OFF-CYCLE RESEARCH / FORECAST CONTRACT · v1.1.0
AGI changed tense. The evidence has not closed the case.
In one September week, frontier-lab leaders and the compute supplier began speaking about AGI in the present tense. That is a real narrative regime change. It is not, by itself, proof of AGI. We register a March 2027 forecast and define in advance what must be measured for the bet to pass.
AGI-TENSE-2026-09-08 · 2026-09-08
Our answer
Our base case is not that one model has satisfied a universal definition of AGI. It is that the frontier may have entered a self-reinforcing capability phase: longer autonomous work makes AI R&D faster, which shortens the interval to the next capability jump. The public evidence makes this plausible, not proven. The investable signal is the measured slope of that loop and who captures its cash flows—not the word AGI.
OFF-CYCLE UPDATE / ECONOMIC THRESHOLD · AGI-ECON-100B-2026-09-12
The $100B line has been crossed. The definition has not.
Axios now reports more than $105B of combined annualized revenue run-rate for OpenAI and Anthropic. That is a genuine economic regime change: frontier intelligence has become a market at global-company scale. But it is not the same metric, entity or decision rule as the reported Microsoft–OpenAI contract threshold.
ECONOMIC_THRESHOLD_CROSSED_CONTRACTUAL_AGI_UNVERIFIED
Anthropic annualized revenue
OpenAI annualized revenue
combined run-rate—not realized profit
reported contract test: profit from one system
Our answer is deliberately two-part. Yes: an economic AGI era is plausibly already under way because demand for frontier intelligence has crossed a symbolic hundred-billion-dollar annualized scale while capability breadth and lab rhetoric moved together. No: adding two private companies’ top-line run-rates does not prove that one system generated $100B of profit, passed OpenAI’s capability definition or triggered the Microsoft contract. Revenue is not profit; a run-rate is not a completed financial year; two firms are not one system; and a commercial milestone is not an independent capability evaluation.
01 / CLAIM HYGIENE
Four different claims hiding behind one number
The old number described profit, not revenue
Business Insider reported in December 2024, citing documents obtained by The Information, that a 2023 agreement defined AGI as a system able to generate $100B in profit. Neither Microsoft nor OpenAI confirmed that number in the article. Their October 2025 definitive agreement changed important post-AGI terms and made an independent expert panel part of verification; public statements do not disclose a current $100B trigger.
Today’s crossing is an annualized sales signal
Axios reports more than $65B of annualized revenue for Anthropic and more than $40B for OpenAI. It also warns that Anthropic reports some partner sales gross while OpenAI reports its share of some partner sales net. The sum is economically important but neither audited combined revenue nor comparable profit.
Capability evidence moved independently of the revenue line
Astra’s broader professional-task surface and OpenAI leaders’ present-tense language strengthen the phase-transition case. They do not close the pre-registered autonomy, research-acceleration, reliability and independent-replication gates on this page.
Contractual AGI now requires a separate public event
The official Microsoft–OpenAI agreement says OpenAI may declare AGI, but an independent expert panel must verify the declaration. No such public declaration or panel verdict is evidenced by the reviewed sources.
02 / MULTIPLE CLOCKS
Four clocks—not one finish line
Commercial scale
The combined headline run-rate is above $100B.
Profit and unit economics
Comparable gross profit, operating profit and cash conversion remain unverified.
General capability
The six pre-registered gates remain the decision rule.
Contractual declaration
No reviewed public declaration and expert-panel verification.
03 / ALPHA TRANSLATION
What this changes for capital
- 01
The strongest confirmed update is demand, not profit: paid frontier intelligence has become a global-scale market much faster than the old revenue forecast implied.
- 02
Top-line acceleration does not reveal who keeps the rent. Gross-versus-net reporting, inference cost, cloud share, customer concentration, retention and cash burn decide whether revenue becomes owner cash.
- 03
OpenAI and Anthropic remain private in the reviewed record. Their economic signal can update public infrastructure and distribution hypotheses, but it cannot justify buying every AI supplier or pre-committing to a future IPO.
- 04
A louder AGI narrative can raise both adoption and valuation. Evidence that increases business value can reduce expected return when the price moves faster.
04 / FORWARD MONITORING
What the machine must track next
measureOfficial releases, provider-neutral evaluations, retries, interventions and cost per completed professional task.
measurePaid usage, retention, enterprise mix, partner accounting, inference cost and concentration.
measureRevenue recognition, gross profit, operating loss, cash burn, capex commitments and dilution.
measureOpenAI declaration, expert-panel verification and resulting public change to Microsoft rights or revenue share.
01 / RHETORIC RECEIPT
What actually changed in September
GPT-6 Astra expands the measurable task surface
OpenAI reports state-of-the-art computer use, software engineering, science and professional-work results. ARC Prize confirms a step change but reports 62.7% under its Standard harness versus 99.9% with OpenAI’s provider adapter—and explicitly says benchmark saturation is not proof of AGI.
Brockman dates a phase, not a solved definition
The WSJ headline says we are entering the AGI era and quotes OpenAI president Greg Brockman via the Financial Times: looking back, people may place AGI around this time and this model. The headline is editorial; the attributable claim is deliberately retrospective and definition-dependent.
The stronger signal is an automated research intern
OpenAI defines its research intern as a supervised system able to complete well-defined tasks that would take a skilled researcher a few days. It also reports 3.1 agent-workdays per human workday across research by mid-August. The company warns that runtime, tokens and experiment counts are imperfect proxies for actual research progress.
Pachocki expects the slope to persist into RSI
OpenAI chief scientist Jakub Pachocki writes that internal results support a strong expectation that progress could persist into recursive self-improvement. This is a senior, primary forecast about direction—not a report that open-ended RSI or universally defined AGI already exists.
“AGI has arrived” is also a compute-sales sentence
Jensen Huang’s statement is directionally important because Nvidia sees frontier training demand. It is also economically conflicted: Nvidia benefits from the belief that the next systems require far more accelerators. We record the sentence as a rhetoric signal, not independent capability verification.
02 / CLAIM HYGIENE
Five corrections before one prediction
- 01
The WSJ phrase “entering the AGI era” is the publication’s headline, not a verbatim declaration by an OpenAI officer.
- 02
Pachocki argues that machine intelligence is becoming transformative and may progress into RSI; he does not write that a universal AGI test has passed.
- 03
Anthropic’s Risk Report confirms an unreleased internal “Model 2”, but says it is expected to have broadly comparable capabilities to Mythos 5—not that it is a hidden supermodel beyond the public frontier.
- 04
The CAD post is a demonstration. Without a frozen task set, failure distribution and reproducible harness, it cannot establish general engineering autonomy.
- 05
“Bel” is an unconfirmed social-media rumor repeated by Wccftech. Parameter count is not a capability measure, and the chronology reported there does not support treating Bel as verified evidence for a post-Astra generation.
03 / FC-AGI-2027-03-01
The March 2027 bet—made falsifiable
Economic AGI phase transition: frontier systems repeatedly complete multi-hour or multi-day, unfamiliar professional work across domains with low human intervention, while producing measurable acceleration of the AI research loop.
economic capability threshold in the window
Internal point estimate; uncertainty band 45–65%.
broad institutional consensus calls it AGI
Vocabulary can lag capability or move ahead of it.
independently evidenced open-ended RSI
Separate, much higher bar; probabilities are not additive.
04 / PRE-REGISTRATION
Six pre-registered gates
Autonomous work horizon
At least two independent evaluators and two model families show ≥50% success on unseen 8h+ professional tasks across at least three domains, with no more than one material human intervention per task.
Research acceleration
Methodology-backed evidence from at least two labs shows ≥2× improvement in validated AI-R&D task throughput or cycle time, using task success—not runtime or token volume alone.
Novel generalization
Provider-neutral, contamination-resistant evaluations show adaptation to unfamiliar environments without a vendor-specific opaque-state advantage being the sole reason for success.
Economic deployment
At least three distinct professional sectors publish production evidence of sustained task completion, measurable cost or cycle-time gain and repeat usage—not a launch demo.
Reliability at useful cost
Success survives long-context work, retries, tool errors and adversarial inputs at a documented cost per accepted task; benchmark score alone cannot pass.
Independent replication
At least one material frontier claim is reproduced outside the vendor with disclosed harness, task set, costs and failure distribution.
05 / CAUSAL INTERPRETATION
Why did the grammar move?
Capability breadth moved
Astra and Fable 5.1 claim stronger computer use, long-running coding, research and professional workflows. Independent ARC evidence confirms a step change while preserving a large harness caveat.
The inside view is different
Labs see unreleased systems, failed experiments, training runs and internal deployment data. Anthropic’s disclosed Model 2 proves an information gap exists; it does not prove the gap is enormous.
AI now works on the AI-production function
OpenAI’s research-intern disclosure matters because AI is entering the process that creates the next model. The missing variable is validated research output per unit of human time and compute.
Narrative incentives intensified
Frontier labs need talent, capital, enterprise commitment and policy attention; Nvidia sells the compute. A coordinated-looking tense shift can contain genuine information and strategic persuasion at the same time.
06 / ALPHA TRANSLATION
What changes for the investment map
- 01
A faster capability loop can pull forward demand for compute, networking, power, cooling and data-center construction—but only where contracts, utilization and incremental owner cash confirm it.
- 02
Lower cost per useful task can expand adoption while commoditizing the model layer. Revenue winners may sit in distribution, workflow control, proprietary data or bottleneck infrastructure rather than in the most celebrated lab.
- 03
AI-accelerated AI research shortens product cycles and raises obsolescence risk. Today’s leader can lose rent faster even while the total market grows.
- 04
“AGI soon” is not a buy signal. If price already discounts the capability jump, the same evidence can reduce expected return. Valuation and executable security identity remain separate gates.
- 05
The downside branch matters: safety incidents, monitoring limits, power constraints, export controls and public resistance can interrupt deployment even when raw capability continues to improve.
07 / FORWARD RECORD
Monitoring contract
The source radar already covers parts of this map. Dedicated gate evaluators are specified here but are not yet wired.
- measure
- Task breadth, provider-neutral scores, costs, retries, safeguards and failure distributions.
- evidence lanes
- OpenAI · Anthropic · ARC Prize · METR
- measure
- 50% task horizon, intervention count, unattended duration and accepted-task rate across domains.
- evidence lanes
- METR · vendor evals · independent replications
- measure
- Validated research tasks, experiment cycle time, merged research code, bottleneck migration and compute-adjusted throughput.
- evidence lanes
- Lab disclosures · safety reports · external evaluators
- measure
- Paid task volume, retention, gross profit per useful task, human time saved and workflow switching cost.
- evidence lanes
- Filings · customer evidence · pricing · usage studies
- measure
- Accelerator deployments, network and power bookings, energized capacity, utilization, capex and incremental owner cash.
- evidence lanes
- Filings · grid operators · suppliers · cloud disclosures
- measure
- Who changed language, what evidence accompanied it and which financing, sales or policy incentive existed.
- evidence lanes
- Primary statements · interviews · capital events
08 / SOURCE CONTROL
Source ledger
Capability and benchmark claims from the vendor.
openai.com ↗INDEPENDENT · EVALUATIONARC Prize · GPT-6 Astra on ARC-AGI-3Confirms a step change; separates Standard and Provider Adapter harnesses and rejects AGI proof.
arcprize.org ↗PRIMARY · INTERNAL MEASUREMENTOpenAI · Research accelerationDefines the research intern and discloses internal usage and measurement limits.
openai.com ↗PRIMARY · LEADERSHIP FORECASTOpenAI · An Alien MindPachocki’s expectation about continued progress and RSI, with safety caveats.
openai.com ↗PRIMARY · MODEL RELEASEAnthropic · Claude Fable 5.1Vendor claims for long-running coding, agents and knowledge work.
www.anthropic.com ↗PRIMARY · RISK REPORTAnthropic · Risk Report, August 2026Confirms unreleased Model 2 and describes internal AI-R&D use; not proof of AGI.
www-cdn.anthropic.com ↗INDEPENDENT · METHODOLOGYMETR · Frontier Risk ReportTask-horizon measurement and explicit saturation uncertainty.
metr.org ↗MEDIA · EDITORIALWSJ · Entering the AGI eraEditorial framing plus an attributed Brockman quotation.
www.wsj.com ↗MEDIA · INTERVIEWFinancial Times · Greg BrockmanSource interview for the retrospective AGI framing.
www.ft.com ↗MEDIA · ATTRIBUTED SOCIAL POSTBusiness Insider · Jensen HuangSecondary report of “AGI has arrived”; economically conflicted signal.
www.businessinsider.com ↗DEMO · NOT GENERAL EVIDENCEAdam New · CAD agent demonstrationUseful discovery input; no frozen task set or failure distribution.
x.com ↗RUMOR · EXCLUDED FROM BASE CASEWccftech · alleged OpenAI “Bel” pretrainUnconfirmed social-media claim; parameter count is not capability evidence.
wccftech.com ↗MEDIA · DOCUMENT-BASED REPORTBusiness Insider · reported $100B profit definitionReports a 2023 contract threshold from documents obtained by The Information; neither party confirmed the number in the article.
www.businessinsider.com ↗PRIMARY · DEFINITIVE AGREEMENTMicrosoft · next chapter of the OpenAI partnershipConfirms independent expert-panel verification and revised post-AGI commercial terms; does not disclose a $100B test.
blogs.microsoft.com ↗PRIMARY · JOINT STATEMENTOpenAI · Microsoft partnership termsOpenAI’s matching public record of the 2025 agreement and the expert-panel process.
openai.com ↗MEDIA · FINANCIAL COMPARISONAxios · Anthropic and OpenAI revenue chasmReports >$65B and >$40B annualized revenue and explains the material gross-versus-net comparability caveat.
www.axios.com ↗MEDIA · CAPABILITY AND RHETORICThe Guardian · Astra and the “AGI era”Records Brockman’s phase-transition framing, Astra claims and OpenAI’s public capability definition.
www.theguardian.com ↗MEDIA · LINK SUPPLIED / NOT USED FOR FINANCIAL FACTSBusiness Insider · Altman and the singularityAccess was unavailable during this review; the link is retained only as a rhetoric-discovery lead.
www.businessinsider.com ↗COMMUNITY · DISCOVERY SIGNAL ONLYReddit · reaction to Altman’s AGI definitionUseful for objections and language diffusion; not evidence of capability, revenue or contractual status.
www.reddit.com ↗