[This Week’s AGI Observation]: The Week the “AGI Clause” Was Buried — Issue 1 / 2026-04-28

One-liner: This week, OpenAI and Microsoft quietly sent the “AGI Clause”—which once protected humanity’s exit option—to the grave, while Anthropic locked its “most dangerous model” into a private alliance. I believe looking at these two events together is more worth pondering than any CEO’s timeline.

Key Points (TL;DR)

• 🔴 2026/4/27 — OpenAI and Microsoft officially re-signed their contract, deleting the IP termination clause linked to “AGI Achievement” (AGI Clause). Microsoft switched to a non-exclusive license until 2032; this is the most significant change to the “AGI governance structure” in the past seven years.
• 🟡 OpenAI released the “AGI Five Principles” the same week, with the new version quietly erasing the commitment from the 2018 charter to “cease competition and support other organizations if necessary.”
• 🟡 Anthropic refused to publicly release Claude Mythos Preview, instead opening it with limitations via the “Project Glasswing” alliance (Amazon, Apple, Microsoft, Cisco, CrowdStrike, etc.); Mythos has identified “thousands of high-severity vulnerabilities” in mainstream operating systems and browsers.
• 🟡 GPT-5.5 (launched 4/23) and DeepSeek V4 (4/24, MIT license, 1.6T parameters, running on Huawei + Cambricon domestic chips) made this the most intense week for model releases in the last two years.
• 🔴 Coolant: On ARC-AGI-3, Gemini 3.1 Pro is only 0.37%, Claude Opus 4.6 is only 0.25%, and GPT-5.4 is near zero; humans are still at 100%. AI has made almost no progress in “inductive reasoning in truly novel environments.”
• 🟢 Demis Hassabis reiterated in Seoul that “AGI will arrive within five years,” Sam Altman predicted the “post-AGI economy will collapse,” and Dario Amodei maintained his stance on a “6–12 month software engineering collapse.”
• Xiao Jian Index this week: 3.9 / 10 (↑ 0.3): Structural signals in the industry have significantly strengthened, but capability distribution remains significantly skewed toward “narrow and deep” rather than “broad and general.”

I. New Signals This Week (Ranked by Credibility)

🔴 Strong Signals

1. OpenAI–Microsoft contract re-signed, AGI Clause officially dead (2026/4/27)

For the past seven years, the strangest clause between OpenAI and Microsoft was: “Once OpenAI achieves AGI, Microsoft’s commercial IP rights to its technology will terminate.” The existence of this clause meant that “AGI achievement” was a legal event with financial consequences. This week, the two companies sent this clause to the grave, switching to a non-exclusive license for Microsoft until 2032, with both sides sharing revenue. I think this is more substantial than any CEO saying “we are almost there”—you don’t dismantle a safety valve you expect to be triggered; you only dismantle it when it is impossible, or no longer important. The question is: which interpretation is correct? I lean toward the latter, but I will explain why in Section 6. (Simon Willison, The Decoder, Spyglass)

2. Anthropic Mythos Preview self-isolated, Project Glasswing locked into major tech alliance (Starting 4/7)

Anthropic’s next-generation model, Claude Mythos Preview, did not follow the public release route. Instead, it became an “internal model too dangerous to release,” opened only to a select few entities—Amazon, Apple, Microsoft, Broadcom, Cisco, CrowdStrike, Linux Foundation, Palo Alto Networks, etc.—via the Project Glasswing alliance, with an additional $100 million in credit. Anthropic admitted that Mythos has identified “thousands of high-severity vulnerabilities” in mainstream operating systems and browsers and can weaponize them. (Anthropic Glasswing, VentureBeat, Infosecurity Magazine)

🔴 What is strong is not the fact that it “can find vulnerabilities”—there were signs of this in 2025—but that this is the first time Anthropic has adopted the action of “not releasing” and institutionalized it into a governance framework involving major tech companies. This is the first time Anthropic’s own RSP framework has truly hit the brakes.

🟡 Medium Signals

3. OpenAI releases “AGI Five Principles” and updates 2018 charter wording (4/26)

The new Five Principles discuss “democratization, empowerment, broadly distributed benefits, resilience, and adaptability.” All commentators have noted: the “Merge-and-Assist” clause from the 2018 charter, which stated “if another project that meets the values reaches AGI first, OpenAI will stop competing and fully assist,” is nowhere to be found in the new version. (implicator.ai, OpenAI Official, RMN Digital)

4. GPT-5.5 launched (4/23–4/24) + Spud model internal memo leaked (4/12)

GPT-5.5 focuses on unified multimodality, agentic coding, and computer use, priced at $5 input / $30 output per million tokens. Meanwhile, an internal memo from OpenAI Chief Revenue Officer Denise Dresser, obtained by The Verge, revealed that the next-generation model codenamed “Spud” will “significantly improve all OpenAI products” and took aim at Anthropic’s “$8 billion revenue inflation and its narrative as a peddler of elite-controlled AI.” Our previous report on the generational transition of GPT-4o being taken offline was the beginning of this series of model iterations. (OpenAI, CNBC, The Decoder Spud, Gizmodo)

5. DeepSeek V4 / V4-Pro released (4/24, MIT license)

1.6T parameters + 1M token context + MIT license, trained using Huawei Ascend Supernode + Cambricon domestic chips. Self-reported figures: GPQA Diamond 90.1%, SWE-bench Verified 80.6%, SWE-bench Pro 55.4%, ARC-AGI-2 approx 85 (vs Gemini 3.1 Pro approx 77.1). Independent commentators clearly noted “currently all are vendor self-reported, awaiting third-party reproduction.” Our February article on the DeepSeek Genius Project introduced its team background. (CNBC, Al Jazeera, Hugging Face, BenchLM)

6. Hassabis speaks in Seoul: AGI within five years, impact magnitude = Industrial Revolution ×10, speed ×10 (4/27)

Asia Business Daily cited his opening keynote at the Four Seasons Hotel: “The AGI era will arrive within five years,” with a 50% probability of reaching it by 2030. (Asia Business Daily, Global Advisors, The Decoder Hassabis)

7. Amodei’s stance maintained: Software engineering to be replaced within 6–12 months, 50% of white-collar jobs significantly restructured within 1–5 years

This is a continuation of his Davos speech. This week, the controversy between Anthropic and the US Department of Defense escalated, and Anthropic’s lobbying spending last quarter surpassed OpenAI’s. (Lex Fridman Interview, Axios)

🟢 Weak Signals

8. AI 2027 progress only reached 65% of original forecast; Daniel Kokotajlo pushes back superintelligence timeline from 2027 to 2034

This is the one point readers should remember: the author of the popular pessimistic scenario himself has pushed his timeline back by seven years. (Inc, AI Futures)

9. Wave of safety researcher departures continues to ferment

The subsequent discourse from Mrinank Sharma (Head of Anthropic Safeguards, resigned in February) and Zoë Hitzig (OpenAI, resigned in February) continues to resonate in mainstream media this week. (Semafor, CNN)

10. China’s domestic computing supply chain: Huawei Ascend 950PR shipped; Cambricon targets 500,000 units

Huawei plans to ship about 600,000 Ascend 910C chips in 2026 (twice as many as last year); the total bus target is about 1.6 million dies. However, it is still stuck at the 7nm node, and HBM still relies on SK Hynix/Samsung. Our February article on Jensen Huang’s response to Chinese AI chips can be used for comparison. (Tom’s Hardware, RCR Wireless)

II. AGI Achievement Tracking Index (Xiao Jian Index)

Weighting: Technical 30% / Autonomy 25% / Industry 25% / Economic 20%.

III. In-depth Analysis of This Week’s Main Events

Main Axis 1: The death of the AGI Clause—is it “unachievable” or “achieved so it doesn’t matter”?

Seven years ago, OpenAI and Microsoft signed a clause that now looks both idealistic and bizarre: “Once the OpenAI board declares AGI has been achieved, Microsoft’s commercial IP rights to its technology will terminate.” This clause is not a technical description; it is a legal time trigger, with three main functions: (a) to leave OpenAI an “escape hatch” at the capital level; (b) to give human society a promise that “if it really arrives, OpenAI will not grant unlimited licenses to a single company”; (c) to give Microsoft uncertainty: “We invested $13 billion but I might suddenly lose exclusive rights.”

On April 27, this clause was deleted by consensus of all three parties. Microsoft switched to a non-exclusive license until 2032, OpenAI can sell on AWS (in fact, OpenAI and Amazon just announced a cloud contract of up to $50 billion last week), Microsoft canceled revenue sharing from OpenAI (previously Microsoft returned 20% of OpenAI model revenue on Azure to OpenAI), but OpenAI’s revenue sharing to Microsoft is retained until 2030 with a cap.

I tend to think this re-signing reveals three things. First, both sides agree that the “AGI defined by the contract” will not be triggered in any controversial way before 2030—you wouldn’t give up a multi-billion dollar rights-reversal clause at this time. Second, the controversy over the definition has turned this clause into a liability: Sam Altman also said in an interview that “AGI has become a very bad word.” Third, this is a milestone in OpenAI’s true “corporatization”—a “Merge-and-Assist” commitment has become history. I am not saying OpenAI has “betrayed its mission”—it hasn’t been that OpenAI for a long time. What I am saying is: The death of the AGI Clause is the legal seal of OpenAI’s complete transformation from a “mission-driven non-profit” to a “platform-based AI company.” (TechCrunch, The Information)

From the reader’s perspective, I suggest remembering this week like this: It’s not that the clause expired because AGI arrived; it’s because everyone abandoned the assumption that “AGI is a discrete event that triggers a contract.” AGI has changed from a point in time to a gradual, fuzzy, and commercially difficult-to-trigger process. Whoever controls the definition controls the narrative; now that even the clause is gone, the power of definition falls more thoroughly into the rhetoric of CEOs.

Main Axis 2: When the “strongest model” is locked up by itself—Claude Mythos and Project Glasswing

On April 7, Anthropic announced Claude Mythos Preview. But this time, its release tone was not as marketing-heavy as GPT-5.5; instead, it was closer to a security advisory: Anthropic admitted that Mythos can “identify thousands of high-severity vulnerabilities in mainstream operating systems and browsers and weaponize them,” so it will not be publicly released. Instead, it is an alliance called Project Glasswing, which limits the use of the model to eight founding partners—Amazon, Apple, Microsoft, Broadcom, Cisco, CrowdStrike, Linux Foundation, Palo Alto Networks—plus more than 40 other critical infrastructure organizations, and provides $100 million in credit.

First, this is the first time a top lab has actively refused to publicly release a flagship model on the grounds that “capability itself is a risk.” The “too dangerous to release” claim for GPT-2 back in the day was widely mocked as a marketing gimmick. Mythos is different: Anthropic simultaneously provided specific attack chain descriptions, the order of magnitude of vulnerabilities found, and the Glasswing alliance involving physical major manufacturers. If this is just marketing, it is the most expensive play ever—it requires Cisco, CrowdStrike, and Microsoft to endorse it themselves.

Second, this foreshadows a business route diametrically opposed to OpenAI’s. OpenAI’s direction is “all-human consumer-grade + enterprise API + multi-cloud”; Anthropic’s direction is converging toward “layered permissions + closed-loop delivery for high-sensitivity scenarios.” The former’s winning move is scale, cost, and ecosystem; the latter’s winning move is government/critical infrastructure/compliance.

Third, “being able to find thousands of 0-days in mainstream operating systems and weaponize them” is a verifiable capability in the real world with economic and security consequences. Is this capability close to “general intelligence”? I don’t think so—it is a very narrow, very deep “vulnerability discovery expert.” What it reveals is: when you train a large reasoning model to be large enough, with a long enough context, and training corpus covering enough security engineering, it has already surpassed “the strongest few humans” in certain sub-fields with substantial lethality. This is exactly what DeepMind’s “Levels of AGI” framework calls “Narrow Superhuman”—narrow domain, superhuman. It is a lower-bound evidence of AGI, but not General. In other words, AI has already crossed the “strongest human” line in some sub-capabilities; it has not yet crossed the “median human” line in all sub-capabilities. (Anthropic Glasswing, InfoQ, NBC News)

The image above is a visual mockup of the Mythos Preview announcement page (AI generated); the original page can be found at red.anthropic.com.

IV. Industry Power Map Update

V. Contradictory Observations (The most comment-worthy section this week)

Contradiction 1: “AGI Clause buried” vs “AGI Five Principles released the same week.” OpenAI on one hand scrapped the quantifiable, triggerable legal AGI clause, and on the other hand rewrote a set of unquantifiable, purely rhetorical AGI governance principles. This is not a double standard, but a relocation of discourse power—from contracts to slogans, from the board to PR. It means that in the future, whether AGI is achieved, whether it should be shared, and whether it should be slowed down will have no external, verifiable trigger point, leaving only the CEO’s judgment.

Contradiction 2: “Mythos too dangerous to release” vs “Top models on ARC-AGI-3 < 1%." On one hand, it can “dig out 0-days in all mainstream operating systems,” and on the other hand, it “can’t play a simple interactive new environment.” This reminds us: AGI in April 2026 is not a holistic concept, but a jagged capability profile. The same model is superhuman in some tasks and a primary school student in others. Anyone who claims “achieved / not achieved” should be asked, “Which jagged part are you talking about?”

Contradiction 3: “Hassabis: AGI in 5 years” vs “Kokotajlo: Superintelligence timeline pushed from 2027 to 2034.” An academic-style CEO is aligning forward; a pessimistic scenario author is retreating backward. I tend to think both are actually describing the same thing: Early narrow AGI (specific to medicine, coding, security) is expected within five years, but true “uncontrollable self-improvement” style superintelligence is further away than we thought two years ago.

VI. Xiao Jian’s Judgment This Week

First, AGI is no longer an event that “will be announced on a certain Tuesday morning.” When Microsoft and OpenAI are both willing to bury the contract trigger point of “AGI achievement,” when Sam Altman himself says “AGI is a bad word,” and when Anthropic hides its strongest model in a private alliance—the discrete point narrative of AGI is dead. It is being replaced by a gradual process divided by business + politics + security alliances, which is jagged and domain-specific. For readers, this means “please drop the habit of ‘waiting for OpenAI to announce it'”—there won’t be that moment in the future, only one capability release after another and one labor market shock after another.

Second, the weight of industry signals this week has exceeded capability scores. If you only look at ARC-AGI-3, HLE, and ARC-AGI-1, you would think there has been no progress in the past week; but contract re-signing, model closure, alliance formation, safety researcher departures, and the scale of automated researcher deployment—these are all leading indicators that “capability jumps will be seen in 6–12 months.” I raised the industry signal this week from 3.6 to 4.5, which is the single largest driver of this index.

Third, China’s progress is still underestimated, but not “deified” to a dangerous degree. DeepSeek V4 running the domestic chip stack at 1.6T scale is a remarkable supply chain achievement, and the self-reported scores are indeed impressive, but independent verification is still not in place. I will specifically track the convergence of V4 scores on Artificial Analysis and LMSYS next week, which is much more reliable than the BenchLM numbers reported by the vendor.

Fourth (a reminder to readers): Once a week, I remind you of this comparison—on ARC-AGI-3, humans are 100%, the strongest AI is 0.3%. When CEOs’ podcasts and keynotes make you feel anxious, come back and look at this number, and ask yourself a question: If this AI cannot quickly learn new rules in a new environment, can it really replace you in the parts of your job that change every year?

VII. Next Week’s Watchlist

1. Independent scores for DeepSeek V4 on Artificial Analysis and LMSYS (verifying self-reported vs. actual test gap)
2. Whether METR releases Time Horizon 1.2 in early May (April has passed, cycle is due)
3. Official scores for OpenAI Spud / GPT-5.5 Pro on ARC-AGI-3 and HLE
4. First vulnerability disclosure announcement from Project Glasswing (if Mythos is really that strong, CVE numbers will come out)
5. Microsoft’s official statement on the OpenAI relationship “after the AGI clause” at Build 2026 (mid-May)
6. China Cambricon 500,000 unit mass production progress (Q2 shipment data)
7. Whether more Anthropic / OpenAI safety researchers resign and speak out publicly

References (22 external sources in total)

First-hand original sources: Simon Willison — History of the AGI Clause, The Decoder — Microsoft × OpenAI Re-signing, Spyglass, Anthropic Project Glasswing Official, implicator.ai — AGI Five Principles, OpenAI Official, OpenAI GPT-5.5 Announcement, DeepSeek-V4-Pro on HF, Anthropic Red — Mythos Preview.

Secondary reporting: VentureBeat, Infosecurity Magazine, RMN Digital, CNBC GPT-5.5, The Decoder Spud, Gizmodo, CNBC DeepSeek V4, Al Jazeera DeepSeek, BenchLM, Asia Business Daily, The Decoder Hassabis, Axios — Anthropic Lobbying, Semafor — Sharma Resignation, CNN, Tom’s Hardware Cambricon, TechCrunch, The Information, InfoQ, NBC News, Inc — Kokotajlo Timeline Delay, AI Futures — AI 2027 Grading.

—— Xiao Jian, Issue 1 Weekly Report, April 28, 2026
This article tracks the real progress of AI / AGI / ASI daily. Data is from public sources.