Latest / Today’s AI News / DeepSeek V4 Is Back and It’s Cheap, Claude Agents Broker Real Deals, Google Bets $40B on Anthropic

Transcript

DeepSeek V4 Is Back and It’s Cheap, Claude Agents Broker Real Deals, Google Bets $40B on Anthropic · Today’s AI News

• 0:00
  Welcome back to today's AI News, brought to you by 9X
• 0:02
  Productions. Imagine a heavyweight boxing
• 0:05
  match where the underdog didn't just show up with cheaper
• 0:07
  gloves, but built a completely separate training gym to bypass
• 0:10
  the Champions rules. That's exactly what's happening
• 0:13
  today. As a powerful new model emerges,
• 0:16
  running entirely outside the usual tech ecosystem.
• 0:19
  We're taking a deep dive into how the AI race just became a
• 0:23
  battle of infrastructure and price.
• 0:25
  Yeah. And the mission for a deep dive
• 0:28
  today is to really unpack these latest developments for you,
• 0:31
  focusing on the new capabilities, the practical use
• 0:33
  cases and, you know, these massive industry shifts.
• 0:36
  We want to make sure you get the absolute most important insights
• 0:39
  without like the total information overload that
• 0:42
  usually comes with this stuff. Right, because the new cycle
• 0:44
  right now is just relentless. So let's jump right into that
• 0:47
  that new separate training gym we are talking about the arrival
• 0:50
  of DeepSeek V4. Oh yeah, the whale.
• 0:53
  Exactly. The whale.
• 0:55
  If you hang around developer circles at all, you know they
• 0:57
  affectionately call it the whale.
• 0:59
  And I mean, we all remember their previous R1 release,
• 1:03
  right? That was the big moment when
• 1:04
  everyone realized cheap Chinese AI wasn't just some hypothetical
• 1:08
  future threat. It was here and it was highly
• 1:11
  competitive. Yeah, it was real wake up call
• 1:13
  for the industry and now V4 Pro just dropped as a preview and
• 1:18
  the raw metrics are frankly pretty intimidating.
• 1:21
  They really are. I mean right out of the gate V4
• 1:23
  Pro features a 1,000,000 token context window.
• 1:26
  Wait a full million just to visualize that for you.
• 1:29
  What does that actually look like in practice?
• 1:30
  So it basically means you could upload, say, the entire text of
• 1:34
  the Harry Potter series, plus a few dense legal textbooks on top
• 1:38
  of it, and the model holds all that information in its active
• 1:41
  memory simultaneously. That is just.
• 1:43
  Wild it is, but memory is really just one piece of the puzzle
• 1:47
  here. Internal evaluations are placing
• 1:49
  V4 Pro's raw reasoning capabilities in the exact same
• 1:52
  neighborhood as GBT 5.4 and Gemini 3.1 Pro.
• 1:56
  But you know, I do want to look closely at those evaluations
• 1:59
  though because it's not exactly a clean sweep across the board.
• 2:03
  Like depending on who is actually running the benchmark,
• 2:06
  the results look very different. V4 Pro takes the absolute top
• 2:10
  spot on Vals AI's Vibe Code Bench, which is a test that
• 2:15
  measures real world coding and debugging.
• 2:17
  Which is huge for developers. Definitely.
• 2:19
  But then on A's Intelligence Index, which tests a much
• 2:22
  broader spectrum of general reasoning and logic, it lands
• 2:26
  all the way down in Tier 4. It's sitting right next to
• 2:30
  Meta's muse, Spark. Yeah.
• 2:32
  So it has these laser focused areas of brilliance,
• 2:35
  particularly in coding, but it hasn't quite achieved that
• 2:39
  undisputed generalized crown yet.
• 2:41
  Right. And here's where it gets really
• 2:43
  interesting. It's the pricing, because
• 2:45
  comparing this to Western models is like comparing budget
• 2:47
  airlines to private jets. Oh, the cost difference is
• 2:50
  staggering. It's almost hard to believe.
• 2:52
  V4 pro costs just $1.74 for one million input tokens and $3.48
• 2:58
  for output right? And when you look at GPT 5.5
• 3:02
  open AI is charging 5 bucks for input and 30 for output.
• 3:05
  Opens at 4.7, is 5 and 25 deep. Seek isn't just undercutting the
• 3:10
  market, they are massively undercutting it.
• 3:12
  They're operating at a tiny fraction of the cost.
• 3:15
  But I mean, does being cheap even matter if they can't get
• 3:18
  the hardware to run it? Like with all the US export
• 3:21
  restrictions on NVIDIA chips, how are they sustaining this?
• 3:25
  Well, that is the crucial insight here, and it's why it
• 3:27
  matters so much. Huawei just announced that it's
• 3:30
  a send chips can fully support V4.
• 3:32
  Oh wow, so they don't even need NVIDIA?
• 3:35
  Exactly. I mean, objectively speaking, it
• 3:37
  proves the viability of a domestic alternative to NVIDIA.
• 3:41
  The export restrictions basically drove the creation of
• 3:44
  entirely separate AI infrastructures, so they aren't
• 3:46
  relying on the US ecosystem at all anymore.
• 3:49
  That completely flips the script, and since DeepSeek is
• 3:51
  driving down the cost of intelligence across the board,
• 3:54
  the discussion naturally shifts to how professionals are
• 3:56
  actually utilizing these tools in their daily work right now.
• 3:59
  Right, because it's no longer just a novelty.
• 4:01
  Exactly. I want to share 2 fascinating
• 4:03
  real world case studies of professionals maximizing this
• 4:07
  utility. First, there's a inset manager
• 4:09
  named Mayer. He got hit with a terrifying
• 4:11
  message from his accountant estimating a $10,000 tax
• 4:15
  liability. That is a completely paralyzing
• 4:18
  moment for an independent contractor.
• 4:20
  Right. Total panic.
• 4:21
  But instead of just accepting it, Mayer decides to use Claude
• 4:24
  to challenge the estimate. Yeah, he shared his specific
• 4:27
  contract structures and his actual invoices with the AI, and
• 4:31
  Claude dug into the regulations and found this highly obscure
• 4:35
  government circular. It classified his exact type of
• 4:38
  work as an export of services under the Goods and Services
• 4:41
  tax. So it completely changed his tax
• 4:43
  bracket. Yes, it saved him thousands of
• 4:45
  dollars. See, the deeper implication here
• 4:47
  is so important. AI is shifting from being just a
• 4:50
  text generator to a highly competent navigator of complex
• 4:55
  bureaucratic legal code. It's finding things that humans
• 4:58
  easily miss. It really is.
• 5:00
  And the second case is just as crazy.
• 5:01
  It involves a growth professional named Rishi, who is
• 5:04
  trying to hire a creative strategist.
• 5:06
  OK, so a lot of resumes sifting. Normally, yes, but he gave
• 5:10
  Claude the job description and instead of just asking for
• 5:12
  advice, Claude took control of his web browser.
• 5:15
  Wait, actual browser control? Yeah, it navigated to LinkedIn,
• 5:18
  searched for candidates, built a full spreadsheet of 20 high
• 5:22
  quality matches, and drafted personalized outreach for every
• 5:26
  single one of them. That is wild.
• 5:28
  I mean, he turned, what, two or three hours of manual work into
• 5:32
  a 15 minute review process? Exactly, but it raises this
• 5:35
  crazy question, if AI is doing the recruiting and writing the
• 5:39
  outreach, how long until an AI on the candidate side is the one
• 5:43
  replying? Are we just automating humans
• 5:45
  out of the handshake entirely? I mean, it's a very real
• 5:48
  possibility. It raises a huge question about
• 5:50
  the future of professional networking.
• 5:52
  When agents talk to agents, what do the humans even do?
• 5:54
  Seriously. And Speaking of AI taking over,
• 5:57
  let's talk about brand generation with claw design,
• 6:00
  because moving from administrative tasks to complex
• 6:02
  multi step creative overhauls is a huge leap.
• 6:05
  Oh, the design workflows right now are incredible.
• 6:08
  Right, let me walk you through the step by step guide on how to
• 6:13
  do a complete brand refresh because it's blowing my mind.
• 6:16
  Step 1. You just take a screenshot of
• 6:18
  your current, maybe outdated website and prompt an AI to
• 6:22
  create a refreshed brand description.
• 6:25
  So it sharpens the positioning, visual direction, typography,
• 6:28
  all of it. Exactly.
• 6:28
  It gives you the full strategy. Step 2.
• 6:30
  You generate a new logo and wordmark based on that, and the
• 6:35
  source actually noted that Chat GPT's image model was favored
• 6:38
  for getting the best logo results here.
• 6:40
  Yeah, it tends to be a bit cleaner with text elements.
• 6:43
  Good to know. Then step three, you paste your
• 6:45
  site name and that new bland description into the Design
• 6:47
  Systems tab at Claude dot I slash design.
• 6:51
  OK. And.
• 6:51
  Step 4. You upload the assets and just
• 6:53
  hit enter. It generates a full marketing
• 6:55
  page, a web app page and slide decks.
• 6:57
  All cohesive and you can just export them straight to cloud.
• 7:00
  Code and this matters so much in the bigger picture.
• 7:03
  We are moving way past just generating individual images.
• 7:06
  We're generating cohesive system wide visual identities.
• 7:09
  It's basically a full agency in your browser.
• 7:12
  Right. It radically lowers the barrier
• 7:14
  to entry for entrepreneurship. Anyone can build a polished
• 7:17
  brand now. Absolutely.
• 7:19
  And since we just discussed Claude building a brand, what
• 7:21
  happens when Claude starts managing the brand's finances?
• 7:25
  Yeah, this leads perfectly into Anthropics project.
• 7:28
  Deal. OK, let's unpack this because
• 7:30
  it's like deploying a tiny digital bulldog broker into a
• 7:34
  fleet market on your behalf. That is the perfect analogy.
• 7:37
  So Project Deal was this one week internal experiment at
• 7:40
  Anthropic. 69 employees gave Claude agents $100 budget to buy
• 7:46
  and sell items in a private Slack marketplace.
• 7:48
  Just autonomous trading? Yep.
• 7:50
  The agents conducted quick interviews with their human
• 7:52
  owners to set goals, and then they negotiated on their own.
• 7:56
  In one week, they completed 186 deals worth over $4000.
• 8:01
  The data from that is just so fascinating, especially when you
• 8:04
  look at model capabilities versus human psychology.
• 8:07
  Yeah, what did the numbers show? So agents running on the more
• 8:10
  advanced Opus model secured $3.64 more per item on average
• 8:15
  than the Haiku agents. Wow.
• 8:17
  Just by being better negotiators, yeah.
• 8:20
  There was a specific example where an identical folding bike
• 8:22
  sold for $65 via Opus, but only $38 via Haiku.
• 8:27
  That is a massive difference, but here is the craziest detail.
• 8:31
  Despite Opus getting obviously better prices, the human users
• 8:35
  rated the fairness of the deals almost identically.
• 8:39
  Really. Even with the price gap.
• 8:40
  Yeah, it was 4.06 out of seven for Haiku and 4.05 for Opus.
• 8:47
  The humans literally didn't care about the lost money.
• 8:50
  Which perfectly connects the dots here.
• 8:52
  This proves that an AI commerce, human convenience and friction
• 8:55
  removal might be valued just as highly as extracting maximum
• 8:59
  financial profit. Oh for sure, I would gladly lose
• 9:01
  a few bucks to not have to message people back and forth on
• 9:04
  a marketplace. Exactly.
• 9:05
  That's a convenience fee, though Anthropic did put out a warning
• 9:09
  that legal works for this simply do not exist yet, even though
• 9:13
  46% of users said they would pay real money for the service
• 9:16
  today. I mean if your agent by
• 9:18
  something illegal by accident, who goes to jail?
• 9:20
  You are the AI. Right.
• 9:22
  The law is way behind, and the realization that AI agents are
• 9:25
  now capable of executing legal, design and commerce tasks is
• 9:30
  exactly what is driving today's massive multibillion dollar
• 9:34
  infrastructure scramble. Which is the perfect transition
• 9:37
  into our rapid fire roundup of the day's fastest moving
• 9:40
  developments? Let's do some quick hits.
• 9:42
  First up, Open AI dropped GPT 5.5, establishing a new flagship
• 9:47
  alongside ChatGPT Images 2.0. Massive updates on the
• 9:51
  generative side. Huge and XAI launched Grok Voice
• 9:55
  Thinkfast 1.0 it top speech benchmarks and is actually
• 9:58
  already powering Starlink's phone support That's a.
• 10:00
  Huge real world deployment, but you know, synthesizing all these
• 10:04
  massive financial moves happening simultaneously is
• 10:06
  wild. Yeah, breakdown the big deals
• 10:08
  for us. So Google is investing up to $40
• 10:10
  billion in Anthropic. That's 10 billion now at a $350
• 10:14
  billion valuation, with 30 billion tied to performance
• 10:17
  targets. Unbelievable number.
• 10:19
  And it includes 5 gigawatts of cloud compute.
• 10:21
  Plus Meta just signed a massive deal with AWS for millions of
• 10:24
  Graviton 5 cord chips for agenic AI workloads.
• 10:27
  Millions of chips, they are just hoarding compute right now.
• 10:31
  And it's not just tech companies, it's too political
• 10:33
  too. Oh absolutely.
• 10:34
  The UAE just rolled out a two year plan to deploy a gentic AI
• 10:38
  across 50% of their government services and they are making AI
• 10:42
  training mandatory for federal employees.
• 10:45
  That is aggressive adoption. And Koch here just acquired
• 10:48
  Germany's LF Alpha for $20 billion.
• 10:52
  Which impartially contextualizing that deal, it
• 10:56
  specifically targets governments and companies wary of relying on
• 10:59
  USAI giants for critical infrastructure.
• 11:02
  It perfectly echoes that DeepSeek and Huawei independence
• 11:05
  we discussed at the start. Right, they want their own
• 11:07
  sandbox. Well, after discussing $40
• 11:10
  billion investments in global government mandates, let's bring
• 11:13
  it back down to earth with how everyday listeners are using
• 11:16
  these tools right now in their local communities.
• 11:18
  I love the community workflows. Today's workflow comes from Tim
• 11:21
  M in Lake Luzerne, NY. Tim used Kimi K 2.6 to code a
• 11:26
  Tik to Text app. Oh that is so practical for
• 11:29
  anyone in a wooded area. Yeah, the brilliant simplicity
• 11:32
  of it is what makes it so good. He uses his phone camera to scan
• 11:35
  hard to see places like his back for ticks or those red target
• 11:39
  marks. You just set the phone on the
• 11:40
  bathroom counter and move up and down.
• 11:42
  That's genius, but we really have to emphasize the critical
• 11:46
  privacy design of Tim's app here.
• 11:48
  Oh right, because it's scanning your body.
• 11:50
  Exactly. It runs completely locally.
• 11:52
  It deletes all the images immediately after scanning and
• 11:55
  creates a local PDF detailing the probable percentages of
• 11:59
  areas of concern. It is a perfect synthesis of
• 12:02
  practical health applications and data privacy.
• 12:05
  Absolutely. Well, we have covered so much
• 12:07
  today from global chip wars to tick detection, but I want to
• 12:11
  wrap up our deep dive by directly addressing you, the
• 12:14
  listener with the final thought that builds on what we've talked
• 12:17
  about with agentic commerce and automation.
• 12:19
  It's a big. Philosophical question.
• 12:20
  It is if AI agents soon handle all of our negotiations, our
• 12:24
  initial outreach, and all that daily haggling, do we as a
• 12:27
  society slowly lose the human art of persuasion?
• 12:31
  Or do we simply free ourselves to focus on the things
• 12:33
  algorithms can never do? It really makes you wonder what
• 12:36
  human interaction will look like in 10 years.
• 12:38
  Definitely something for you to ponder.
• 12:41
  Well, Please remember to subscribe to stay updated on all
• 12:43
  these fast moving changes. And if you found this deep dive
• 12:46
  valuable, we'd love it if you could rate the show 5 stars.
• 12:49
  Thank you so much for listening, and we'll see you back for
• 12:52
  tomorrow's episode.