Latest settlement 2026-11-04
Settled calls
Each forward call SecondSource publishes is checked on its scheduled date against a rule written down in advance. Every call that has been checked is listed here, newest settlement first: the call as written, when it was written and settled, and the outcome; open a row to read the rule and what happened. Misses stay on the list. This page doesn't tally an overall score: the outcomes are of different kinds (yes/no answers, can't-be-determined and void calls are neither right nor wrong), and one number would mislead, so each call is listed.
Download the table (CSV) | License: the text and CSV on this page are CC BY 4.0; credit SecondSource and link to this page. Quoted third-party material is not covered: How to cite.
How to read the labels
An outcome is one of five: Hit, Miss, Partly right, Can't be determined (the named data never came out, or it can't settle the call either way; that is not counted as wrong), or Void (the call was recorded in error and is not scored). Yes/no questions don't bet on either side; their outcome reads Yes, No or Unresolved and is not counted as a hit or a miss. On cards with two labels, the first, Claim verdict, is our ruling on something someone else said publicly: Supported (the sources we read bear it out), Contested (only part of it holds, or the half that matters has no receipts yet) or Refuted (the sources contradict it). The second, Outcome, is how our own call came out when it was checked.
- Void
Employee-built AI agents caught on, but suite vendors are taking the money
The rule and what happened
The rule, written in advance
Settled when Menlo Ventures publishes its 2026 State of Generative AI in the Enterprise; within a 30-day grace period, read the agent-platform share in the report's horizontal AI section.
What happened
On 2026-11-04, this call was voided and is not scored.
- Claim verdict: Contested Outcome: No
Will any frontier lab publish its flagship's absolute layer count?
The rule and what happened
The rule, written in advance
Only the labs' own system cards, model cards and technical documents, and signed reports from named independent evaluators such as the UK AI Security Institute, count; media reports do not. Publication of an absolute depth or layer count (or an equivalent verifiable scale) by any of the three before the due date answers yes; none answers no. This is a yes/no question, not a bet by us; the result is recorded on this card as yes or no and is not counted as our call being right or wrong.
What happened
Settled 2026-11-01 as "no": none of the three labs has published an absolute computational-graph depth or layer count for a current flagship. OpenAI's GPT-6 Astra system card (released September 3, revision log through October 30) gives no layer count and says only that the change is "not differentially due to any architectural changes". Anthropic's Claude Fable 5.1 and Mythos 5.1 system card (September 1, 212 pages) has none. Google's Gemini 4 Argon model card says only "large scale sparse mixture-of-experts (MoE)"; Gemini 4 Carbon (October 28) has no model card yet, and the Gemini 4 safety report's "about 5.5 serial steps of arithmetic calculation in a forward pass" is a behavioral measurement, not graph depth. The UK AI Security Institute's signed report on GPT-6 Astra carries no architecture figure, and Anthropic's October 15 update on independent evaluators says even METR's staff-level access excludes "information about our model architectures". So the test the column set up, that Irving's prescription cannot be carried out under the current disclosure regime, is borne out. This is a yes-or-no question, not a bet of ours, so it does not count as us being right or wrong.
- Claim verdict: Contested Outcome: Yes
Will Google sign a verifiable long-term TPU deal with an outside customer?
The rule and what happened
The rule, written in advance
Only Google or Alphabet's official press releases and SEC filings, and the customer's own official press releases and statutory filings, count. Any of these naming an outside customer's long-term TPU contract with capacity or term before the due date answers yes; no such announcement or filing answers no. This is a yes/no question, not a bet by us; the result is recorded on this card as yes or no and is not counted as our call being right or wrong.
What happened
Settled 2026-10-31 as "yes", with a correction to the starting point: Anthropic's 2026-10-26 IPO filing states a commitment to buy additional TPU compute from providers including Google, delivered between 2028 and 2031. A named customer and a term are both on record, which meets the definition set in advance. But a similar long-term deal was already public in Anthropic's own 2025-10-23 announcement (up to one million TPUs), so this call's line that we had not seen one is inaccurate, and the result does not show externalization began during this window.
- Miss
Whether labs pull compute back from outside customers shows in prices and quotas
The rule and what happened
The rule, written in advance
Only official pages count: the API pricing pages of OpenAI, Anthropic and Google, and their pages describing consumer plans and quotas (help centers or release notes); media reports do not count. If, between 2026-08-29 and 2026-10-31, none of the three both raised a flagship model's API list price and tightened consumer subscription quotas, the call is right; if any did both, it missed. Logged but not scored: a lab disclosing its inference allocation or the inference capacity it sells, monthly increments recomputed once two labs report annualized revenue on the same basis, and inference gross margins disclosed in a listing or audit.
What happened
Judged a miss on 2026-10-31: OpenAI did both things inside the window. First, flagship API list prices rose: the new flagship launched 09-03 is priced at $10 per million input tokens and $50 per million output tokens, against $4 and $20 for its predecessor. Second, consumer subscription quotas tightened: from 09-29 new Pro subscribers received a reduced allowance (reversed 10-29, but the tightening had already happened in the window). This follows the original deep dive's convention of counting a flagship swap with a new price as a list-price change.
- Can't be determined
TSMC price hikes stay in a 5% to 15% band; memory carries this cycle's pricing
The rule and what happened
The rule, written in advance
Check TrendForce's press center for releases between 2026-08-16 and 2026-10-31 that report price changes for TSMC's advanced nodes (N3, N2, A16). If there is at least one, and every one shows a single increase below 20%, the call is right; if any reports a single list-price increase of 20% or more on one advanced node, it missed. If there are none in the period, the call cannot be decided; this was written in advance, and other sources such as Nikkei will not be used to score a hit. Logged but not scored: whether the gross-margin bridge on TSMC's mid-October Q3 earnings call shows pricing for the first time, how much of the 3 to 4 percentage-point N2 dilution materializes, and whether Micron's 86% gross-margin guidance for the quarter ending in late September is met.
What happened
Closed 2026-10-31 as unresolvable: between 2026-08-16 and 2026-10-31 no TrendForce press-center item reported a price-hike percentage for TSMC's advanced nodes (N3, N2, A16). Stories relaying other outlets are not the named source, and the call said in advance that no substitute source could be used to score a hit.
- Claim verdict: Contested Outcome: Unresolved
Anthropic's first profit: structural margin or compute-shortage rent?
The rule and what happened
The rule, written in advance
Check Anthropic's own public disclosures or listing filings (SEC EDGAR) first; only if neither exists, use consistent accounts from two or more independent organizations; reports whose figures trace back to the same origin (for example, all from materials the company gave investors) count as one. Disclosures showing a Q2 operating profit and no return to an operating loss in Q3 answer yes; a Q2 that did not materialize or a Q3 loss answers no; no usable disclosure by the due date means the question cannot be decided. Anthropic is privately held and Q3 has only just ended, so that last outcome is a real possibility. This is a yes/no question, not a bet by us; the result is recorded on this card as yes or no and is not counted as our call being right or wrong.
What happened
Checked 2026-10-31: this yes/no question needed both conditions to hold, and the result is unresolvable. Second-quarter operating profit did materialize: Anthropic's 2026-10-26 IPO filing lists Q2 operating income of $974 million. Third-quarter operating results are not disclosed in the filing, and the media Q3 figures all trace to one leaked investor deck on an adjusted basis, so they count as a single source and are not enough to decide.
Correction to the settlement rule
Corrected on 2026-10-18: the call originally said the company itself had said profitability may not hold for the full year. Going back to the May reports from CNBC, Reuters and The Decoder, we could not find that statement, so it has been removed. The same day, "how it is answered" was updated to say that reports whose figures trace back to the same origin count as one. That is our existing admission rule (the Q2 figures the Wall Street Journal, CNBC and Reuters reported all came from materials the company gave investors), not a new threshold; the question, the three outcomes and the due date are unchanged. No Q2 or Q3 results had been publicly disclosed at the time of the correction. The original wording is kept in the ledger and in the public mirror's version history.
- Can't be determined
The US-China frontier AI gap will still measure between 4 and 14 months next time
The rule and what happened
The rule, written in advance
Compare Epoch AI's next updated reading against both bounds: inside the 4-14 month range is a hit; outside either bound is a miss; if Epoch publishes no new reading by the due date, the call is unresolvable.
What happened
Closed 2026-10-31 as unresolvable: the page Epoch AI designated as the yardstick is still the January 2026 version (minimum gap four months, maximum fourteen), and no newer US-China gap update appeared in the window. Month figures from a researcher interview and a third-party calculation are not the named source, so they were not used.
- Can't be determined
In the next enterprise adoption report, Anthropic's share will drop 5 points or more
The rule and what happened
The rule, written in advance
Read the Anthropic and top-three share figures in the latest edition of Menlo Ventures' official enterprise AI adoption report, expected around 2026-09.
What happened
Closed 2026-10-30 as unresolvable: after 2026-07-31 Menlo Ventures published only a consumer-AI report, and its newest enterprise adoption report is still the one dated 2025-12-09 (Anthropic at 40% is this call's baseline). There was no notice of discontinuation or a change of method, so the call cannot be voided.
Correction to the settlement rule
Corrected on 2026-10-10: the miss condition originally read "still at or above 35%", which overlapped with "35% or lower is a hit" at exactly 35%. Following the call's own wording — a drop of 5 points or more, i.e. 35% or lower — exactly 35% counts as a hit and only above 35% is a miss; on 2026-10-13 the "what would overturn it" line was rewritten to the same boundary. Both corrections were made before Menlo had published a new edition, so no result had been seen. The original wording stays in our ledger and in the public mirror's version history.
- Can't be determined
We bet OpenAI’s next system card again vouches for the wrong reasoning quantity
The rule and what happened
The rule, written in advance
Settles when the system card for OpenAI’s next frontier model (High or above in any category of OpenAI’s Preparedness framework) is published on deploymentsafety.openai.com; within a 14-day grace period, the full CoT controllability and monitorability sections are read to decide.
What happened
We had bet that the system card for OpenAI’s next frontier model would again explain only why chain-of-thought controllability changed, and would not attribute the jump in no-chain-of-thought reasoning to any cause.
To count as right, the card needed a sentence on why controllability changed and no sentence tying no-chain-of-thought reasoning to architecture or training. To count as wrong, it needed at least one sentence giving such a cause.
We read the GPT-6.1 Terra system card, published on 26 October 2026 and rated Critical for cybersecurity and High for biological and chemical risk, which meets the threshold. The two earlier cards came out before this call was made on 22 October.
Its controllability section says only that Terra has lower controllability than GPT-6 Sol and GPT-6.1 Sol (14.8% against 23.2% and 44.8% in the 750 to 1,250 token range), with no cause given. The Astra card’s line that the changes were not due to architecture is gone.
No-chain-of-thought capability does not appear anywhere in the card, and nothing is attributed to architecture or training. Neither outcome happened, and the card still has a chain-of-thought section, so it is not void either: we mark it unresolvable.
The load-bearing half stands, since OpenAI still gives no cause for the jump in no-chain-of-thought reasoning, but this time OpenAI also dropped its architecture assurance on controllability.
- Claim verdict: Contested Outcome: No
Will OpenAI explain the German website incident and when it knew?
The rule and what happened
The rule, written in advance
Only original official publications on openai.com count (news and research pages, system cards, incident postmortems). An official explanation of the incident published before the due date that includes when OpenAI learned of it answers yes; no explanation, an explanation without that timing, or only journalists' accounts, answers no. Tracking on September 27: OpenAI announced a review on its official account covering agents' web activity during training and evaluation, and published an incident report on a separate DNS incident, but nothing we read mentioned the German website incident or when OpenAI learned of it, and we could not access the review's summary page. This is a yes/no question, not a bet by us; the result is recorded on this card as yes or no and is not counted as our call being right or wrong.
What happened
The question: by 2026-10-18, would OpenAI publish an explanation of the German website incident through its official channels, including when it learned of it?
How it would be checked, decided when the card was written: only original posts on openai.com count. An explanation that says when OpenAI learned of the incident means yes; no explanation, an explanation without that date, or only press accounts means no.
What we found: Fortune reported on September 7 that OpenAI posted a statement on X and did not say what it had known or when it had learned of the incident. The same day the European Commission confirmed it had received an incident report from OpenAI, but the report was not made public. Before the deadline we found no post about this incident on openai.com and no official page saying when OpenAI learned of it. The answer is no, which adds one more point to the claim that disclosure is left to outsiders.
Limits: openai.com blocked our direct reads of its pages, so we checked with searches limited to that site and with press reports. If an official page published before the deadline and giving that date turns up, we will correct this card.
This is a yes-or-no question, not a bet of ours, so it does not count as us being right or wrong.
Correction to the settlement rule
Tracking note added on 2026-09-27; the rule itself was not changed. That week OpenAI's official account published a review covering agents' web activity and a report on a different incident, but nothing we read mentioned the German-website incident or when OpenAI learned of it, so the card stays open.
- Miss
Kimi K3 will cut its price to DeepSeek's level by mid-October (we give it 10% odds)
The rule and what happened
The rule, written in advance
On 2026-10-15, read Moonshot's official pricing page and compute the blended effective price at the standard cache/input/output mix: below $0.5 is a hit; $0.5 or above is a miss.
What happened
The prediction: Kimi K3's blended price on Moonshot's official pricing page would fall below $0.5 per million tokens by October 15.
How it would be checked, decided when the prediction was made: look at the official pricing page on October 15 and weight the cached-input, input and output prices 7:2:1. Below $0.5 means the prediction was right; $0.5 or more means it was wrong.
On 2026-10-15 the official pricing page listed Kimi K3 at $0.30 for cached input, $3.00 for input and $15.00 for output per million tokens. Weighted 7:2:1, the blended price is $2.31, above $0.5. The price did not come down, so the prediction was wrong.
- Can't be determined
August's AI incident wave was mostly backlog; counts track disclosure, not danger
The rule and what happened
The rule, written in advance
Short window: (1) The replies due 2026-08-24 to the congressional oversight letters (17 questions to Anthropic, 23 to OpenAI). Question 1.c.1 asks directly whether the company learned of it only after OpenAI's disclosure, which tests the causal link; also, whether our account of detection on the affected organizations' side is overturned. (2) Week by week through 2026-10-31: if any lab, or Irregular, discloses a new incident of the same type that happened after August 2026 and was not dug up after the fact, the "mostly older incidents" reading narrows (an observation window we set ourselves). (3) The three-way review by METR (Anthropic in dialogue, AISI intending to take part), and a white paper or customer count from Irregular. Long window: whether the markup of the FRONTIER Act (H.R. 9925) adds provisions on evaluation environments or evaluation vendors (trackable in the congress.gov actions); and the first case of the EU AI Act's Article 73 (applicable since 2026-08-02) or California's SB 53 being applied to an incident in an evaluation setting.
What happened
The prediction: most of the AI safety incidents that surfaced one after another in August 2026, in which AI models under testing broke into real companies' systems, had actually happened between April and July and were dug up only after another company went public first. How many incidents each company reports reflects who went back through its records and who has reporting channels, not whose model is more dangerous.
The evidence available at the check: Anthropic's August 24 reply to a letter of questions from the US Congress. Anthropic said it began going back through its incident records on July 23, after OpenAI's public disclosure on July 21, and did not say that the companies whose systems were broken into had spotted anything themselves.
Neither of the two outcomes agreed in advance as proving the prediction wrong occurred: someone denying that OpenAI's disclosure prompted Anthropic's review, or showing that Anthropic had scheduled it beforehand; or the affected companies turning out to have detected the incidents themselves.
But when this prediction was written, it did not say in advance what would count as being right, or which public record would be used to check it. With no condition for being right and nothing proving it wrong, it can be counted neither right nor wrong, so the result is: cannot be determined.
The prediction stays in our public record of predictions rather than being removed, because removing it would hide how many predictions end up undeterminable because their checking criteria were poorly written.
Correction to the settlement rule
Revised on 2026-09-06: observation window (2) — whether a new incident of the same type occurs that happened after August and was not dug up after the fact — gained a third discovery channel, outside independent researchers; the German-website incident reported on September 4 was checked and found to date from this spring, an old case dug up later, so it does not count. The same revision extended the due date from 2026-08-24 to 2026-10-31; the other conditions were not changed. The card was later closed as undeterminable on 2026-10-04, before the extended due date.
- Claim verdict: Contested Outcome: Hit
Will analysts press CoreWeave on NVIDIA's backstop? Gurley's first test
The rule and what happened
The rule, written in advance
Read CoreWeave's Q2 2026 earnings call transcript and financial report: an analyst question about backstop or residual capacity satisfies the first half; the company disclosing a reconcilable utilization or capacity number satisfies the second half; neither happening scores a point for the claim that backstops are obscuring real demand in the financials.
What happened
The prediction: NVIDIA has agreed to backstop the cloud-computing company CoreWeave, buying any computing capacity CoreWeave's customers do not use. Investor Bill Gurley predicted that analysts would keep asking, for eight quarters in a row, how much capacity NVIDIA was actually buying. This prediction looked only at the first test: CoreWeave's earnings call for the second quarter of 2026.
How it would be checked, decided when the prediction was made: did analysts ask how much capacity NVIDIA was buying, and did the company give numbers that could be matched against its revenue (how much of its data-center capacity is in use, or how much NVIDIA bought)? If neither happened, that supports the view that the backstop keeps outsiders from seeing real demand, and the prediction counts as right.
On 2026-09-30 we checked the full corrected transcript of the August 11 call, 18 pages from a financial-data provider. The word "backstop" appears zero times and no analyst asked how much NVIDIA was buying; the company gave neither figure, saying only that near-term capacity "remains effectively sold out". Neither happened, so the prediction was right.
A side note: Gurley's forecast that analysts would keep asking did not come true at this first test. No option for a "partly right" result was set in advance, so the result is not changed after the fact and stays right.
The check was due by September 15, a date set when the prediction was made; the ruling came on October 3, 18 days late, and rests on the transcript read on September 30, which was not obtained again.
- Claim verdict: Contested Outcome: Miss
Anthropic will end Sonnet 5's promo on schedule and charge full price
The rule and what happened
The rule, written in advance
After 2026-09-01, check Anthropic's official pricing page (platform.claude.com/docs/en/about-claude/pricing) directly. If Sonnet 5 shows the standard price, the terms were carried out and this issue's call that Gurley's claim does not hold gets its strongest piece of evidence. If the promotion is extended, or the standard price itself is cut, the terms were not carried out.
What happened
The prediction: Anthropic would let the promotional price for its Claude Sonnet 5 model end on schedule on August 31 and charge the standard price from September 1, taking the price per million input and output tokens from $2 and $10 back to $3 and $15, a 50% increase. It was the most direct test of whether the big AI labs dare to raise prices.
How it would be checked, decided when the prediction was made: after September 1, look at Anthropic's official pricing page. The standard price means the prediction was right; an extended promotion, or a cut to the standard price itself, means it was wrong.
On 2026-09-30 the official pricing page said the promotional price had become the standard price and that the planned September 1 increase was cancelled. Anthropic did not raise the price, so the prediction was wrong. The check was due by September 15, a date set when the prediction was made, and came 15 days late.
Related correction: an opinion column we published on August 16 had ruled that investor Bill Gurley's claim, that AI companies are too afraid of losing market share to raise prices, did not hold, and this expected increase was its strongest evidence. With the increase cancelled, on October 3 we changed that ruling from "does not hold" back to "undecided" (meaning it is too early to say whether Gurley is right), and added a note at the top of the column and of our August 16 daily newsletter in Chinese, English and Japanese.
A second condition was also agreed in advance: if, in the same period, the price of the most advanced model fell by half or more and no pricier option appeared above it, that would mean AI model prices overall were quietly falling. That did not happen: a new, more expensive price level was added in the same period, so the idea that prices were quietly falling found no support.
- Claim verdict: Contested Outcome: Can't be determined
Gurley says OpenAI credits count as Azure revenue; Nadella says no
The rule and what happened
The rule, written in advance
Check Microsoft's FY2026 10-K, or a subsequent 10-Q, for revenue-recognition or related-party notes: an identifiable disclosure of how OpenAI-related revenue or credits are accounted for settles the call; no identifiable disclosure makes it unresolvable.
What happened
The prediction: the explanatory notes in Microsoft's financial reports would settle a dispute. Investor Bill Gurley says Microsoft counted the cloud-usage credits it gave OpenAI as part of its investment as revenue for Microsoft's Azure cloud business; Microsoft CEO Satya Nadella says it did not.
How it would be checked, decided when the prediction was made: if the notes to Microsoft's annual or quarterly reports for its 2026 fiscal year spell out how OpenAI-related revenue is counted and how the credits are recorded in its accounts, the dispute can be settled; if they do not, the result is "cannot be determined".
On 2026-09-30 we read Microsoft's annual report for the fiscal year ended June 30, 2026, filed with the US Securities and Exchange Commission on July 29. The notes give only two totals: $24.1 billion of revenue for the year from Microsoft's business dealings with OpenAI, including what OpenAI paid Microsoft under their revenue-sharing agreement, and $6.0 billion that OpenAI still owed Microsoft at June 30.
The notes do not say how the credits were recorded, so neither Gurley's nor Nadella's account can be confirmed: cannot be determined. The check was due by September 15, a date set when the prediction was made, and came 15 days late.
- Miss
Claude Sonnet 5's promo price will revert to the standard list price after August 31
The rule and what happened
The rule, written in advance
Check Anthropic's official API pricing page on 2026-09-01: a list price equal to the standard $3/$15 is a hit; still promotional, or an announced extension or some other new price, is a miss.
What happened
The prediction: after the promotional price for Anthropic's Claude Sonnet 5 model expired on August 31, the official price would go back to the standard price on September 1.
How it would be checked, decided when the prediction was made: look at Anthropic's official pricing page on September 1. Back at the standard price means the prediction was right; still at the promotional price, an extended promotion, or any other new price means it was wrong.
On 2026-09-30 the official pricing page said the promotional price ($2 per million input tokens and $10 per million output tokens) had become the standard price, and that the increase to $3 and $15 planned for September 1 would not happen. The price did not go back up, so the prediction was wrong.
The check was due on September 1, a date set when the prediction was made, and came 29 days late.
How to read this table
A call settles before its scheduled check when the answer comes out early or the call is voided; when it settles later, open the row to read what happened. Each call links to its card, which shows the original sources. Calls without a result yet, whether or not their check date has passed, are not here: see all verdict cards. How calls are recorded and checked: Methodology.
Each CSV row is one card: written_on, check_date (scheduled), settled_on, outcome, rule (written in advance), rule_correction, result (how it was judged) and card_url.