Latest / Connecting the Dots / Anthropic Drops Three New Models and Ups the AI Stakes

Transcript

Anthropic Drops Three New Models and Ups the AI Stakes · Connecting the Dots

• 0:00
  Today's tech news moves fast, but understanding
• 0:03
  it shouldn't be hard. That's why Snarful Solutions
• 0:06
  Group brings you Connecting the Dots, where we
• 0:09
  break down what's happening in tech and why it
• 0:11
  matters to you. Now, let's dive in with your
• 0:14
  hosts. Welcome to Connecting the Dots. I'm Alex.
• 0:17
  And I'm Morgan. Good morning, everyone. It is
• 0:20
  Tuesday, November 25th, 2025. And we've got a
• 0:24
  lot to talk about today. It looks like Anthropic
• 0:26
  is making some pretty big waves. They certainly
• 0:29
  are but before we get to that just a quick look
• 0:31
  at our local situation here in Sacramento If
• 0:33
  you're heading out this morning that dense fog
• 0:35
  advisory is still in effect until 11 zero zero
• 0:39
  a .m. So, you know take it easy on the roads
• 0:42
  Yeah, please do and for anyone traveling for
• 0:44
  Thanksgiving. It looks like a bit of a messy
• 0:47
  day nationally There's heavy snow and some pretty
• 0:49
  high winds up in the northern plains and the
• 0:52
  upper Midwest and some thunderstorms down in
• 0:54
  the southeast too. So just A travel day to be
• 0:57
  extra careful. Absolutely. Safe travels to everyone
• 0:59
  out there. Speaking of movement, let's take a
• 1:01
  quick peek at the markets. Looks like a bit of
• 1:03
  a mixed bag to start the day. It is. The Dow
• 1:06
  Jones and the S &P 500 both kicked off in the
• 1:09
  green. The Dow is up about 283 points. That's
• 1:12
  a nice little gain of 0 .61%. And the S &P 500
• 1:15
  is up slightly as well, about 0 .20%. But crypto
• 1:19
  seems to be taking a bit of a breather. It does.
• 1:21
  Bitcoin has pulled back a little. It's down about
• 1:23
  one point three two percent. That's what around
• 1:25
  eleven one hundred sixty two towers. Just a good
• 1:27
  reminder that volatility is always part of the
• 1:30
  game in that space. Always. Now, before we jump
• 1:33
  into our main analysis, we want to give a quick
• 1:35
  shout out to our sponsor, Old Glory. If you're
• 1:39
  looking for iconic music in sports merch. they
• 1:41
  are the place to go and you know with the holidays
• 1:44
  coming up it's perfect for that techie friend
• 1:46
  who says they don't need anything right they
• 1:48
  have some really unique stuff and we've got a
• 1:50
  great deal for you just use the promo code SNARFL
• 1:53
  at checkout and you'll get 15 % off your entire
• 1:56
  purchase you can find them at OldGlory .com definitely
• 1:59
  go check them out all right let's get into the
• 2:01
  core of it The Morningstar report today is, well,
• 2:04
  it's all about Anthropic. It feels like they
• 2:07
  just unleashed a major offensive. That's a great
• 2:09
  way to put it, an offensive. They've just dropped
• 2:11
  Claude Opus 4 .5, and it's a clear shot across
• 2:14
  the ballot OpenAI and Google. And it seems like
• 2:16
  it's all about a three -pronged strategy to really
• 2:20
  capture the enterprise market. Exactly. I mean,
• 2:23
  the pace of these releases is just relentless.
• 2:25
  We just got GBT 5 .1, we just got Gemini 3, and
• 2:28
  now Anthropic is back saying, nope, we have the
• 2:30
  best model in the world. So let's start with
• 2:32
  that first story, the AI price war. Because this
• 2:35
  isn't a minor adjustment, is it? This is huge.
• 2:38
  Oh, it's a bombshell. Yeah. The headline here
• 2:40
  is a massive 67 % price cut. 67%. So what were
• 2:46
  we looking at before? Well, with the last Opus
• 2:48
  model, It costs you $15 for a million input tokens.
• 2:52
  Okay. And get this $75 for a million output tokens.
• 2:57
  Wow. Okay. So that's, that's boutique pricing
• 3:00
  or only using that for very specific, very high
• 3:03
  value tasks. Precisely. And now those prices
• 3:06
  have absolutely plummeted. We're talking $5 for
• 3:09
  input and 25 for output. So for anyone listening
• 3:11
  who manages a corporate budget, what does that
• 3:14
  actually mean? Cutting your costs by two thirds.
• 3:16
  It means AI is moving from being this interesting
• 3:18
  experiment to a genuine production -ready utility.
• 3:22
  That cost of inference just running the model,
• 3:25
  that's what adds up. And when you can cut that
• 3:27
  cost so dramatically, you can suddenly afford
• 3:30
  to use it for everything. For customer service,
• 3:33
  for internal coding, for compliance, you name
• 3:36
  it. It kind of reminds me of the cloud storage
• 3:38
  wars. you know a decade ago where the price per
• 3:42
  gigabyte just kept falling and falling that's
• 3:45
  the perfect analogy but what's really interesting
• 3:47
  is that even with this massive cut opus 4 .5
• 3:51
  is still positioning itself as the premium option
• 3:54
  yeah that's what i found fascinating it's still
• 3:56
  more expensive than its rivals oh yeah By a good
• 3:58
  margin, GPT 5 .1 is way down at like $1 .25 and
• 4:03
  $10. Gemini 3 Pro is around $2 and $12. So Anthropic
• 4:07
  is basically saying, we're more affordable for
• 4:09
  production now, but you still have to pay up
• 4:13
  for our quality. Exactly. It's a huge vote of
• 4:15
  confidence in their own model. But the problem
• 4:17
  is that justifying that quality is getting harder
• 4:20
  and harder Because the benchmarks are only showing
• 4:22
  tiny improvements a few percentage points here
• 4:24
  and there, right? So if you're a developer, you're
• 4:26
  not gonna pay three times more for a two percent
• 4:29
  better score on a test The real difference is
• 4:32
  shifting to the application features. Ah Okay,
• 4:34
  so things like The ability to preserve thinking
• 4:39
  blocks. What exactly is that? Think of it like
• 4:41
  a persistent memory for a project before the
• 4:45
  model would kind of start fresh with every single
• 4:47
  request. Which is really inefficient for a long
• 4:50
  complex task. Totally. Now, by keeping those
• 4:53
  thinking blocks, it remembers the plan, the context,
• 4:57
  the reasoning from previous steps. It makes complex
• 5:00
  debugging or data analysis so much more reliable.
• 5:03
  OK, that makes sense. And then there's this Zoom
• 5:05
  tool for computer use. How does that change things?
• 5:08
  It's all about efficiency. instead of the AI
• 5:11
  having to look at the entire computer screen
• 5:12
  at once, which is computationally expensive.
• 5:15
  And can be confusing for it. Right. It can now
• 5:17
  zoom in on just one specific button or one data
• 5:20
  field, just like a person would. It's faster,
• 5:23
  cheaper, and way more accurate for automation.
• 5:25
  So the big takeaway from Story 1 is that the
• 5:27
  battleground is shifting from raw intelligence
• 5:30
  to developer -friendly features and efficiency.
• 5:33
  That's it. Which, of course, leads us right into
• 5:36
  Story 2. Because when you give an AI all this
• 5:39
  power, you have to talk about the security, the
• 5:42
  elephant in the room, the big one, prompt injection.
• 5:45
  So Anthropic is claiming this new model is the
• 5:47
  most secure ever, the hardest to trick. But the
• 5:50
  fine print tells a different story. It does.
• 5:53
  They were very clear. They said it's harder to
• 5:56
  trick than any other frontier model, which is
• 5:59
  a strong claim. Yeah. But they also released
• 6:01
  the data, which is. The reality check, we all
• 6:03
  need it. So prompt injection. This is basically
• 6:06
  when an attacker tries to trick the AI with hidden
• 6:09
  instructions to get it to do something malicious.
• 6:12
  Right, exactly. It's like social engineering
• 6:13
  for an AI. And the numbers, they're a little
• 6:16
  concerning. Even with this safest model, a single
• 6:19
  attempt to trick it still works one out of 20
• 6:22
  times. OK, that's not great. But what if the
• 6:25
  attacker is persistent? Well, that's where it
• 6:27
  gets really scary. If an attacker tries, say,
• 6:29
  10 different ways to trick it. The success rate
• 6:32
  shoots up to one out of three. One in three.
• 6:34
  So a 33 % chance of success for a motivated attacker.
• 6:39
  If you're running this on your company's critical
• 6:41
  systems, that's an unacceptable risk. It's a
• 6:45
  huge risk. And it gets worse when you look at
• 6:47
  what they were asking the model to do. Right.
• 6:49
  Let's look at the tests. So for generating malicious
• 6:51
  code things like malware or DDoS attack scripts,
• 6:55
  it only refused about 78 % of the time. meaning
• 6:58
  one out of every four times, it might actually
• 7:00
  help you build a cyber weapon. And for general
• 7:03
  computer use, like surveillance or spreading
• 7:05
  misinformation, it was a little better, but still
• 7:08
  only refused about 88 % of requests. This is
• 7:11
  why you can't just rely on the model's built
• 7:12
  -in safety features. It's not enough. They gave
• 7:15
  some pretty chilling real -world examples, too.
• 7:17
  They did, like tricking the model into scraping
• 7:20
  forums to find people with gambling addictions.
• 7:22
  and then compiling their usernames for a predatory
• 7:26
  marketing campaign. That's just awful. It's incredibly
• 7:28
  unethical. Or the other one, drafting a blackmail
• 7:31
  email demanding Bitcoin. These aren't theoretical
• 7:34
  risks. These are things that can happen. So the
• 7:36
  connecting thought here for any business listening
• 7:38
  is that you have to design your systems assuming
• 7:40
  an attacker will get through. Exactly. The model
• 7:43
  is just one layer. You need your own security,
• 7:45
  your own verification, your own access controls
• 7:47
  built around the AI. You have to plan for that
• 7:50
  one in three failure. Which brings us perfectly
• 7:52
  to our third and final story. The efficiency
• 7:55
  engine. If Opus is the big powerful and slightly
• 8:00
  risky model. Then you need a smaller, faster,
• 8:03
  cheaper model to do the grunt work. Yeah. And
• 8:05
  that's where Heiko 4 .5 comes in. The rise of
• 8:08
  the small language model, the SLM. And this one
• 8:09
  is so efficient, they're just giving it away
• 8:12
  for free on their main site. And the performance
• 8:14
  is incredible. It's got the same coding ability
• 8:16
  as Sonnet 4, which is a much bigger model from
• 8:19
  just four months ago. But it's twice as fast.
• 8:21
  Twice as fast and one third the cost. We're talking
• 8:23
  one dollar for input, five for output. That's
• 8:26
  a game changer. And it enables this whole new
• 8:29
  strategy that they're calling the team of agents.
• 8:31
  Right. You don't use your most expensive, brilliant
• 8:34
  tool for every single little job. Yeah. You wouldn't
• 8:36
  hire a brain surgeon to put on a band -aid. That's
• 8:39
  a great way to put it. So you use the big model
• 8:41
  like Opus or Sonnet as the leader. It does the
• 8:44
  high level thinking. It makes the strategic plan.
• 8:47
  It does the complex analysis. The expensive stuff.
• 8:50
  And then Haiku 4 .5 comes in as the worker. It's
• 8:54
  the sub -agent. It handles the millions of tiny
• 8:57
  repetitive tasks, monitoring data streams, gathering
• 9:01
  research, answering simple chatbot questions.
• 9:04
  all the high -volume work. And this has a massive
• 9:07
  impact on cost and something we don't talk about
• 9:09
  enough, energy consumption. A huge impact. The
• 9:12
  numbers are staggering. A single query to a giant
• 9:15
  model can use something like 6 ,700 joules of
• 9:18
  energy. Wow. 6 ,700 joules. Contrast that with
• 9:22
  a small efficient model like Haiku. It might
• 9:24
  only use 114 joules. That's almost 60 times less
• 9:27
  energy. So if you can shift most of your workload
• 9:29
  from the big model to the small one, you're saving
• 9:32
  a fortune on your electricity bill and your server
• 9:34
  costs. It's the only way this whole AI ecosystem
• 9:37
  becomes sustainable in the long run, both financially
• 9:40
  and environmentally. This team of agents approach
• 9:42
  is, without a doubt, the future. Before we wrap
• 9:45
  things up, we want to thank our other wonderful
• 9:47
  sponsor, Scony. Scony is a premium golf shoe
• 9:50
  store. If you're a golfer, you know how important
• 9:53
  good shoes are for comfort and performance on
• 9:55
  the course. They have incredible stuff. They
• 9:57
  really do. And like with our other sponsor, you
• 9:59
  can get a great deal. Just use the promo code
• 10:01
  SNARFL at checkout for 15 % off your entire purchase.
• 10:05
  You can find them at scony .com. OK, so let's
• 10:08
  quickly recap the three big takeaways from Anthropix
• 10:11
  News today. All right. First, Opus 4 .5 and its
• 10:15
  huge price drop is turning high end AI into more
• 10:18
  of a commodity. The price war is on and large
• 10:22
  scale enterprise adoption is the end goal. Second,
• 10:25
  security is still a major concern. Even the best
• 10:28
  models can be tricked, so you absolutely must
• 10:30
  build your own security architecture around them.
• 10:32
  Don't rely on the model alone. And third, the
• 10:35
  future is a team of agents. Using small efficient
• 10:39
  models like Haiku for the bulk of the work is
• 10:41
  the only way to make AI cost -effective and sustainable
• 10:44
  at scale. It really shows the whole industry
• 10:46
  is shifting from just chasing benchmark scores
• 10:48
  to building practical, secure, and efficient
• 10:50
  solutions. The market is definitely maturing.
• 10:53
  It's all about utility now. We're here to help
• 10:55
  at Snarful, so reach out on our website if you
• 10:57
  have any more questions. That's a wrap on today's
• 11:03
  top stories. Remember, tech doesn't have to be
• 11:06
  complicated and we're here to help connect the
• 11:08
  dots. If you've got questions or want to learn
• 11:11
  more, visit us at snarful .com. Thanks for listening
• 11:15
  and we'll catch you next time.