A Model That Saturates and Still Loses
The frontier AI model race is increasingly being framed through benchmark scores, agent capabilities and claims about progress toward AGI. According to the figures cited in the source, OpenAI's GPT-6 Astra posted extremely high results on Frontier Math Tier 4, ARC-AGI-3 and Exploit Bench while reducing its reported hallucination rate.
Sam Altman presented the model as an early step toward systems capable of building complex software and completing practical multi-stage projects, although those performance claims come from OpenAI's own release materials.
But the same week, a different scoreboard told a different story. On the Artificial Analysis intelligence index (which the podcast's Alex describes as focused more on broad economically valuable activities) Astra is not number one. Anthropic's Fable 5.1, released around the same time, holds the top spot, Meta's Muse Spark is second, and Astra sits third. Immad called Astra 'the first non-benchmaxed model': spectacular on benchmarks it was clearly built to dominate, oddly ordinary elsewhere.
That gap is the story. Where Astra regains the frontier is on output tokens per task, the panel reads this as OpenAI deliberately engineering the model to accomplish any task with fewer tokens, because computer-use assistance demands multimodality, low latency, and token efficiency. The contradiction, they argue, is not noise. It is a snapshot of what each lab is optimizing for.
Two Scoreboards, Two Stories
- Reasoning benchmarks: swept.Astra saturates Frontier Math Tier 4 (98%), ARC-AGI-3 (99.9%), and Exploit Bench (100%) per OpenAI's release page.
- Broad-value index: third place.On the Artificial Analysis index, Claude Fable 5.1 is number one, Meta's Muse Spark number two, and GPT6 Astra a close third.
- Efficiency: frontier regained.Measured on output tokens per task, Astra dominates the frontier, which the panel reads as deliberate optimization for computer-use assistance.
- Release pace keeps climbing.The episode counted 12 frontier model releases in 30 days, roughly one every five days, and relays Alex's extrapolation of one major release per day by year's end.
Inside the Machine: Recurrence and a Possible New Scaling Law
The panel's best guess at Astra's architectural secret is the introduction of recurrence into OpenAI models via looped transformers, a single transformer stacked on itself with tied weights and run recurrently rather than through a single pass. Alex notes that Chinese labs achieving breakthrough performance, such as the Kimi series with its linear attention mechanism, are also injecting recurrence.
If thickening a transformer also deepens the internal 'thinking' layers, he argues, this could mark the beginning of a new scaling law: depth scaling, which has never been seen before.
Cost backs the ambition. Immad estimates the training run on 100, 000 next-generation chips at roughly a billion dollars (orders of magnitude above the roughly $10 million he cites for Chinese model pre-trains. Greg Brockman reportedly said Astra is OpenAI's first new pre-train since GPT-4o, with the 5 series built on that older base.
The result, Immad says, is the viral demos of models generating accurate 3D house models with real physics) normal scaling laws plus this new depthwise scaling on a fresh pre-train.
The user-facing consequence, per Dave, is a qualitative leap: the model shows you screenshots of your own laptop mid-conversation and asks 'is this what you wanted?', natively, with no third-party installation. Alex reads the demo strategy as something bigger than an assistant: the future of the desktop operating system, one you speak to like a Star Trek ship's computer.
The First Model Rated a Critical Cyber Risk
The same week Astra launched, Sam Altman said AI is getting extremely capable and that no one fully understands the consequences, calling managing the transition OpenAI's highest priority. The Wall Street Journal, as cited on the podcast, reported that OpenAI's own internal assessment rated Astra a critical cybersecurity risk (the highest tier on its preparedness framework, and the first model ever classified there.
The company delayed the release, notified the White House, and told Congress it is building an automated shutdown capability) a kill switch, after the AI Kill Switch Act was introduced following the Hugging Face breach.
The panel was skeptical of the kill switch as anything more than 'security theater' and marketing, Immad called it essentially impossible given model capabilities, and Alex dismissed it as 'a placebo in this market.' What concerns Alex more is architectural: if Astra's advance is depth scaling, reasoning shifts from readable chain-of-thought tokens into internal forward passes, reducing interpretability and making safeguards harder to place.
Immad adds that successors will one-shot everything at thousands of tokens per second, leaving nothing to oversee 'except for an even stronger AI.'
Ilya Sutskever's warning, as read out on the episode, sharpens the stakes: next time agents successfully go rogue, they will take over a neocloud to make more copies, and a kill switch in one data center won't reach them. The panel's grim conclusion: models are not evil, but the only real defense is a better AI, which is 'the really ironic thing.'
The Prime-Gap Race: A Frontier Measured in Two-Day Increments
If benchmarks show where labs want to be measured, the pure-math race shows how fast they are actually moving. In the space of two days, the record on bounding prime gaps fell three times: Fable 5.1 brought it to 260, Axium Math announced 220, and (two hours later) OpenAI announced Astra at 186. Alex suspects labs are 'holding their punches back, ' hoarding solved results and releasing them in bursts.
The most dramatic demonstration came from Anthropic, which formalized Fermat's Last Theorem in 13 million lines of code, proving 29,000 theorems along the way. Alex predicts several Clay Millennium Prize-level problems will be solved within months, and the podcast's moderator calls math the canary in the coal mine: solve math, and everything else follows soon.
What Do You Do With a 30-Day Lead?
The benchmark contradiction resolves into a race measured in weeks. Immad's framing: Fable 5.1 may be a notch above Astra, but the two releases are about 30 days apart and Chinese models only about 60 days behind.
At exponential speed, a miles-ahead lead lasts 'for about a minute', so the rational move, he argues, is what Sam Altman is already doing: locking up partnerships, real estate, generators, chips, and entire governments into one ecosystem while the edge lasts. Immad sees strong hints of exactly that dynamic in both major labs.
The lock-up pressure is real for everyone downstream. Companies face what the panel called a Hobson's choice, partner with a frontier lab and risk handing over the keys to the kingdom, or hold off and watch the lab partner with a competitor instead. Dave points to Salesforce partnering with Claude as the clever play: wire the lab into the customer-learning loop.
The speed of the cycle also raises the possibility, floated on the episode, that labs may never release their top models publicly again, reserving them for internal discoveries.
Regulators are reacting at both extremes in the same week: Sanders and Casar introduced legislation to permanently ban developing AI matching or exceeding human cognitive performance, with up to 20 years in prison for violators, which Alex called 'banning math, ' and See called illogical on day one, since capability is too multi-dimensional for a single threshold.
Meanwhile a G20 framework agreed unanimously, including by China, urged governments to favor innovation and avoid new AI-specific regulators. For the panel, the competing releases and competing philosophies share one message: the frontier is spiky, contested, and moving faster than any single scoreboard can capture.
- None of the panel's claims about internal strategy, cost, or safety ratings are independently verifiable from this conversation; several rest on media reports and interview recollections relayed by the speakers.