OpenAI's Great Reset: Killing Every Path That Doesn't Lead to AGI
The story didn't land in one episode. It landed across two.
On April 7, 2026, Greg Brockman sat with Alex Kantrowitz on Big Technology Podcast and laid out the product architecture — "fat layer," Superapp, 70–80% of the way to AGI. Two weeks later, on April 21, he and Sam Altman sat with Ashlee Vance on Core Memory EP 67 — Altman's first long-form joint sit-down with Brockman since Reid Hoffman's Possible in May 2023 — and used the same stage to open three fights: against Anthropic, against Elon Musk, and against the narrative that AI comes without political cost.
Put the two episodes on one timeline and the picture is unambiguous. OpenAI is resetting. Not pivoting — resetting:
- Sora is dead. The video model is shutting down; compute is being redirected.
- The social product is dead. The Sora-content feed experiment is gone with it.
- "Personal AGI" — an assistant that knows your calendar and taste, and books the concert ticket without asking — is now, for the first time, a product term rather than a vision word.
- Altman opened three fronts at once — calling Anthropic's Mythos release strategy "fear-based marketing," saying he's worried Musk will drop the lawsuit before trial, and admitting — a week and a half after a Molotov cocktail was thrown at his San Francisco home — that AI is starting to incur a real political cost.
The whole 90 minutes (and the supporting Big Technology arc), in one piece. And at the end — the ironies Altman and Brockman glided past.
1. The First Cut: Sora Dies, Not Because It's Bad, Because It Isn't AGI
Two years ago Sora was OpenAI's answer to Gemini, Runway, Pika, and Veo. In April 2026, it's dead — the app is shutting down April 26, the API on September 24.
Brockman's rationale on Big Technology was unusually direct: text is the path to AGI, video is not, and you cannot push on every track simultaneously.
"There's been this debate about how far text models can go — we have definitively answered that question. It is going to go to AGI." — Greg Brockman, on Big Technology Podcast, April 7
Translation: Sora's compute got reallocated. This isn't a product failure or a DAU problem. It's strategic opportunity cost — in a world where GPUs are being fought over globally, burning hundreds of thousands of them on video generation is a tax on the AGI target.
Also killed: the quiet social-network experiment OpenAI had spun up as a distribution surface for Sora content. Brockman confirmed: done.
One caveat worth naming. "All compute to text" is OpenAI's tagline, but the same interview also announces a robotics hardware play (see §7). The honest reading is compute reprioritized to text and agentic/coding workloads, with robotics as the single exception — not an absolute.
The signal is bigger than Sora itself:
- OpenAI is not trying to be Meta. No recommendation-algorithm bet.
- OpenAI is not trying to be Midjourney or Adobe. Video generation is ceded.
- The remaining battle is: who ships Personal AGI first.
2. Personal AGI: The Product OpenAI Now Defines Itself By
On Core Memory, Altman used "Personal AGI" as a product term, not a vision term. His definition is concrete:
An AI that knows your calendar, knows your taste, and books the concert ticket without asking.
This is a form-factor shift:
- Memory: deep opt-in long-term memory. You choose what it remembers, and it remembers nearly everything. Once you load your history, decisions, and trust, the cost of leaving is your entire digital life.
- Autonomy: from "write me an itinerary" to "plan my whole trip" — book flights, hotels, restaurants, draft emails, produce PDFs, with approvals only where needed.
- Personalization: it just gets you.
Altman's tell, in the interview:
"ChatGPT was by far not the most impressive technological thing we had done. But it was the thing that hit."
Read that as strategic confession. OpenAI isn't playing benchmark; it's playing who-bonds-deepest-to-the-user. Personal AGI is the answer — you pour your digital context in, and it becomes the OS you can't leave.
Worth naming immediately: the article describes this as lock-in — correctly. A product that's hard to leave is, by construction, a platform with switching costs. OpenAI's anti-Anthropic positioning later in the interview turns on "broad diffusion," but the core product mechanic is the same one Meta and Google use. Diffusion with gravity is still gravity.
3. Superapp and the "Fat Layer": Brockman's Product Thesis
The delivery vehicle for Personal AGI, per Brockman's Big Technology appearance, is a Superapp — three things fused into one:
- Chat — personal memory + context.
- Codex — a general-purpose problem solver for anyone, not just developers.
- Browser / Computer Use — the AI directly operates your machine, across apps and across sites.
Brockman, verbatim on Big Technology:
"One application where anything you want your computer to do, you can ask it."
The cognitive flip sits inside Codex: it was positioned for engineers; non-engineers flooded in to automate spreadsheets, edit video, run knowledge work. Brockman's takeaway:
"The underlying technology we produced is mostly not about code at all. It's mostly about solving problems."
This reframes OpenAI's product architecture. Brockman's new term for it is the fat layer — model + skills + connectors + computer-use, stacked. The competition isn't parameter count; it's the stack's depth and integration.
The implied corollary, which Brockman states directly: thin-wrapper startups — a prompt plus GPT — are toast.
A founder should read that claim twice before internalizing it. Platforms with "fat layers" have been declared wrapper-killers before — GPTs store, plugins, Apple Intelligence, Salesforce Einstein — and the wrappers that survived did so by owning distribution, workflow, data, or regulatory surface that the platform couldn't replicate. The real question isn't "will fat layers kill thin wrappers"; it's "which wrappers have a moat OpenAI actually can't reach." Brockman has an incentive to say the answer is "none." It isn't.
4. Spud / GPT-5.5 — Brockman Puts a Number On It: 70–80%
Also on Big Technology, Brockman offered the most interrogatable line of this April:
"I'd say I'm basically like 70–80% there" — meaning on the path to AGI.
He attached this estimate to Spud, the internal codename for the new base model shipping publicly on April 23 as GPT-5.5. Not a 5.4 patch — a new pre-train, with, in his framing, "two years of research coming to fruition."
He also named a December 2025 inflection point:
"Something happened — for the first time, people said, 'I can let this thing code on its own for hours, and I don't need to watch it.' It went from doing 20% of your tasks to 80%."
Take the 70–80% claim with proper skepticism. The denominator is undefined — 70–80% of what exactly, measured how? Directional estimates of this sort from frontier labs have a track record of aging awkwardly: the last 20–30% is typically the hardest part (see self-driving, see fusion), and "we're almost there" has preceded multi-year plateaus in every prior AI wave. Brockman may be right. But the number is a vibe, not a forecast, and it should be treated as one.
The December inflection is easier to stand behind: coding benchmarks and agentic-task completion rates did step-change in late 2025, and Spud is the first base model trained on the other side of that step. "Designed to move the economy" is Brockman's positioning, not a fact — but the positioning itself is new. OpenAI is no longer selling a chat product. It's selling a work substitute.
5. Firing at Anthropic: Mythos as "Fear-Based Marketing"
The sharpest stretch of Core Memory is Altman's take on Anthropic's cyber model, Claude Mythos.
Context first, because it's load-bearing. Anthropic's Responsible Scaling Policy defines AI Safety Levels (ASL-1 through ASL-5) as capability thresholds that trigger escalating deployment restrictions. In early April, Anthropic announced Mythos had crossed the ASL-3 threshold — specifically on cyber-offensive capability, after internal red-teaming surfaced autonomous zero-day discovery — and would not ship publicly. Instead Mythos was moved into Project Glasswing, a closed program with 12 launch partners including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — a list that is conspicuously critical-infrastructure. That selection is exactly the positioning Altman proceeded to attack.
Altman's attack — "fear-based marketing":
"It is clearly incredible marketing to say — 'We have built a bomb. We will sell you a bomb shelter for $100 million.'"
"There are people in the world who have, for a long time, wanted to keep AI in the hands of a smaller group of people. Justifications vary. But if what you want is 'we need control of AI, just us, because we're the trustworthy people' — fear-based marketing is probably the most effective way to justify that."
This does three things:
- Ad hominem: Anthropic's safety narrative reduced to a sales tactic.
- Positional contrast: OpenAI stands for "broad, personal, AGI for everyone." Anthropic stands for "few, controlled, AGI for institutions."
- Historical irony: Altman has spent years promoting AI existential risk. Coming from the CEO who spent 2023 warning Congress AI could end humanity, "fear-based marketing" is rich.
To be fair to the side Altman is attacking: ASL-3 gating isn't something Anthropic invented for Mythos. The RSP was published in September 2023, the tiers and triggers have been public for nearly three years, and Glasswing is consistent with that prior policy. A skeptic would say: "fear-based marketing" describes something real (safety language does move enterprise procurement) but applied to Mythos specifically, the more honest read is "Anthropic is doing what they said they'd do, which happens to be good marketing." Both things can be true.
Strategically, though, this is now in the open. Altman has explicitly staked OpenAI's 2026 narrative on diffusion; Anthropic's Mythos + Glasswing is its structural opposite. H2 2026 enterprise procurement cycles will price this split.
6. Musk v. Altman: The Most Counterintuitive Line of the Episode
Musk v. Altman goes to federal court in Oakland on April 27. The core dispute is OpenAI's non-profit-to-for-profit restructuring. Musk is the plaintiff; by default the defendant is supposed to be nervous.
Altman's framing in the interview:
"My biggest worry is that Elon drops the suit before it gets to court."
Read this twice. Because the obvious interpretation — "I'm so confident I want my day in court" — is also exactly the line a defendant's PR team writes to project confidence regardless of the actual legal posture. Defendants don't usually want verdicts. They want dismissals, or settlements. Wanting a verdict is an unusual posture that itself deserves interrogation: OpenAI's IPO path needs this cloud cleared; a judgment does that cleanly; dismissal leaves Musk free to re-file and re-frame.
So the read is underdetermined. Either Altman has seen his discovery and likes what's there, or he's performing a confidence he doesn't privately hold because the alternative makes IPO bankers nervous. The Core Memory framing doesn't let us tell them apart. Worth noting which one it is, in public, before ratifying it as analysis.
Either way, bracketed against §5: in late April 2026, Altman is firing outward on three fronts at once. This isn't a defensive crouch. It's also, plausibly, what a defensive crouch looks like when played for the cameras.
7. The Political Cost: A Crime at His Door, an Underclass in the Forecast, and Actuators
Before the analysis: a Molotov cocktail was thrown at Sam Altman's San Francisco residence on April 10, with a second incident (reported as gunfire) within 48 hours. Federal agents subsequently arrested a Texan carrying an anti-AI manifesto. This is a crime against his family, household, and neighbors, and the fact that it lines up with a thesis about political backlash doesn't change that. It's worth acknowledging as such before drawing analytical conclusions.
The interview handles the incident briefly but uses it to set up a broader admission: AI is starting to generate political exhaust that didn't exist on Twitter.
Two pieces of that exhaust in the episode:
The permanent-underclass scenario. Altman invoked Sam Lessin — the former Facebook VP and well-known early-stage investor and commentator — who has publicly argued that AI breaks the industrial mobility story and replaces it with two equally bad options: stay alive and receive abundance, but lose meaning from work, or the work ladder is gone and you fall into a permanent underclass. Altman named this as something he worries about. In his own blog post responding to the recent attacks, he separately conceded he had "underestimated the power of words and narratives" — a quiet apology wrapped around months of blowback.
Manufacturing actuators. Altman said OpenAI will go as far as manufacturing its own robot actuators. Not full robots — actuators, the motors and joints that are the physical bottleneck. The US supply chain does not have cheap, reliable, high-volume actuator capability. If you want physical AGI, you vertically integrate from hardware through algorithms. It puts OpenAI on a direct path to compete with Tesla, Figure, and Boston Dynamics — and undermines, quietly, the "broad + distributed" narrative from §5. A company vertically integrating from models to actuators to social-feed-as-needed is not a plural ecosystem. It's a vertically integrated platform.
8. Sam's Two Futures vs Greg's Third — the Ideology Comes Out
Altman sketches two futures:
- Future A: AGI succeeds, abundance arrives, humans lose the meaning that work provided.
- Future B: AGI creates a permanent underclass; society splits into those with access and those without.
Neither is good.
Brockman interrupts — he wants a third. He calls it a resilience ecosystem: instead of a small cluster of companies holding AGI-tier capability, many AI builders coexist, distributed like the power grid or the internet, and accountability stays human:
"You can't abdicate responsibility to an AI agent. You can delegate the mechanical parts, but accountability has to stay with a human."
This is where the two companies' ideologies come out in the open. OpenAI's narrative is "broad + distributed"; Anthropic's Mythos + Glasswing is "concentrated + controlled." The piece has been setting this up since §5; §8 lets it land.
It is also where the contradiction gets loudest. A company that kills Sora because it can't justify compute against AGI, that builds its own actuators, that owns a Superapp's entire fat layer, and that wants to live inside your calendar — is not a plural resilience ecosystem. It's a vertically integrated platform describing other labs' concentration as the problem. Brockman's vision is coherent. OpenAI's execution of it is not the same thing.
9. The Partnership Itself Is a Moat
Brockman, asked about the Altman-Brockman dynamic:
"Sam raises the target. I force the refocus."
Ten years in, the operating mechanism is that tension: ambition × execution discipline.
The reason this matters, and the reason it's worth one paragraph not ten: every other frontier-lab founder pair has fractured. Google Brain, DeepMind, Anthropic, Inflection, Mistral — every single one has had a founder exit or breakup. Altman-Brockman survived the 72-hour board coup in November 2023. Continuity is itself a moat, and it's one OpenAI's competitors can't buy.
10. The Ironies the Interview Glides Past
The case for a Great Reset is real. The case as Altman and Brockman tell it leaves several ironies on the floor. Worth naming before closing:
- OpenAI's own safety gating. GPT-4 was capability-gated for months. Voice was restricted. Image generation rolled out in tiers. "Broad personal AGI" is a brand claim. OpenAI is the company whose name stopped meaning "open."
- Personal AGI = lock-in. The same article describes Personal AGI approvingly as an OS you can't leave. That's Meta/Google-class switching cost, named approvingly while Anthropic's much more modest Mythos gating is named as elitism.
- Vertical integration ≠ pluralism. Brockman's resilience ecosystem and OpenAI's actuator-to-Superapp vertical are in tension. You don't get to be both.
- "Fat layer kills thin wrappers" is a centralization argument. A few companies with deep capability stacks, everyone else dependent. Structurally identical to the Anthropic critique OpenAI just deployed.
- "Fear-based marketing" — coming from the CEO who, in 2023, told the US Senate that AI could cause "significant harm to the world," is the most-interviewed prophet of existential risk in the industry.
Holding these open doesn't invalidate the Reset. The product moves are real. The ideology is coherent on its own terms. But readers should see what's being asserted, and what's being elided.
11. The Deeper Read: This Reset Was Forced By Anthropic
What follows is my (LJ's) reading — not what's said in the interview. It's the version you get when you stitch this Reset together with the H1 2026 ARR numbers, the past six months of other podcast appearances, and the model-release cadence on both sides.
1) OpenAI didn't kill Sora because they believed in Personal AGI. They killed it because the H1 2026 revenue split with Anthropic made the misallocation visible. Anthropic's ARR pushed into the 1M/year. OpenAI's top-line ARR is also growing, but the structure is consumer subscriptions — and the per-account ARPU on a high-value B2B agentic customer is orders of magnitude above a ChatGPT Plus seat. One enterprise contract = hundreds of consumer subs. Sora's problem wasn't quality. It was that its user profile sat on the wrong price curve.
2) The cut isn't about Sora. It's about freeing up compute and product surface for Codex to take over. Codex is the under-narrated centerpiece of this Reset, but the interview wraps it in Personal AGI language. Codex is OpenAI's answer to Claude Code — the product Anthropic has actually used to eat the developer market over the last twelve months. Read Brockman's "Codex isn't for engineers, it's for everyone" through this lens: he's trying to bracket Anthropic's core stronghold under a wider story.
3) OpenAI hasn't actually been spending serious resources on agentic until now. The biggest agentic capability jump of the past year was made by Anthropic — starting from coding, then generalizing. OpenAI got pulled bottom-up by Sora's early traction and let the coding/agentic frontier slip. Anthropic stayed top-down: from day one, agentic + coding was the line and they didn't deviate. OpenAI is now playing catch-up.
4) Anthropic's "strategic clarity" is partly market-forced. When they came up, the consumer market was already saturated by OpenAI and Google. Survival required B2B + coding focus, which is why they didn't even build a serious multimodal product. The clarity is real, but the market structure handed it to them.
5) "AI for every user" is a placeholder. The real target is high-value agentic B2B. No consumer is going to spend serious money letting an agent grab concert tickets and order groceries — that price-density doesn't pay GPU bills. The customers who actually pay are enterprises: code, customer service, knowledge work, process automation. "Personal AGI" is the public-facing story. B2B agentic is where the money is.
6) Which is why the Anthropic attack was necessary — true or not. Read §5's "fear-based marketing" segment through this frame and it becomes completely coherent: the goal is to chip away at Anthropic's B2B market share. Mythos + Glasswing has been too successful at building Anthropic's "we take enterprise safety more seriously" brand, and OpenAI has to disrupt that narrative before H2 2026 enterprise procurement cycles. Whether the criticism is fair, whether Anthropic is just executing a published RSP — doesn't matter on this battlefield. What matters is putting one more question mark in a CIO's head.
7) Chinese models are catching up fast, but OpenAI's structural advantages still hold. MiniMax, Z.ai, and Kimi already locked into the same B2B + coding niche (Chinese big-tech doesn't leave room for startups in the consumer space, so paths converge naturally). But on talent density, top-tier compute, and entrenched enterprise contracts, OpenAI still leads Anthropic and the Chinese frontier labs. Late doesn't mean lost — provided this Reset actually executes.
Stitched together, the honest description of the Great Reset is:
In H1 2026 OpenAI looked at the revenue structure and saw they had handed the agentic home turf to Anthropic. So they ran a company-wide reset — kill Sora-tier consumer exploration, redirect all resources to Codex/agentic to claw back B2B share, wrap the pivot in "Personal AGI for everyone" public narrative, and attack Anthropic's B2B moat with "fear-based marketing."
Altman's story in the interview is "we are building Personal AGI." The underlying story is "we are taking enterprise agentic back from Anthropic." Both can be true. Don't only see the top layer.
What to Watch
Skip the recap. Three things will tell you whether this Reset is substantive or theater:
- Does Sora's replacement model actually ship with the promised image/writing capability, and does it perform against Gemini / Imagen / Midjourney — or does "we cut Sora to focus" turn out to be "we cut Sora because it was losing"?
- Does Glasswing expand or contract through H2 2026? If Anthropic keeps growing the enterprise book through capability-gated access, "fear-based marketing" will look less like a joke and more like the market signal it is.
- Does the actuator team get announced — real hires, a plant location, a robot in 2027? If yes, Personal AGI is a whole-stack play. If it never shows up, the robotics line was rhetoric.
Ten years ago OpenAI was a non-profit research lab. Ten years later, it's a company simultaneously manufacturing robot actuators, litigating with Elon Musk, and trying to live inside your calendar.
Great Reset — the thing being reset is not the product line. It's what kind of company OpenAI now is.
Broad and distributed, or vertically integrated and bonded to you for life? The Core Memory interview insisted on the first story. The product moves all point to the second.
Sources:
- Core Memory EP 67 — The Great Reset at OpenAI (YouTube)
- Big Technology Podcast — Greg Brockman on fat layer, Spud, 70–80%
- TechCrunch — "fear-based marketing" coverage
- Anthropic — Project Glasswing
- SF Standard — Musk v. Altman trial
- TechCrunch — Altman responds after attack on home
- Axios — GPT-5.5 "Spud" release
