10:14Chase AI · Tutorial
How Claude Fable 5.1's Text Watermark Actually Works (And How to Get Around It)
It's not a hidden mark in the text, it's a loaded die on word choice, and only a key Anthropic won't hand out can read it.
September 2ndThe argument in one line.
Anthropic's newly announced Claude Opus 5 matches or beats the rival Fable 5 model across nearly every benchmark category while costing roughly half as much per task, and the jump from Opus 4.8 to Opus 5 is a bigger leap than the gap between competing labs' flagship models.
Read if. Skip if.
READ IF YOU ARE…
- You're choosing between Claude and a competing frontier model for agentic coding or agent-building work and want current benchmark and pricing specifics.
- You build long-horizon autonomous agent tasks and care about a model's ability to verify its own work across many iterations, not just one-shot accuracy.
- You want a fast read of what actually changed in Anthropic's own announcement without opening the source page yourself.
SKIP IF…
- You want a hands-on Opus 5 coding tutorial — this is a benchmark read-through of an announcement page, not a build-along.
- You don't care about model-vs-model cost-per-task comparisons or effort-level tuning.
The full version, fast.
Anthropic's Opus 5 announcement shows the model matching rival Fable 5 at high and extra-high effort settings while costing about half as much per task, and only trailing at max effort by roughly 0.5% on CursorBench. The bigger story per the host is the jump from Opus 4.8 to Opus 5 itself — a larger leap than the gap to competing labs — driven by stronger self-verification on long-horizon agentic tasks: the model checks its work against a goal and iterates rather than one-shotting. Anecdotal examples include writing its own computer-vision pipeline to reconstruct a machine part from a drawing it couldn't view, fixing a bug at its root cause rather than the symptom, and building its own test harness to validate a market data feed with no live reference to check against. Pricing lands at $5/$25 per million input/output tokens, unchanged from Opus 4.8 and about half of Fable 5's rate.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:46
01 · Opus 5 drops — beating Fable 5 on most benchmarks
Comparison table across agentic terminal coding, knowledge work, novel problem-solving, agentic search, multidisciplinary reasoning, computer use, agentic coding, business workflows, legal, health, biology. Opus 5 beats Fable 5 and Opus 4.8 in most categories, trails only on multidisciplinary reasoning. CursorBench: within 0.5% of Fable 5 at max effort, at half the cost; roughly equal at high/extra-high effort.
01:46 – 02:12
02 · Token efficiency vs. Sonnet 5
The Opus 4.8-to-Opus 5 jump is framed as the bigger story than the Fable 5 comparison. Unlike Sonnet 5 (described as disproportionately expensive per task despite improvements), Opus 5 is token-efficient relative to both Sonnet 5 and Fable 5.
02:12 – 02:44
03 · Visual output: wind tunnel and 3D cell demos
Two embedded Anthropic demo clips: Opus 5 generating a wind-tunnel airflow visualization over a car model, and building an interactive 3D cell artifact. Framed as closing a longstanding gap since Claude models can't natively generate images.
02:44 – 03:55
04 · Working with Opus 5: self-verification examples
Three anecdotal examples from Anthropic's page: reconstructing a machine part in FreeCAD from a drawing it couldn't view (wrote its own CV pipeline), fixing an open-source bug at its root cause instead of the symptom, and building its own test harness to validate a market data feed with no live reference. Misaligned-behavior audit and vulnerability-finding also improved.
03:55 – 04:18
05 · Safeguards and pricing
Cybersecurity classifiers are less restrictive than Fable 5's (85% less likely to trigger) via a Cyber Verification Program for vetted users. Opus 5 is priced at $5/$25 per million input/output tokens — same as Opus 4.8, about half of Fable 5.
Lines worth screenshotting.
- Opus 5 performs within 0.5% of Fable 5's peak CursorBench score at max effort, but at half the cost per task.
- At high and extra-high effort settings — where most users actually operate — Opus 5 and Fable 5 produce roughly equivalent output, with Opus 5 costing about half as much.
- The jump from Opus 4.8 to Opus 5 is described as a bigger leap than the gap between Opus 5 and its closest competitor, Fable 5.
- Opus 5 is Fable 5's near-equal on nearly every benchmark shown except multidisciplinary reasoning, where Fable 5 still leads unless tool use is allowed.
- Given a machine-part drawing but no way to actually view it, Opus 5 wrote its own computer-vision pipeline to extract geometry from raw pixels and rebuilt the part as a 3D CAD model.
- On a real open-source bug, a competing model's patch fixed only the surface symptom; Opus 5 found and fixed the underlying root cause.
- With no live data feed to validate against, Opus 5 built its own test harness to confirm a market-data-feed integration was parsing correctly.
- Opus 5 scores lowest (best) of Opus 4.8, Mythos 5, Sonnet 5, and itself on an automated misaligned-behavior audit.
- Opus 5's cybersecurity safety classifiers are described as 85% less likely to trigger than Fable 5's equivalent guardrails.
- Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8, and about half of Fable 5's rate.
What Opus 5's benchmarks actually tell you
MODEL COMPARISON
Benchmark wins depend heavily on effort level and cost framing, and the real story in this announcement is a self-verification jump, not just a leaderboard position.
- Benchmark leaderboards shift with effort level, not just model choice — Opus 5 matches Fable 5 at high effort but only Fable 5 pulls ahead at max effort, by about half a percent.
- A model's true cost advantage shows up in cost-per-task charts, not headline win/loss counts — Opus 5 reaches similar scores at roughly half the spend.
- The jump between two versions of the same model line (Opus 4.8 to Opus 5) can be larger than the gap between competing labs' flagship models.
- Long-horizon agentic tasks reward a model's ability to verify its own work and iterate, not just first-pass accuracy — that's a different skill than one-shot benchmark performance.
- When a model isn't given a way to directly perceive its input, like a drawing it can't view, it can still solve the problem by building its own tool to extract the missing information first.
- Safety classifier strictness and raw capability are separate axes — a model can get more capable and less restrictive at the same time, which changes how usable it is for edge-case-adjacent work like cybersecurity research.
Terms worth knowing.
- CursorBench
- A benchmark (version 3.2 referenced here) that scores agentic coding performance against cost per task at different effort levels.
- Frontier-Bench
- An Anthropic-referenced benchmark suite used to evaluate agentic coding and problem-solving performance, including the machine-part CAD reconstruction example.
- ARC-AGI-3
- A benchmark used here for 'novel problem-solving,' plotting score against total evaluation cost.
- Agentic effort level
- A setting (e.g. high, extra high, max) that trades more compute/cost for higher task performance on a given model; benchmark charts in the video compare models across these levels rather than at a single fixed setting.
- Misaligned behavior audit
- An automated evaluation Anthropic runs to score how often a model's behavior diverges from intended goals; lower scores are better.
- Cyber Verification Program (CVP)
- An Anthropic program mentioned in the safeguards section that gives enterprises and researchers who are already vetted early access to a version of the model with narrower cybersecurity safety classifiers.
Things they pointed at.
Lines you could clip.
00:00
“Claude Opus five just dropped in. Anthropic is claiming we now have a model that is getting close to Fable five outputs at half the cost.”
01:32
“At max max effort, Fable five does pull ahead, but most people aren't actually operating at max. They're either operating at high or extra high.”
02:55
“It went ahead and wrote its own computer vision pipeline to pull the geometry from raw pixels and then reconstructed the whole machine part.”
04:16
“If the numbers are true, we kind of almost have a better Fable for half the cost.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
So Claude Opus five just dropped in. Anthropic is claiming we now have a model that is getting close to Fable five outputs at half the cost. So let's take a look at what they're giving us.
So let's take a look at some of the benchmarks. And right away, kind of wild numbers. We see that it beats Fable five and blows OPUS 4.8 out of the water on a number of these tests, specifically agentic terminal coding, knowledge work, agentic search.
Also, it's beating things like GPT's 5.6 SOL in these categories as well. The only areas that we see OPUS five doing worse than Fable five is multidisciplinary reasoning, which isn't by much.
In fact, it beats it when it use uses tools as well as legal and health. But every other spot, it's actually beating their top model.
And if those numbers are to be believed, that's crazy. Right here, it's saying on CursorBench 3.2 at max effort, the model performed within point 5% of Fable five's peak score, but at half the cost per task. Nuts.
And you can see that right here. We have Opus five in the red and then Fable five in the orange, and down here in the blue is Opus 4.8. So, yes, at max max effort, Fable five does pull ahead, but most people aren't actually operating at max.
They're either operating at high or extra high. And when we look at extra high and high compared to Fable, we're pretty much getting the same outputs with OPUS five, but again, at, like, half the cost. And in some of these other benchmarks, like artificial analysis coding agent index, Opus five is just straight up beating Fable five.
And same thing with the frontier bench. Like, it's it's actually whooping Fable, and it's cheaper. Now let's take a look at some knowledge work and problem solving tasks.
And, I think what's really crazy isn't just this whole, like, Fable five, Opus five dynamic. It's the jump from Opus 4.8 to Opus five. Like, all of these across the board, just a massive, massive leap forward.
And again, if this was like a Sonnet five thing where we saw it doing better than the previous model, yet it was just like prohibitively expensive because Sonnet five is somehow the most expensive model out there when we break it down by task, that's not the case here. Opus looks to be extremely token efficient compared to these two other models.
And again, like, the fact that it's actually beating Fable five, I really can't get over that. Another interesting improvement is visual output. So here's a look at Opus five doing some wind tunnel work.
And then over here, we have it building a three d interactive animal cell artifact. Now, one of the downsides with Claude is it, like, can't create its own images and stuff.
So it being able to do better with these sort of, like, three d created SVG, almost like HTML graphics is huge. Another big improvement is that Anthropic is saying it's much stronger at verifying its work and iterating carefully until it succeeds.
What does this mean? This means when we use Opus five on long horizon agentic tasks, things that just aren't one shot. These are like loops.
Right? This is like, you know, the whole graph engineering thing people are talking about. Anything that requires Claude to essentially check its work against some sort of goal we give it and keep doing that over and over and over again, well, OPUS five is much much better than OPUS 4.8 in that regard.
And one interesting example here is OPUS five was given a drawing of a machine part and told to write code to rebuild it as a three d free CAD model, but the model wasn't given any way to actually view the drawing. So, you know, it was told to reach this goal. It wasn't told specifically how to do it.
So it went ahead and wrote its own computer vision pipeline to pull the geometry from raw pixels and then reconstructed the whole machine part. Again, anything that is long horizon, given some sort of north star to reach for, OPUS five is much better than 4.8 and actually completing that. In terms of misaligned behavior, it's also a step above.
All these previous models, lower is better here. And we see a huge increase in its capability from OPUS 4.8 to find vulnerabilities and exploits in open code. Now, other cool thing is the safeguards.
This isn't a Fable scenario where if you mention anything about cybersecurity biology, it's going to get shut down. Those guardrails are much less strict with OPUS five versus Fable five. Now, while there are still some classifiers in there, they say it's 85% less likely that they actually trigger versus Fable.
Now lastly, the big thing is cost. Like I said, this is half the price of Fable. We're looking at $5 per million input and 25 per million output.
And like we saw from the benchmarks in the charts, this is a relatively token efficient model. So I'm really excited to try this one out because if the numbers are true, we kind of almost have a better fable for half the cost.
The Hook
The bait, then the rug-pull.
Anthropic just published its Opus 5 announcement, and the host reads it live on screen — a benchmark table and a run of cost-per-task charts that put the new model within half a percent of rival Fable 5's peak score, at roughly half the price per task.
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
