Latest / AI Fire Daily / #99 Robin: Fable 5.1’s Price Trap, Gemini 3.8’s Coding Upset, and the Dark Side of OpenAI Astra

Transcript

#99 Robin: Fable 5.1’s Price Trap, Gemini 3.8’s Coding Upset, and the Dark Side of OpenAI Astra · AI Fire Daily

• 0:00
  I often look at this endless stream of AI releases.
• 0:04
  It feels like an exhausting cycle of pure hype.
• 0:07
  Oh, yeah. It gets overwhelming really fast. Right.
• 0:10
  Every single week, you know, another lab drops
• 0:14
  a foundational model, and there's a brand new
• 0:16
  set of benchmark charts, plus this loud claim
• 0:19
  that your workflow is suddenly obsolete. Right.
• 0:21
  But when you actually sit down to write, nothing
• 0:24
  has changed. Yeah. For most daily tasks, the
• 0:27
  needle hasn't really moved at all. No, it hasn't.
• 0:30
  But there is one area experiencing a massive,
• 0:33
  quiet revolution. Absolutely. The real paradigm
• 0:35
  shift is happening right now in coding. So welcome
• 0:38
  to this deep dive into our latest research stack.
• 0:40
  Yeah, we are cutting through all that algorithmic
• 0:43
  noise today. We really want to evaluate the actual
• 0:46
  real -world utility of three specific models
• 0:49
  for you. We're going to look at Anthropic's expensive
• 0:51
  powerhouse, Fable 5 .1. Right. We'll examine
• 0:55
  Google's underhyped workhorse, Gemini 3 .8 Flash.
• 0:58
  Yes. And we'll explore the delayed high -risk
• 1:01
  wildcard, which is OpenAI Astra. The mission
• 1:04
  here is actually incredibly straightforward.
• 1:07
  We need to figure out how you can choose a model,
• 1:09
  right? Yeah. And you need to do it without getting
• 1:11
  trapped by leaderboard hype. So let's start with
• 1:14
  the loudest release of the week. That would be
• 1:17
  Fable 5 .1 from Anthropic. It definitely made
• 1:21
  waves across the developer community. We need
• 1:23
  to see if it actually justifies all this noise.
• 1:26
  Fable 5 .1 is mostly restricted to paid anthropic
• 1:29
  plants, right? Yeah, that's right. They also
• 1:31
  dropped a parallel model called Mythos 5 .1,
• 1:34
  but that one is heavily restricted to cyber defense
• 1:37
  organizations. So most everyday developers will
• 1:39
  never even touch Mythos. Exactly. Fable 5 .1
• 1:43
  is the real story for the public. But it isn't
• 1:45
  actually a completely brand new model from scratch.
• 1:49
  No, it's not a totally new base architecture.
• 1:52
  It's basically a highly refined version of Fable
• 1:55
  5. Right. They just added some intense post training
• 1:58
  and reinforcement learning layers. That feels
• 2:01
  like a crucial distinction to make here. It's
• 2:03
  a calculated refinement. Not a massive architectural
• 2:06
  revolution, but Anthropic did share some seriously
• 2:10
  impressive benchmark charts this week. The charts
• 2:13
  look incredible at a first glance. Their scores
• 2:16
  for agentic scientific research roughly doubled
• 2:19
  overnight. Wow. We should probably define what
• 2:23
  that actually means in practice. Right. So it
• 2:25
  means letting AI autonomously plan and run complex
• 2:29
  laboratory experiments. Which is a huge leap
• 2:32
  for highly specialized independent tasks. Oh,
• 2:34
  massive. And their agentic coding scores also
• 2:37
  saw a very clear upgrade. But I noticed their
• 2:39
  scores for general knowledge work. basically
• 2:42
  flat -lined. Yeah, that chart definitely feels
• 2:45
  very carefully selected. They want to show the
• 2:47
  biggest possible gap in performance. The gain
• 2:50
  encoding ability is very real and measurable.
• 2:53
  Right. But for everyday knowledge work, the jump
• 2:56
  is surprisingly small. Which really separates
• 2:58
  a great benchmark model from a daily driver.
• 3:01
  Absolutely. If you just want to draft an email,
• 3:03
  this feels like overkill. Totally. But they also
• 3:05
  made some fascinating structural safety changes,
• 3:07
  didn't they? They did. Fable 5 .1 is far less
• 3:10
  restrictive for general cybersecurity work. The
• 3:13
  internal safety safeguards are much more precise
• 3:16
  now. Meaning it won't just blindly refuse to
• 3:19
  help you analyze code? Exactly. There are about
• 3:21
  60 % fewer false positive safety interventions.
• 3:25
  Wow. Yeah. It helps developers identify system
• 3:28
  vulnerabilities much easier now, but dual -use
• 3:32
  cyber tasks still get pushed down to older Opus
• 3:34
  models. Which brings us to a deeply controversial
• 3:37
  shift in their strategy. Oh boy, yeah. They introduced
• 3:40
  a strict anti -distillation move this week. Anthropic
• 3:44
  actually blocked new API accounts from keeping
• 3:47
  prior thinking transcripts. This is a massive
• 3:50
  philosophical shift in the industry. It really
• 3:52
  is. We should define what distillation actually
• 3:54
  means in this context. Go for it. Distillation
• 3:56
  is using a stronger model to train a newer or
• 3:59
  smaller one. Perfect. You basically take the
• 4:01
  massive teacher model's outputs, right? And then
• 4:04
  you train a smaller model to perfectly copy that
• 4:06
  behavior. So Anthropic is aggressively blocking
• 4:09
  access to those prior thinking transcripts. They
• 4:12
  really want to protect their incredibly expensive
• 4:15
  technological moat. actively making it harder
• 4:18
  for other labs to copy Fable 5 .1. Right. And
• 4:22
  you can certainly understand why Frontier Labs
• 4:24
  care about this. If an open source lab copies
• 4:27
  your model, your billion dollar lead vanishes
• 4:29
  fast. Yeah. But there is a deeply uncomfortable
• 4:33
  irony happening right here. Oh, absolutely. Frontier
• 4:36
  Labs trained all these models on scraped public
• 4:38
  data. They essentially vacuumed up the entire
• 4:40
  open Internet for free. Yeah. But now they actively
• 4:44
  block others from analyzing their own models.
• 4:47
  It is a fascinating and glaring contradiction.
• 4:50
  They can gladly use public data to build their
• 4:53
  massive moat, but now they're aggressively pulling
• 4:56
  up the drawbridge behind them. Is this anti -distillation
• 4:59
  move actually going to protect their technological
• 5:01
  moat or just create friction that annoys developers?
• 5:04
  Honestly. It's mostly just going to frustrate
• 5:07
  the developer community. Dedicated competitor
• 5:09
  labs will always find complex technical workarounds,
• 5:12
  but regular users will just hit these artificial
• 5:15
  and annoying roadblocks. So protecting their
• 5:18
  data creates friction that actually drives everyday
• 5:21
  developers away. A beat. That is exactly what's
• 5:24
  happening right now across developer forums.
• 5:27
  And the friction goes way beyond just losing
• 5:29
  those transcripts. The sheer economics are a
• 5:32
  real problem for Fable 5 .1. Right. Let's talk
• 5:35
  about the cost. Anthropic claims it's 25 % cheaper
• 5:39
  for typical coding workloads. Yes, they do. They
• 5:42
  say enhanced context cache reads drive the total
• 5:45
  cost down. Let's explain how context caching
• 5:48
  is actually supposed to work. You upload a massive
• 5:51
  dense code base to the model once. The AI keeps
• 5:56
  that specific data sitting in active memory.
• 5:59
  So subsequent questions about that code base
• 6:00
  shouldn't cost as much computationally. Right,
• 6:03
  because the model isn't re -reading the entire
• 6:05
  massive file every single time. They say that
• 6:07
  drastically lowers the cost for heavy users.
• 6:09
  But, you know, a discount on something outrageously
• 6:11
  expensive is still expensive. Exactly. The base
• 6:13
  sticker price didn't actually change at all.
• 6:15
  The input price is still $10 per million tokens,
• 6:19
  and the output price remains an incredible $50
• 6:22
  per million tokens. That is incredibly steep
• 6:26
  for any kind of daily development work. And the
• 6:29
  benchmark cost per task tells a very revealing
• 6:32
  story. Which completely contradicts their claims
• 6:35
  about typical workload savings. It really does.
• 6:37
  The average benchmark cost sits at $3 .69 per
• 6:41
  task. Wow. That is literally the highest cost
• 6:44
  in its entire peer group. It's sitting entirely
• 6:46
  ahead of Opus 5 in raw cost. Yeah. That creates
• 6:50
  a massive barrier for practical everyday use.
• 6:53
  Yeah. And the independent hands -on tests proved
• 6:55
  this point perfectly this week. Oh, they definitely
• 6:57
  did. The pure output quality in the tests was
• 7:00
  undeniably excellent. The model generated a structurally
• 7:04
  brilliant SVG image from scratch, but generating
• 7:07
  that single vector image cost $4 .35. And it
• 7:11
  took 18 painful minutes to fully generate. What
• 7:14
  on earth was MATL doing for 18 minutes? It was
• 7:17
  obsessively refining the mathematical coordinates
• 7:19
  of the vector paths. It's incredibly thorough,
• 7:22
  but that speed is entirely unusable for rapid
• 7:25
  prototyping. Totally. It also built an incredible
• 7:29
  one -shot game clone. We should quickly define
• 7:32
  what a one -shot game clone is. It's building
• 7:34
  an entire playable video game from a single prompt.
• 7:37
  Exactly. It built a complex browser game called
• 7:40
  Megabonk. Right. The custom characters, the interactive
• 7:43
  of settings, the physics controls, they were
• 7:45
  genuinely fantastic. But then the final API bill
• 7:48
  landed on that heavy coding test. It was undeniably
• 7:51
  painful to look at. Running that single test
• 7:54
  on the ultra code integration environment was
• 7:56
  disastrous. It completely burned through a $200
• 7:59
  monthly subscription plan. And then it automatically
• 8:02
  hit $114 in extra credit. Yes. It ran its autonomous
• 8:07
  coding loops for over 90 minutes. Wow. And after
• 8:11
  all that time and money, it still did. finish
• 8:13
  the task. It's an absolute powerhouse of reasoning
• 8:16
  capability, but it's just entirely too expensive
• 8:19
  for normal integration. Way too expensive. Using
• 8:22
  Fable 5 .1 for daily coding feels like using
• 8:25
  a Formula One car for a grocery run. It's an
• 8:27
  incredible machine, but completely impractical.
• 8:30
  That is the perfect analogy for this specific
• 8:33
  model. You really don't need a Formula One car
• 8:36
  to get milk. No. You need a highly reliable,
• 8:39
  incredibly efficient daily vehicle. If Fable
• 8:42
  5 .1 is an unaffordable luxury, what is the practical
• 8:45
  alternative? We need to pivot from luxury cost
• 8:49
  to extreme efficiency. We definitely need to
• 8:52
  talk about Google right now. Right. They've been
• 8:54
  missing from the serious AI coding conversation
• 8:56
  for a while. Cloud Code and OpenAI have completely
• 8:59
  dominated that specific developer space. They
• 9:02
  absolutely have. But Google's new Gemini 3 .8
• 9:04
  Flash changes that conversation completely. It's
• 9:07
  the quiet, underhyped workhorse we've actually
• 9:10
  been waiting for. The deep SWE score is the massive
• 9:13
  signal here. We should clarify what deep SWE
• 9:16
  actually measures for developers. Right. It's
• 9:18
  a benchmark testing AI on real software engineering
• 9:21
  GitHub issues. It literally strips away all the
• 9:25
  abstract theoretical logic puzzles. It forces
• 9:28
  the model to solve actual bugs in real code basis.
• 9:31
  And Gemini 3 .8 Flash scored an impressive 73
• 9:34
  .7 on it. That officially rounds up to a 74 on
• 9:37
  the public leaderboard. This is a fundamentally
• 9:40
  critical metric to pay attention to. It perfectly
• 9:43
  matches the highly capable Claude Opus 5 model.
• 9:46
  Wow. And it actually beats Anthropix Fable 5,
• 9:48
  which only scored around 70. Which means it handles
• 9:51
  real -world software engineering incredibly well.
• 9:54
  And the sheer cost efficiency of this model is
• 9:57
  just staggering. The average cost per task is
• 10:00
  roughly $2 .36. But artificial analysis shows
• 10:04
  optimized runs can drop as low as $0 .58. Compare
• 10:07
  that to directly to Claude Opus 5. That model
• 10:10
  averages around $11 .84 per task. Right. And
• 10:14
  Fable 5 was sitting over $21 per task. The generation
• 10:17
  speed is also a massive operational advantage
• 10:20
  here. Gemini 3 .8 Flash averages a lightning
• 10:23
  fast two and a half minutes per task. Wow. Opus
• 10:25
  5 and Fable both take roughly 7 .4 minutes on
• 10:29
  average. It's essentially matching top tier coding
• 10:31
  performance at a fraction of the cost. Exactly.
• 10:33
  The independent hands -on tests completely back
• 10:36
  this up too. For that same SVG image test, it
• 10:39
  was incredibly efficient. The API call cost just
• 10:43
  nine cents to run. Nine cents. And it took only
• 10:46
  one minute and 32 seconds to finish. Sure. It
• 10:48
  wasn't mathematically perfect like the fable
• 10:50
  generation. Some of the visual vector details
• 10:53
  were a little strange. Yeah. But you simply cannot
• 10:55
  ignore that aggressive price and speed. The Megabont
• 10:59
  game clone was arguably even more impressive.
• 11:02
  It actually outperformed the first complex attempt
• 11:05
  from Fable 5. It generated surprisingly smart
• 11:08
  enemies with decent pathfinding logic. Yes. The
• 11:12
  core character interaction work was remarkably
• 11:14
  solid. The dynamic camera control actually followed
• 11:17
  the player perfectly. It truly felt like a robust
• 11:20
  interactive game prototype. Yeah. It wasn't just
• 11:23
  a fragile rough toy demo script. Yeah. But we
• 11:26
  do need to talk about its glaring weaknesses.
• 11:29
  Because it certainly isn't a perfect foundational
• 11:31
  model. It ranks around 10th overall on the artificial
• 11:34
  analysis leaderboard. Right. It sits at a generalized
• 11:37
  knowledge score of about 59 there. It struggles
• 11:40
  heavily with something called humanity's last
• 11:43
  exam. That's a test of extremely difficult questions
• 11:46
  that stump human experts. Right. It completely
• 11:49
  fails to reason through those abstract academic
• 11:52
  problems. Yeah. It also fails terminal bench
• 11:54
  2 .1. pretty significantly. Which evaluates an
• 11:57
  AI's ability to navigate complex command line
• 12:00
  systems. Exactly. It struggles to chain terminal
• 12:03
  commands together autonomously. Okay. So it really
• 12:06
  is mostly a highly optimized coding specialist.
• 12:09
  It isn't the smartest general model in the world
• 12:11
  for broad tasks. And it does have one other frustrating
• 12:14
  operational weak spot. It burns a massive amount
• 12:17
  of output tokens on extended reasoning. Oh, it
• 12:21
  definitely likes to think out loud in the output
• 12:23
  window. I still wrestle with DropDrift myself.
• 12:25
  Yeah. You asked for a simple button script. The
• 12:28
  model hallucinates a database error and tries
• 12:30
  to fix it. Oh, yes. Suddenly, it's designing
• 12:33
  a brand new server architecture you didn't ask
• 12:35
  for. You get lost in these long loops and waste
• 12:37
  so many tokens. It happens to absolutely all
• 12:40
  of us. Gemini 3 .8 flash can definitely spiral
• 12:43
  into those rambling internal monologues It generates
• 12:46
  thousands of tokens before finally arriving at
• 12:49
  the answer if Gemini uses so many more output
• 12:52
  tokens for reasoning does the cheap price actually
• 12:55
  hold up during heavy real -world coding? It genuinely
• 12:58
  does hold up remarkably well because the base
• 13:01
  cost per output token is just so aggressively
• 13:04
  low. You can burn through thousands of rambling
• 13:07
  tokens and still save money. Yeah, higher token
• 13:10
  volume doesn't matter if the final bill stays
• 13:12
  incredibly low. Two -sec silence. It completely
• 13:15
  balances out in the end. Insert mid -roll sponsor
• 13:18
  RID here, provided separately. So moving on from
• 13:21
  the daily drivers. Right. We've looked deeply
• 13:23
  at the practical everyday models. Now we really
• 13:26
  need to move our focus toward the horizon. Yes.
• 13:29
  We have to look at the implications of OpenAI
• 13:31
  Astra. This is where the landscape gets genuinely
• 13:33
  intense and slightly alarming. What actually
• 13:36
  happens when extreme efficiency meets autonomous
• 13:39
  power? OpenAI didn't actually release the Astra
• 13:42
  model this week. They just published a technical
• 13:45
  path to Astra research update. Right. They absolutely
• 13:48
  have to delay the public release. And the core
• 13:51
  reason they delayed it is profound safety concerns.
• 13:54
  The model can actively find unknown zero -day
• 13:57
  security flaws. It can autonomously build complex
• 14:00
  exploits without any human guidance. That is
• 14:03
  a terrifying and fascinating leap in system capability.
• 14:06
  It really is. We're not just talking about a
• 14:08
  helpful coding assistant anymore. We are talking
• 14:10
  about an independent autonomous digital agent.
• 14:14
  Let's look closely at the massive capability
• 14:16
  jump here. The GPT -5 .6 sole model. had an exploit
• 14:20
  success rate of around 11 .5%. Yeah. And it used
• 14:24
  roughly 140 ,000 tokens to achieve that. Astra
• 14:27
  immediately jumps to a 17 .5 % exploit success
• 14:30
  rate. Wow. But it remarkably only uses 18 ,500
• 14:34
  tokens to do it. It's achieving higher success
• 14:36
  while reasoning far more densely. And at a higher
• 14:39
  computational run, it gets even crazier. Oh,
• 14:41
  yeah. At a higher run, Astra hits nearly a 40
• 14:44
  % exploit success rate. And it achieves that
• 14:47
  using roughly 76 ,000 tokens. Whoa. Imagine an
• 14:50
  autonomous agent finding unknown security exploits
• 14:53
  at a 40 % success rate. It is genuinely wild
• 14:57
  to think about. Yeah. It is wildly more capable
• 14:59
  at navigating complex networked systems. And
• 15:03
  it's vastly more token efficient at the exact
• 15:05
  same time. This means it can generate devastatingly
• 15:08
  effective results while doing far less computational
• 15:11
  work. Exactly. For internal safety teams, this
• 15:14
  creates a complete nightmare. scenario. Right.
• 15:17
  It means the model needs incredibly tight control
• 15:20
  before a public release. Which brings us to the
• 15:22
  biggest technical concern with the Astra architecture.
• 15:25
  Yes. It relies heavily on a mechanism called
• 15:27
  recurrent depth. Also known in the research as
• 15:30
  a looped transformer architecture. We really
• 15:33
  need to explain what this does inside the neural
• 15:35
  network. Let's do it. Recurrent depth is letting
• 15:38
  an AI process the same text repeatedly to improve
• 15:41
  answers. Exactly. A standard model predicts the
• 15:44
  next word in a straight line, but this model
• 15:47
  loops back over its own latent mathematical states.
• 15:50
  It repeatedly processes the information before
• 15:53
  it ever outputs a human -readable word. In theory,
• 15:56
  this helps the model reason through incredibly
• 15:58
  complex problems much better. It's like stacking
• 16:02
  Lego blocks of data over and over again. But
• 16:04
  there is a massive, deeply problematic trade
• 16:07
  -off happening here. This looped process completely
• 16:11
  hides the model's internal chain of thought.
• 16:13
  It effectively obscures the underlying logic
• 16:16
  from the researchers. This means human safety
• 16:19
  teams cannot read the model's reasoning process
• 16:22
  anymore. The actual reasoning becomes a completely
• 16:25
  opaque black box. You cannot simply inspect the
• 16:29
  text to see how it reached a specific answer.
• 16:31
  Which means you can't spot exactly where the
• 16:34
  internal logic went fundamentally wrong. Right.
• 16:36
  and you can't smoothly redirect it before it
• 16:38
  outputs something dangerous. For simple everyday
• 16:41
  writing tasks, that lack of transparency doesn't
• 16:43
  really matter. Sure. But for autonomous cybersecurity
• 16:46
  or synthetic biology tasks, it's incredibly risky.
• 16:50
  Unreadable reasoning makes the model almost impossible
• 16:53
  to steer reliably. It becomes incredibly hard
• 16:56
  to monitor for dangerous, deceptive behavior.
• 16:59
  OpenAI used to actively champion readable model
• 17:02
  thinking for safety alignment. Yeah, they do.
• 17:04
  They wanted us to see exactly how the AI was
• 17:06
  making decisions. OpenAI used to champion readable
• 17:09
  model thinking for safety. If we can't read this
• 17:13
  new model's reasoning, how can we possibly trust
• 17:16
  its code? We really can't trust it inherently
• 17:19
  anymore. We have to rely entirely on rigidly
• 17:22
  testing the final output. We can no longer guide
• 17:26
  the step -by -step thought process itself. Right.
• 17:28
  Hidden logic means we lose the ability to catch
• 17:31
  dangerous mistakes early. Beat. This creates
• 17:35
  a massive unresolved tension across the entire
• 17:38
  AI industry. It really does. We are willingly
• 17:40
  trading vital transparency for raw unbridled
• 17:44
  performance. It makes Astra the ultimate high
• 17:47
  -risk wild card in this entire stack. Oh, absolutely.
• 17:49
  It might genuinely be the most important model
• 17:51
  in this group. But it's also clearly the one
• 17:53
  we must be the most careful with. We desperately
• 17:56
  need to see the real product operating in the
• 17:58
  wild first. We need to see the real operational
• 18:00
  limits and the actual pricing structure. Yeah.
• 18:03
  Until then, it's just an incredibly powerful,
• 18:06
  slightly terrifying promise. So let's synthesize
• 18:08
  this overarching theme for a moment. The massive
• 18:11
  AI industry desperately wants you to chase the
• 18:13
  loudest launch. They actively want you to stare
• 18:16
  obsessively at the highest benchmark scores.
• 18:19
  Right. They want you to constantly fear missing
• 18:22
  out on the next big revolution. But generating
• 18:24
  real productivity is an entirely different game.
• 18:26
  It is. Especially when we are talking about daily
• 18:29
  software development. Real productivity is just
• 18:32
  about finding the perfect operational fit. Fable
• 18:35
  5 .1 is the brilliant but completely unaffordable
• 18:39
  genius. It really is a Formula One car meant
• 18:42
  for a simple grocery run. Right. Astra represents
• 18:46
  the incredibly powerful but potentially dangerous
• 18:48
  autonomous future. Yeah. It is highly capable,
• 18:52
  extremely efficient, and currently locked away
• 18:54
  for global safety. And Gemini 3 .8 Flash is the
• 18:58
  quiet, highly efficient winner today. Exactly.
• 19:01
  It's the practical model that actually helps
• 19:02
  you finish your real work. And it manages to
• 19:05
  do that without completely draining your cloud
• 19:07
  budget. It effortlessly ignores the industry
• 19:09
  hype and just delivers solid code. Thank you
• 19:12
  so much for joining us on this deep dive. Take
• 19:14
  a quiet moment to reflect on your own actual
• 19:16
  workflows. Yeah. Look closely at your real daily
• 19:19
  needs before you rush to upgrade. Check our show
• 19:22
  notes for the direct links to these specific
• 19:24
  benchmarks. It's always a better strategy to
• 19:27
  optimize for your actual daily tasks. Don't just
• 19:30
  blindly optimize for the top of the leaderboard.
• 19:32
  As frontier models like Astra aggressively move
• 19:35
  toward hidden, unreadable reasoning to achieve
• 19:38
  higher performance, we have to ask ourselves
• 19:40
  a deeply difficult question. Are we rapidly approaching
• 19:44
  a future where we have to choose between having
• 19:45
  the smartest AI in the world and actually understanding
• 19:48
  how our own software works?