# OpenAI's $7 Billion Buyback and Claude's Breakthroughs

An episode of [AI Chat: AI News on Claude, ChatGPT, Gemini, Grok, Perplexity](https://invisiblenetwork.ai/podcast/1669990482), 11 Aug 2026, 15 min.

Audio: https://content.rss.com/episodes/365073/3061853/ai-chat-podcast/2026_08_11_20_26_23_d7d4daae-3ef1-499a-9c2d-a2ef88e89e53.mp3

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! In this episode, we explore OpenAI's recent $7 billion employee share buyback amidst an $852 billion valuation, highlighting its implications for the company's future as it competes with Anthropic. We also discuss groundbreaking advances in AI models tackling unsolved mathematical problems, particularly Anthropic's efforts with the Riemann hypothesis and OpenAI's Astra model achievements. Chapters 00:00 OpenAI's $7 Billion Buyback 03:38 Advances in AI Mathematics 08:10 Comparative AI Model Performance 12:04 Revealing Hidden Reasoning 14:00 Latest AI Developments Show Links • Get the top 80+ AI Models for $8.99 at AI Box: https://aibox.ai • How I Grow and Scale My Business with AI: https://www.skool.com/aihustle • Get the AI Chat Daily Newsletter: https://www.aichatdaily.com/newsletter

## Transcript

[0:02] Welcome to the AI Chat Podcast. I'm your host,
[0:04] Jaden Schaefer. Every day I cover cutting -edge
[0:07] AI news and talk with the leaders behind it,
[0:09] breaking down what it means for your life and
[0:11] business. OpenAI has just bought back $7 billion
[0:22] worth of employee shares at about an $852 billion
[0:26] valuation. Anthropic's unreleased model has advanced
[0:31] the Riemann hypothesis. They spent 31 million
[0:33] tokens to do this. OpenAI's Astra model has solved
[0:37] a 10 long open math problem for $2 ,000 worth
[0:41] of tokens. So it wasn't cheap, but it was able
[0:43] to get it done. Alibaba released Quen 3 .8 Max,
[0:47] which is a 3 .4 trillion parameter model. It's
[0:50] basically competitive with Claude. Researchers
[0:53] right now are extracting hidden reasoning from
[0:55] Claude, ChatGPT, and Gemini. They're using a
[0:58] trick with the API, and they're comparing the
[1:00] reasoning of different models with the open source
[1:03] or open weight model. that are Chinese versions
[1:05] to see if they ripped off those models. Almost
[1:08] everyone I know is paying $20 a month for some
[1:11] AI tool, whether that's ChatGPT, Cloud, Gemini,
[1:13] Canva, there's so many out there. And the one
[1:16] thing that I will say to all of them is I would
[1:18] love for them to check out... AI box .ai, which
[1:21] is my own startup that lets you access 80 different
[1:23] AI models all on one platform. Our cheapest tier
[1:26] is $8 and 99 cents a month. It is incredibly
[1:29] cheap. It's the price of a coffee and you get
[1:31] access to over 80 different AI models. There's
[1:33] image models, there's audio models, there's video
[1:36] generation models like Google VO three. And of
[1:39] course you have all of the text and reasoning
[1:40] models like cloud chat, GPD and Gemini. If you
[1:42] want to get access to all of that in one place
[1:44] and also have an MCP, which basically is a link
[1:47] that lets If that's your main model that you
[1:50] use, generate images or let's cloud generate
[1:52] audio or let's cloud generate video. So you can
[1:55] pull any model into any model. If you want to
[1:57] check that out, it is AI box .ai. I'll leave
[1:59] a link in the description. OpenAI has just completed
[2:02] a $7 billion buyback of employee shares. They
[2:06] did this, which I think a lot of people are kind
[2:08] of shocked by, at an $852 billion valuation.
[2:12] Now, if you remember, this is the exact same
[2:14] valuation that they had back in March. We are
[2:16] now many months later. And if you look at a company
[2:18] like Anthropic, which month over month is adding
[2:20] like $100 billion to their valuation every single
[2:24] month, basically, this is not, I mean, you could
[2:27] look at this a couple different ways, but like
[2:29] technically that's not great. Their valuation
[2:31] hasn't grown since March. I think this is basically
[2:35] they're trying to keep the staff from leaving
[2:37] and going to Anthropic or going to Gemini. But
[2:40] also, I think this is basically showing they're
[2:42] not in a big rush to go public anytime soon.
[2:45] This $7 billion tender offer is one of the largest
[2:48] employee buybacks a private company has ever
[2:51] done. I think that's no shocker. But this is
[2:53] very consistent with what OpenAI has done. They
[2:56] have a pattern of doing really regular secondary
[2:58] sales. They've done this for the last three years
[3:01] where they'll let the employees sell at a regular
[3:03] basis. Sam Altman told the staff last month that
[3:07] OpenAI missed a bunch of internal financial targets
[3:10] over the last year, but he expects that the next
[3:12] year is going to be stronger. I mean, there's
[3:14] the elephant in the room and the no surprise
[3:16] is that Anthropic just absolutely mopped the
[3:19] floor with their growth. And OpenAI, I don't
[3:21] think has actually shrunk. They probably just
[3:23] went stagnant or they kind of plateaued for a
[3:25] minute. And the reason why is just because Anthropic
[3:28] grew so fast, they just took a lot of the oxygen
[3:30] out of the room. Is it going to stay that way
[3:32] forever? I personally don't. think so. I'm heavily
[3:35] testing both Claude and chat GPT. And right now,
[3:38] my main tool that I'm using is chat GPT work,
[3:41] I switched from Claude co work about two weeks
[3:44] ago. And I've been really impressed Claude co
[3:46] work or chat GPT work can do a lot of things
[3:49] that Claude co work didn't, it feels like opening,
[3:51] I felt pretty threatened, and they put a lot
[3:53] of time and energy into getting this right. So
[3:55] I love a good healthy rivalry, I think it was
[3:57] going to go back and forth. I'm excited if we
[3:59] have other competitors, if it feels like Google
[4:01] gets a little bit more in this space. You have
[4:03] perplexity in there. But overall, I think OpenAI
[4:06] has a superior product right now. And so I think
[4:09] they might start getting back a little bit of
[4:11] that enterprise that they've lost. Anthropic
[4:13] was reportedly profitable earlier this year.
[4:16] I think that's definitely going to threaten OpenAI
[4:18] as far as if Anthropic can reach the stock market
[4:22] first. If they can do their IPO first, they're
[4:24] going to get a lot of... a lot of buyers that
[4:27] would possibly have invested in OpenAI if not.
[4:30] So I think definitely there's a timing issue
[4:33] or there's a motivation to get to an IPO first
[4:37] for OpenAI and Anthropic. But right now, if they're
[4:40] holding their valuation flat and they're doing
[4:42] their tender offer for all their employees now
[4:44] rather than waiting for the IPO, OpenAI is basically
[4:47] buying time to prove that their business model
[4:50] is going to work at scale, especially for enterprise
[4:52] and their API products that have real... recurring
[4:55] revenue also i mean if you want to be like sort
[4:58] of pessimistic maybe you could say look, perhaps
[5:00] Sam Altman believes that the company is worth
[5:03] more and it's growing, but he just wants to buy
[5:05] back all the shares at a really good deal so
[5:07] he can go buy $7 billion worth of shares from
[5:10] the employees. And when they IPO, they can bump
[5:13] up the IPO price 50 % and now those $7 billion
[5:17] of shares will be worth $14 billion. The company
[5:19] that's worth over $850 billion, I don't know
[5:22] if that's really the main incentive here, but
[5:24] there's definitely a case to be made that perhaps
[5:27] those shares are at a discount. I think more
[5:29] like - Likely he's just trying to keep his employees
[5:30] happy and keep them around. But, you know, it's
[5:34] got to make you think that an IPO is not just
[5:36] weeks away if they're able to or if they're going
[5:38] to be buying back these shares from their employees.
[5:40] You know, if an IPO is around the corner, they
[5:42] probably just let them all sell on the open market.
[5:45] But if it's going to be months away, they probably
[5:47] will just buy them and wrap that up now. There's
[5:50] an unreleased anthropic model that is advancing
[5:53] the Riemann hypothesis. So basically this is
[5:55] a 150 -year -old math problem. There is a $1
[5:58] million bounty on this math problem, by the way,
[6:00] which is kind of cool that this exists. But by
[6:03] extending the range of numbers for... which it's
[6:07] been verified. This new model from Anthropic
[6:09] ran itself for a day and a half. It organized
[6:12] 60 subagents to test 650 different ideas and
[6:17] formalize the results in a way that mathematicians
[6:20] could check because not only do you have to solve
[6:21] these problems, but you have to explain how you
[6:23] got it. So this is a big part of it. The model
[6:25] orchestrated its own workflow without a human
[6:28] mathematician directing it. It was assigning
[6:30] tasks like idea generation, validation, paper
[6:33] drafting to different subagents. It used about
[6:36] 31 million tokens to produce this. Two anthropic
[6:40] mathematicians confirmed the results, which was
[6:42] then formalized in Lean, which is an open source
[6:45] proof assistant that lets other researchers verify
[6:47] it step by step. OpenAI's internal Astra model
[6:51] proved 10 major results this year and a separate
[6:55] anthropic effort disproved the Jacobian conjecture,
[6:58] which is a 1939 problem. So I think we're seeing
[7:02] a big shift in how AI is tackling big problems
[7:06] in math right now. I think AI is producing verifiable
[7:09] mathematical progress on a bunch of unsolved
[7:12] problems, but there's a lot of people in the
[7:15] math field that are debating what it means for
[7:17] credit and accountability when a machine is coordinating
[7:20] all this work. instead of a person, but also
[7:22] you can imagine, well, what if a person just
[7:23] used? You know, something like open AI, like
[7:26] right now, this is, you know, being done by the
[7:28] actual researchers are anthropic. But what would
[7:30] happen if, you know, they put out this unreleased
[7:32] model that's good at math and a regular person
[7:34] or a regular mathematician uses it to solve one
[7:37] of these problems? Who gets the credit for it?
[7:39] And maybe they say like, oh, no, I did it all
[7:41] by myself and I never even used AI. Would they
[7:44] get credit for it for solving these really hard
[7:47] kinds of problems? This is what everyone in the
[7:49] math field is currently debating. Now, I mentioned
[7:52] that. problem or those 10 problems that Astra
[7:55] solved. It basically the way this worked is this
[7:59] is also an unreleased model, by the way. So OpenAI
[8:01] and Anthropic, it's so funny. It feels like they
[8:03] do all their PR stunts like in at the same time.
[8:07] So it's like OpenAI goes and hacks Hugging Face
[8:09] and Anthropic's like, we hacked people. And then
[8:11] Meta's like, we hacked people too. And then Quinn's
[8:14] KimmyK3 is like, we hacked people or not Quinn,
[8:17] Moonshot's KimmyK3 is like, we hacked people
[8:18] as well with our unreleased model. And now we
[8:21] have like, I don't know, all of these different
[8:22] models are like, we solved, you know, these unsolvable
[8:25] hundreds. year old math problems. And they're
[8:27] like, Oh, we did too with our with our unpublished
[8:30] model. Anyways, I find it funny. But I mean,
[8:32] this one's pretty cool. So in this case, though,
[8:34] the Astra model, so not released, it solved 10
[8:37] long standing math problems that mathematicians
[8:39] had worked on for decades. It published a 250
[8:42] page proofs that verified with formal proof checking
[8:47] software, the solution like to actually get the
[8:50] solution to this cost about $2 ,000 in tokens
[8:52] to generate it. And on the one hand, I'm sure
[8:55] people are like, well, that's a lot of money
[8:56] to solve, you know, solve these math problems.
[8:58] I was like, if these are like famous math problems
[9:01] that have never been solved, they've been around
[9:02] for like hundreds of years, I mean, 2000 bucks,
[9:05] that seems pretty cheap, especially for a PR
[9:07] stunt like this. I mean, this is worth a lot
[9:08] more than $2 ,000 to open AI, this is probably
[9:11] worth, you know. uh, like 50 or a hundred million
[9:15] dollars worth of good PR for them. So it makes
[9:17] sense why they would do this. Um, but yeah, these
[9:19] are, the math is getting cracked by, um, these
[9:23] AI models. And this is stuff that's been, you
[9:25] know, pretty tricky for a very long time. Alibaba
[9:27] has just released Quen 3 .8 Max on Monday. This
[9:31] is a huge AI model. It has 2 .4 trillion parameters.
[9:34] It is on a bunch of benchmarks. It is right behind
[9:38] Anthropix Claude Fable 5 on the public leaderboard.
[9:42] So it's, I mean, it's right up there as number
[9:44] two. The company said that they're going to release
[9:46] the model's code and weights next week, making
[9:48] it free for developers. So any developer can
[9:50] go and download it. They can go modify it. I
[9:53] mean, the thing that's cool here, it's not like
[9:54] they're like, oh, we released a number two. model,
[9:57] that's great meta just released like a number
[9:58] four model yesterday or whatever. The thing that's
[10:01] amazing about it is that it's a number two model,
[10:04] and they're going to release the codes, open
[10:06] weights and like give this to anyone to go and
[10:08] download and use. on arena .ai's text leaderboard.
[10:12] It is behind Fable 5 and 3 Cloud Opus models.
[10:18] It's ahead of a bunch of other ones. Alibaba
[10:20] is coming to the openweight distribution. They
[10:23] had a little bit of a proprietary pivot earlier
[10:25] this year, which was kind of showing how a lot
[10:28] of Chinese labs like Moonshot and ByteDance were
[10:31] ever, there's like kind of this weird, and it
[10:34] feels like it happens in sync, but there's this
[10:35] weird moment for a lot of these. companies that
[10:38] are making these open weight models it feels
[10:39] like if they fall a little bit too far behind
[10:41] then they're like they go proprietary they fall
[10:43] a little bit behind then they're like okay now
[10:45] that we're getting caught up we got our mojo
[10:47] back we're going to release an open weight model
[10:48] even meta did that moonshots AI's Kimi K3 released
[10:53] last week had a 2 .8 trillion parameter model.
[10:57] So, I mean, this is 2 .4 trillion. It's very
[10:59] similar, but it's definitely, you know, Kimi
[11:02] K3 is a little bit more than Quen 3 .8 max. But
[11:06] the parameters don't 100 % matter, especially
[11:08] because this is currently ranking above it or
[11:11] on a bunch of different benchmarks. Researchers
[11:13] have found a way to extract hidden reasoning
[11:16] from Claude, GPT, and Gemini. This is a really
[11:19] interesting one for me. What they're doing is
[11:22] they go and trick smaller AI model variants into
[11:25] decrypting what larger models keep secret. So
[11:29] inside of any time that you ask an AI model like
[11:32] ChatGP or Cloud a question, it's not just running
[11:34] it to one model and spitting back the results.
[11:38] It takes your question and it gives it to like
[11:39] 16 or a whole bunch of different sub models that
[11:43] all review it and they all come up with their
[11:45] own response. And then it sends that to another
[11:48] model that looks at all the responses and it
[11:50] can solve Anyways, there's like these orchestrations
[11:53] of models in the background looking at it and
[11:55] reasoning through it and trying to come up with
[11:58] the best response possible for you. But apparently
[12:01] that is the vulnerability. So the same flaw has
[12:05] also leaked API keys and passwords before the
[12:09] companies patched it last month. So essentially
[12:11] the main model was kind of better protected,
[12:14] but these little sub models where they could
[12:16] actually go and exploit them. This particular
[12:18] attack exploits shared decryption keys between
[12:21] large and small model versions. So you're feeding
[12:24] encrypted reasoning to a less restricted. sibling
[12:28] model and then you're making it decrypt and expose
[12:30] the hidden chain of thought from the bigger and
[12:33] better model that is you know more encrypted
[12:35] or you know higher security but the little models
[12:37] like you're using their own models against them
[12:39] kind of moonshots open weight model kimmy k3
[12:42] produced reasoning outputs that were very similar
[12:44] apparently to claude opus 4 .8 and gpt 4 .0 their
[12:48] reasoning traced across 90 different test questions
[12:52] so basically the idea here is that kimmy k3 probably
[12:56] use model distillation. I mean, this isn't a
[12:58] shocker. Why wouldn't they? They obviously use
[13:00] model distillation on Opus 4 .8 and GPT -4 .0.
[13:05] OpenAI, Anthropic, and Google all patched that
[13:07] particular vulnerability after they were notified.
[13:10] Researcher Alexander Panfulov says that some
[13:12] reasoning content can still be reconstructed
[13:15] and closing the hole entirely would require rebuilding
[13:18] how APIs handle offloaded computation. So in
[13:21] some of these cases, it's really, really hard
[13:24] to patch this, this kind of exploit. And, you
[13:27] know, they'd have to really rebuild a lot of
[13:28] their infrastructure and rebuild how APIs handle
[13:31] stuff altogether, which, you know, obviously
[13:33] that's a mess. problem. I think this exposes
[13:36] a really big structural weakness in how frontier
[13:40] models offload reasoning. The companies share
[13:42] infrastructure for cost reasons, but that shared
[13:45] infrastructure becomes a backdoor when alignment
[13:48] training is different. So the distillation angle
[13:50] matters most for policy. I think when we're talking
[13:52] about distillation, if U .S. models are being
[13:54] systematically copied, you know, by all these
[13:56] open weight competitors, which some of them are
[13:58] American, right? We have Meta doing that with
[14:00] Meta's Muse Spark. And then, of course, we have
[14:03] K3. But basically, this is a technique that is
[14:08] this particular. method is a very plausible avenue.
[14:12] A couple other interesting things that happened
[14:13] this week. Brad Lightcap, OpenAI's longtime COO,
[14:17] is leaving after eight years to start something
[14:19] new. And Anthropic is going to start watermarking
[14:22] claw -generated text. They have to do this to
[14:24] comply with the EU AI Act. So there's a lot going
[14:27] on. Thank you so much for tuning into the podcast.
[14:30] If you enjoyed this episode and you haven't left
[14:32] a review on the show yet, it helps the show out
[14:35] so much. I'm sure you hear me say this all the
[14:37] time, but honestly, if you're one of the people
[14:38] that hasn't left a review yet. It really helps
[14:41] me show up in the algorithm. I read them all.
[14:44] So I just appreciate hearing from you guys. If
[14:46] there's any topics you want me to cover or any
[14:47] things that you enjoy, drop it in, drop it in
[14:50] a review, leave a comment. I will read it. And
[14:52] I really appreciate it. Also make sure to check
[14:54] out AI box .ai. If you want to get access to
[14:56] over 80 different image, audio, video, and text
[14:59] models, all in one platform for $8 and 99 cents
[15:01] a month. We also have 20 % off annual subscriptions.
[15:04] And if you want to get more tokens, there's a
[15:05] bunch of different tiers. You could do some really,
[15:07] truly incredible. things with AI box dot AI,
[15:10] especially our MCP that lets you connect AI box
[15:13] to Claude and Claude can generate images, video
[15:15] and audio all inside of Claude, which is what
[15:18] I use it for a ton. All right, catch you guys
[15:20] all in the next episode.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.