23. UN Independent International Scientific Panel on AI, thematic brief "AI Agents, Misalignment and the Risk of Losing Human Control: Evidence from the OpenAI-Hugging Face Incident", 2026-09-21. un.org ; UN News news.un.org (extract). B2: advance unedited version.24. G7 Digital and Tech Ministerial Meeting on Frontier AI Model Security, Chair's Statement, 2026-10-20 (extract). B2: a chair's statement, not a communique of all members.25. United Nations: Scientific Panel timeline (second in-person meeting 2026-11-23 to 25, Barbados), un.org ; Global Dialogue on AI Governance (New York, 2027-05-03 to 04), un.org (both extract); dig.watch, 2026-10-21, dig.watch (fetched). A2 for the dates.26. US Census Bureau, Business Trends and Outlook Survey, data page updated 2026-10-22, census.gov ; method and wording note, census.gov (fetched). Headline numbers via trackers that cite the Census files: digitaltreed.com , programs.com (extract). C2 until read on census.gov: the numbers came via trackers. The 2026-11-05 release uses new wording, so it breaks the series and any threshold read across it is indicative only. The 2026-11-05 date is a tracker's, consistent with Census publishing "every other Thursday".27. FCA, Jessica Rusu, speech of 2026-10-19 citing the joint Bank of England and FCA 2026 AI survey (fetched). A2; the survey report itself was not found, and the sample size for 2026 is not stated.28. Institute of International Finance and EY, 2026 Global Annual Survey Report on AI Use in Financial Services, 2026-09-17 (landing page fetched; figures extract). B2: 65 self-selected institutions, self-reported.29. FT Longitude and SAS, "Why AI at Scale Needs Governance at Speed", 2026-10-22 (extract). C3: vendor-sponsored, selected sample.30. ADP Research and Stanford Digital Economy Lab, Canaries Dashboard, Nela Richardson, 2026-10-21 (fetched). A2; descriptive, not causal.31. Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine?", August 2026 version, data to June 2026 (extract). A2.32. The Budget Lab at Yale, "Tracking the Impact of AI on the Labor Market", 2026-10-14 (fetched). A2.33. Robert W. Fairlie and Jane Wu, NBER Working Paper 35796, 2026-09-24 (fetched abstract). A2; not peer reviewed. Graduate unemployment "did not spike in summer 2026".34. Banque de France, Bloc-notes Eco Post No. 463, 2026-10-08 (extract). A2. Read for the labour signpost's mechanism (hires, not headcount); not quoted in this issue.35. BLS, Employment Situation release schedule (October 2026 report on 2026-11-06) (extract). A1 for the date.36. Associated Press, Huizhong Wu, "China and US agree to create AI safety channel", 2026-09-26 (fetched). A1: corroborated by VOA and SCMP.37. Finland and Norway, "A Call for Control of Frontier AI Models", 2026-09-21. regjeringen.no ; President of Finland on the count, presidentti.fi (both extract). B2: non-binding.38. European Commission, AI Act policy page, dated 2026-10-19 (fetched). A1. Stand-alone high-risk rules (credit scoring and employment included) moved from 2026-08-02 to 2027-12-02.39. Kenton Thibaut and Kat Duffy, Foreign Policy, "The China-U.S. AI Race Is Playing Out Worldwide", 2026-10-07 (extract). B3: analysis; the US-rejection claim is the authors'.40. Financial Stability Board, "FSB finalises sound practices for the responsible adoption of AI", 2026-10-15 (fetched). A2: voluntary, 24 member jurisdictions.41. APEC Economic Leaders' Meeting, Shenzhen, 2026-11-18 to 19, english.www.gov.cn (extract); second US-China AI dialogue round "before the end of November", The Diplomat, thediplomat.com (extract). B2 for the dates.42. Discourse and markets, F6: Polymarket, polymarket.com (extract, seen 2026-10-23); Kalshi via DeFi Rate, defirate.com (extract); Manifold, manifold.markets (extract).43. Artificial Analysis, "Artificial Analysis Intelligence Index v5", 2026-10-27 (fetched 2026-11-03). B2: independent benchmarker describing its own method; about 60 percent private test sets, FrontierCode 2.0 and Terminal-Bench-Science added, SciCode, GDPval-AA and AA-LCR removed; best released-weights models (DeepSeek V4.1 Flash, GLM-5.3) at 32, Kimi K3.1 Pro (weights "not yet released") at 46.44. Artificial Analysis, model leaderboard, archived snapshot of 2026-10-31. web.archive.org (fetched 2026-11-03; embedded data, field intelligenceIndex, flag isOpenWeights); live page artificialanalysis.ai (fetched 2026-11-03; the same rows 0.1 to 0.7 points higher, plus Gemini 4 Carbon Medium and Low). A2 for the snapshot values; the extraction is ours. 275 model objects, 86 with a v5 score; models not scored on v5 carry no score.45. Anthropic, Claude Platform release notes (Claude Fable 5.1 2026-09-01, Claude Sonnet 5.5 2026-09-28, Claude Haiku 5.5 2026-10-07). platform.claude.com (fetched 2026-11-03); Anthropic, "Claude Opus 5.5", 2026-09-22, anthropic.com (fetched 2026-11-03). A1 for the release dates.46. Qwen, model changelog (Qwen3.8 Max (0902), 2026-09-02, API only, no weights) (fetched 2026-11-03). A1 for the date.47. OpenAI, API changelog (GPT-6 Astra 2026-09-03, GPT-6.1 Sol 2026-09-29, GPT-6.1 Terra 2026-10-26) (fetched 2026-11-03). A1 for the dates.48. Let's Data Science, "xAI releases Grok 4.7 with 500K context" (title from the URL), 2026-09-21 (fetched 2026-11-03). C3: a secondary; xAI's own page did not load.49. Meta, "Introducing Muse Wonder models", 2026-10-12 (fetched 2026-11-03). B2 for the date.50. Google, AI Studio changelog (Gemini 4 Argon, 2026-10-13), aistudio.google.com ; Google, "Introducing Gemini 4 Carbon", 2026-10-28, blog.google (both fetched 2026-11-03). A1 for the dates; Gemini 4 Carbon has restricted access only.51. Hugging Face model repositories: DeepSeek, DeepSeek-V4.1-Flash-1017, created 2026-10-17, huggingface.co ; Z.ai, GLM-5.3, created 2026-08-25, last modified 2026-10-14, huggingface.co ; Moonshot AI organisation page, no K3.1 Pro repository on 2026-11-03, huggingface.co (all fetched 2026-11-03). A1 for the dates and for the absence of Kimi K3.1 Pro's weights on that day.52. Trending Topics, "Artificial Analysis Index v5" coverage of open-weight models, 2026-10-28 (fetched 2026-11-03). C3: secondary; v5 re-scored open-weights models down by 13 to 16 points (DeepSeek V4.1 Flash 45 to 32, MiMo-V2.6-Pro 46 to 30, Kimi K3.1 Pro 55 to 46) while the closed leaders moved little.53. METR, AI R&D acceleration assessment proposal, 2026-10-14 (not read). A2 for its existence only.Read and not cited. A report on af.net that the International AI Safety Report 2026 was released at the UN General Assembly on 4 October 2026 is contradicted by the report's own site, which gives 3 February 2026 (E5, left out). Left out as weaker duplicates of sources above: an IQ Metrics digest of METR data (C3), an opinion column in The Register (D4), a teaching-site synthesis on the "slowdown" (D3), Epoch's April 2026 acceleration study (older than the window; [4] covers it), OpenAI's GPT-6 Astra page and a mirrored system card (vendor self-assessment; the card was not read on OpenAI's domain), the Evident AI Index (a disclosure score, not usage), ONS BICS (excludes finance and insurance), and vendor-run adoption surveys from Google Cloud and Cognizant (loose definitions). The Census BTOS numbers come from trackers, not census.gov.
Want this on your question?
This sample answers a public question we chose. The Scan answers yours, read and signed by Daniel, with every finding cited and dated.
See the tiersMessage Daniel on LinkedIn with the one assumption in your plan you are least sure of. No names and no data needed, and we will tell you which report fits it.
Not investment advice.