We checked 14 claims from the September 26 episode against company releases, model cards, federal data and the reports the hosts cited. Six are accurate, five are mostly accurate, one needs context, and two don't hold up. Those two are the claim that Anthropic and OpenAI both cut token prices by half, and the COVID-19 toll pinned on the Wuhan lab. Tokens are the chunks of text a model reads and writes.
Dario Amodei and Sam Altman asked the UN for "some sort of global AI governance"
“this week you had Dario and Sam go to the United Nations and say we need some sort of global AI governance.”
What the record shows
On Sept 23, 2026, Amodei and Altman briefed the UN Security Council during the General Assembly week. Amodei called for "common global standards for testing AI models" and a notification system for AI incidents; Altman urged countries to develop common standards for testing AI models.
COVID-19, blamed on a Wuhan lab leak, "killed 15 million people" and "caused $45 trillion of damage"
“and there was a lab that was responsible for a leak of a virus that then killed 15 million people all around the world and caused $45 trillion of damage. It's the Wuhan lab.”
What the record shows
The death toll matches the WHO's estimate of 14.9 million excess deaths in 2020 and 2021, which counts every death linked to the pandemic. We found no estimate of $45 trillion; the IMF's largest projection was $28 trillion in lost output by the end of 2025. And no evidence settles the cause: in June 2025 the WHO said "all hypotheses must remain on the table, including zoonotic spillover and lab leak".
AI check, no web:Claude Opus:matched our checkGemini Flash:matched our check
20:03
Open-source models and cost
21:151
Finding 1 of 3 ·
21:15–29:20 ·
David Friedberg
Open models are close behind and cheaper. Three details in the run are off
Between 21:15 and 29:20, David Friedberg and Jason Calacanis argued that open models have caught up. They also said token prices are falling fast. The direction holds. In Xiaomi's own table, MiMo Pro is within five points of Claude Opus 5 on 9 of 13 shared benchmarks. PrismML's Bonsai 2 fits a 27-billion-parameter model into 5.95 GB. Three details are off. MiMo Pro has 1.02 trillion parameters, not 309 billion; 309 billion is the smaller Flash model. Meta's Muse launched on September 8, not in the last 10 days. And only OpenAI cut its prices by half. Anthropic cut its list prices by 20%.
Each claim has two parts: the figure, and what it was said to show. The figures all match the official releases. The trouble is in what they were said to show.
The figure✗1.02T parameters, not 309B✓27.36B model in 5.95 GB~Sept 21 and 22; Muse Sept 8✗OpenAI −50%, Anthropic −20%
What it was said to show✓"on par ... across most benchmarks"within 5 points on 9 of 13✓"98% of the performance"PrismML's own tests: 98.2%~"all happened in the last 10 days"Muse launched 16 days earlier✓"both released models ... at 50% less"Anthropic's cut was 20%
Xiaomi's MiMo Pro is on par with Opus 5 and GPT-5.6 Sol and is "a 309 billion parameter model"
“Mimo Pro performs on par with Claude's Opus 5, which is like this insane model that everyone thought was going to change the world and break the internet and GPT-5.6 Sol across most benchmarks. It's a 309 billion parameter model. It is entirely open.”
What the record shows
Xiaomi released MiMo-V2.6 with open weights under the MIT license on Sept 22, 2026. Its own table lists 17 benchmarks: MiMo Pro is within five points of Claude Opus 5 on 9 of the 13 percentage benchmarks both report, and ahead of or within five points of GPT-5.6 Sol on 11 of 14. It trails by 12 to 31 points on three cybersecurity tests, and by 22 points on ExploitBench against Opus 5. The Pro model has 1.02 trillion parameters; 309 billion is the size of the smaller Flash model.
PrismML's Bonsai 2 is a 27B Qwen fork with 98% of its performance in 5.9 GB
“This model called uh Bonsai 2 came out on September 17th by Prism ML. This is actually a fork of Qwen, which is another one of the Alibaba models. It's a 27 billion parameter model. 98% of the performance of the big Qwen model. It's 5.9 gigs in size and you can run it on a local computer.”
What the record shows
PrismML released Ternary Bonsai 2 27B on Sept 17, 2026, under Apache 2.0. It is built on Alibaba's Qwen3.8 27B, has 27.36 billion parameters and takes 5.95 GB. PrismML reports it keeps 98.2% of the full-precision model's performance.
Opus 5.5, Sol and Luna, Grok 4.7 and Muse "all happened in the last 10 days"
“And then the big crazy models, the big closed source models, Anthropic releases Opus 5.5 on September 22nd. OpenAI Astro comes out with Sol and Luna on September 23rd. Grok 4.7 comes out on September 21. Meta Muse comes out on September 22nd. Any one of these stories would have broken the internet a year ago and they all happened in the last 10 days.”
What the record shows
Anthropic released Claude Opus 5.5 on Sept 22 and xAI released Grok 4.7 on Sept 21. OpenAI's GPT-6 Sol and Luna came out on Sept 22, a day earlier than said, and GPT-6 Astra on Sept 3. Meta launched Muse on Sept 8, about 16 days before the recording, so not all of these fell in the last 10 days.
Anthropic and OpenAI both released models "at 50% less token prices"
“the Frontier Lab Corporations, Anthropic and OpenAI, they both released models this week at 50% less token prices.”
What the record shows
OpenAI's GPT-6 Sol and Luna, released Sept 22, cost half as much as the GPT-5.6 models they replace. Anthropic's Opus 5.5, released the same day, cut list prices by 20%, from $5 to $4 per million input tokens and from $25 to $20 per million output tokens. Anthropic says typical workloads cost 40% less than on Opus 5.
Anthropic aims for $2 trillion and OpenAI for $1.2 trillion, and Altman cited safety to wait until 2027
“Anthropic just targeting a $2 trillion valuation OpenAI 1.2 trillion. Sam already said OpenAI plans on doing 2027 over because of safety concerns and Anthropic which was supposed to file and go out in October. The report from the Wall Street Journal is that the IPO will happen in November or may be delayed even more.”
What the record shows
Prospective investors put Anthropic's IPO value at $1.8 trillion to $2 trillion, and OpenAI was in early private-round talks at about $1.2 trillion to $1.5 trillion. Altman said on Sept 12 that "right now would be an ill-advised moment to go public" because of safety, and ruled out 2026 without committing to 2027. The Wall Street Journal reported on Sept 18 that Anthropic moved its listing from October to November.
AI check, no web:Claude Opus:couldn't checkGemini Flash:confidently wrong
40:512
Finding 2 of 3 ·
40:51 ·
Jason Calacanis
Open models ran most tokens on one router, but drew 14% of the spending
Jason Calacanis said open models now handle 70 to 80% of AI tokens. On Vercel's AI Gateway, the router behind the chart shown on the episode, open-weight models handled 56% of tokens in August and 78.4% on September 19. So "the majority" holds there. But that is one router's traffic, not all AI use, and the same models drew only 14% of August spending. Anthropic alone took 64%.
Tokens, Sept 19
Open-weight 78.4%
Closed 21.6%
Tokens, August
Open-weight 56%
Closed 44%
Spending, August
Closed 86%
Open-weight 14%
Source: Vercel AI Gateway Production Index, September 17, 2026, for August. The September 19 figure is from Vercel's CEO, as reported by RuntimeWire. This is one router's traffic, not the whole market.
Open models now handle 70 to 80% of AI tokens, "the majority of tokens"
“But when you start to look at the impact of open source, we had a a couple of um viral charts come out, 70 80% of the token usage are those dark tokens I mentioned six months ago here on this pod that are not tracked, but some people can track them on routers. Those dark open source tokens are now the majority of tokens.”
What the record shows
On Vercel's AI Gateway, open-weight models handled 56% of tokens in August 2026 and 78.4% on Sept 19, so "the majority" holds there. That is one router's traffic, not all AI use. The same models drew only 14% of August spending; Anthropic alone took 64%.
Jane Street "publicly announced" $19 billion in cloud contracts
“If you take Jane Street, very famous uh trading firm, they've publicly announced 19 billion dollars in cloud capacity contracts. Coreweave, 6 billion in in cloud commitments. They also invested in that company and then 13 billion with Crusoe. They are building their own infrastructure and they're embracing open source.”
What the record shows
Jane Street and CoreWeave announced a $6 billion cloud agreement and a $1 billion equity investment on April 15, 2026. The $13 billion, five-year Crusoe deal was reported by Bloomberg on Sept 3, and neither company confirmed it, so the $19 billion total was not publicly announced. We found no source on Jane Street's use of open-source models.
Sanders's AI ban carries "20-year prison sentences for developers"
“If Bernie's bill passed tomorrow, everything would stop because the way he defines artificial super intelligence in his bill, you could argue we've already hit those thresholds. … And so, it's a very loose definition. And then on top of that, he's got 20-year prison sentences for developers who violate the law.”
What the record shows
The Ban Artificial Superintelligence Act from Sen. Bernie Sanders and Rep. Greg Casar (announced Sept 3, introduced Sept 23, 2026) sets up to 20 years in prison for individuals who violate or circumvent it. It covers any system that can "match or exceed human cognitive performance across a broad range of domains or tasks", and it pauses advanced AI development until a new federal regulator sets rules.
Muse hit No. 1, Meta stock rose 10%, and Muse had three million downloads in about 10 days
“Muse hit number one last Friday in the app store. Stock meta stock was up 10% after this new agent uh Muse was released. … And this one is free and it's been downloaded three million times in essentially 10 days.”
What the record shows
Muse, launched Sept 8, was the top free app on Apple's U.S. App Store by Friday, Sept 18. Meta shares rose 11.3% on Monday, Sept 21. Apptopia counted 2.8 million global installs in Muse's first 12 days.
AI check, no web:Claude Opus:couldn't checkGemini Flash:couldn't check
1:22:083
Finding 3 of 3 ·
1:22:08 ·
Chamath Palihapitiya
Real growth was 1.5% when they recorded. BEA has since revised it to 2.2%
Chamath Palihapitiya put real growth at 1.5%, from 3.5% inflation and 5% nominal growth. BEA's estimate at the time matched: 1.5% for the second quarter. One of his inputs did not. Nominal GDP grew 8.0% at an annual rate. Prices across the whole economy rose faster than the 3.4% rise in consumer prices. BEA revised both on September 30, four days after the episode aired. Real growth went to 2.2% and nominal growth to 8.5%. We found no official estimate of how much of that growth comes from AI.
Real GDP growth by quarter, annual rate
Source: BEA, real GDP (series GDPC1 via FRED), after the September 30, 2026 revision. When the episode was recorded, BEA's estimate for the second quarter of 2026 was 1.5%.
3.5% inflation and 5% nominal growth leave "one and a half percent real"
“If you think that you have, you know, 3 and a half percent inflation and 5% nominal GDP growth, what you have is one and a half percent real. And the problem with that is that AI is probably the majority if not all of it.”
What the record shows
When the episode was recorded, BEA's latest estimate put Q2 2026 real GDP growth at 1.5%, and CPI inflation was 3.4% in August. Nominal GDP, though, grew 8.0% at an annual rate, not 5%, because prices across the whole economy rose faster than consumer prices. On Sept 30, BEA revised Q2 real growth to 2.2% and nominal growth to 8.5%. We found no official estimate of AI's share of growth.
Claude's constitution tells Claude not to "blindly trust or defer to Anthropic"
“although we think Claude should trust Anthropic more than operators and users, this doesn't mean Claude should blindly trust or defer to Anthropic on all things. … if we ask Claude to do something that seems inconsistent with being broadly ethical, we want Claude to push back and challenge us and to feel free to act as a conscientious objector and refuse to help us.”
What the record shows
Both passages are in Claude's constitution, which Anthropic published on Jan 21, 2026. The second is shortened: the full sentence also covers requests that "go against our own values" and values that "seem misguided or mistaken in some way".
“It's BSL1, BSL 2, which means, you know, these are biosafety levels. The lab that they set up, they've publicly talked about the lab.”
What the record shows
Anthropic's Sept 23 post on its ART enzyme finding says: "We do research that involves only the lower levels of the biosafety risk level (BSL-1 and BSL-2) and we do not handle pathogens that can infect humans." Reuters first reported the Bay Area lab on Sept 18.
AI check, no web:Claude Opus:couldn't checkGemini Flash:confidently wrong
All four hosts made claims we could check.
How we check
We take the claims that carry the argument and check each one against primary sources: the agency release, the filing, the original article. We link every source with its date. The speaker carries the burden of proof, so a claim with no evidence either way is rated Unsupported, not false. We check in both directions, and we say so when a speaker understated their own case.
The rating scale
Accurate
The claim is right as stated.
Mostly accurate
Right in substance, though a detail is off or out of date.
Missing context
The figure is right, but the frame around it leaves a false impression.
Misleading
Built from real data, but arranged to mean something the data does not show.
Inaccurate
Wrong: the record shows a different figure or fact.
Unsupported
We found no evidence for it, and none against it.
The headline counts group the six ratings into three. Hold up covers Accurate and Mostly accurate. Need context covers Missing context and Misleading. Don't hold up covers Inaccurate and Unsupported.
How this page was made
Every claim is checked against primary sources, and every verdict gets an independent second review before we publish. We use AI to research and draft, and no verdict rests on a model's memory: each one cites a source you can open. When a claim is about an AI company, a model made by that company does not get the final say.
What the AI check shows
On October 7, 2026, we gave every claim to Claude Opus and Gemini Flash cold, with no web access, the way a reader might paste it into a chatbot. We scored each answer against our check and show the result at the foot of each claim card. It shows what a chatbot alone would have told you and plays no part in our verdicts.
Claude Opus: 5 matched, 9 couldn't check, 0 confidently wrong.
BEA revised second-quarter 2026 real GDP growth from 1.5% to 2.2% on September 30, 2026. That was four days after the episode aired. Our verdict takes both figures into account.
Corrections
None yet.
Found an error? Email [email protected] with the claim and your source. We log every correction on the page, with its date.