Opus 5 Is Here And Its BETTER THAN FABLE?!

Englishالعربية Deutsch Español Français हिन्दी Bahasa Indonesia 日本語 한국어 Русский 中文

컴퓨터/소프트웨어경제 뉴스AI/미래기술

Transcript

00:00:00So Claude Opus 5 just dropped in, and Anthropic is claiming we now have a model that is getting

00:00:04close to Fable 5 outputs at half the cost. So let's take a look at what they're giving us.

00:00:09So let's take a look at some of the benchmarks, and right away, kind of wild numbers. We see that

00:00:14it beats Fable 5 and blows Opus 4.8 out of the water on a number of these tests, specifically

00:00:20agentic terminal coding, knowledge work, agentic search. Also, it's beating things like GPT's 5.6

00:00:27in these categories as well. The only areas that we see Opus 5 doing worse than Fable 5 is

00:00:33multidisciplinary reasoning, which isn't by much. In fact, it beats it when it uses tools, as well as

00:00:39legal and health. But every other spot, it's actually beating their top model. And if those numbers are to

00:00:46be believed, that's crazy. Right here, it's saying on cursor bench 3.2 at max effort, the model performed

00:00:51within 0.5% of Fable 5's peak score, but at half the cost per task. Nuts. And you can see that right

00:00:57here. We have Opus 5 in the red, and then Fable 5 in the orange, and down here in the blue is Opus 4.8.

00:01:03So yes, at max, max effort, Fable 5 does pull ahead, but most people aren't actually operating

00:01:08at max. They're either operating at high or extra high. And when we look at extra high and high

00:01:13compared to Fable, we're pretty much getting the same outputs with Opus 5, but again, at like

00:01:17half the cost. And in some of these other benchmarks, like artificial analysis coding agent index,

00:01:23Opus 5 is just straight up beating Fable 5. And same thing with the frontier bench. Like

00:01:27it's, it's actually whooping Fable and it's cheaper. Now let's take a look at some knowledge

00:01:32work and problem solving tasks. And again, I think what's really crazy isn't just this whole

00:01:37like Fable 5, Opus 5 dynamic. It's the jump from Opus 4.8 to Opus 5. Like all of these

00:01:43across the board, just a massive, massive leap forward. And again, if this was like a Sonnet

00:01:495 thing where we saw it doing better than the previous model, yet it was just like prohibitively

00:01:53expensive because Sonnet 5 is somehow the most expensive model out there when we break it down

00:01:57by task. That's not the case here. Opus looks to be extremely token efficient compared to these

00:02:03two other models. And again, like the fact it's actually beating Fable 5, I really can't get

00:02:07over that. Another interesting improvement is visual output. So here's a look at Opus 5 doing

00:02:13some wind tunnel work. And then over here, we have it building a 3D interactive animal cell

00:02:19artifact. Now, one of the downsides with Cloud is it like can't create its own images and stuff. So

00:02:24it being able to do better with these sort of like 3D created SVG, almost like HTML graphics is huge.

00:02:31Another big improvement is that Anthropic is saying it's much stronger in verifying its work

00:02:35and iterating carefully until it succeeds. What does this mean? This means when we use Opus 5

00:02:40on long horizon agentic tasks, things that just aren't one shot. These are like loops, right? This

00:02:45is like, you know, the whole graph engineering thing people are talking about. Anything that requires

00:02:49Claude to essentially check its work against some sort of goal we give it and keep doing that over and

00:02:53over and over again. Well, Opus 5 is much better than Opus 4.8 in that regard. And one interesting

00:02:58example here is Opus 5 was given a drawing of a machine part and told to write code to rebuild it as a

00:03:033D free CAD model. But the model wasn't given any way to actually view the drawing. So, you know,

00:03:09it was told to reach this goal. It wasn't told specifically how to do it. So it went ahead and

00:03:14wrote its own computer vision pipeline to pull the geometry from raw pixels and then reconstructed the

00:03:19whole machine part. Again, anything that is long horizon, given some sort of North Star to reach for,

00:03:25Opus 5 is much better than 4.8 in actually completing that. In terms of misaligned behavior,

00:03:29it's also a step above all these previous models. Lower is better here. And we see a huge increase in its capability

00:03:35from Opus 4.8 to find vulnerabilities and exploits in open source code. Now the other cool thing is the

00:03:40safeguards. This isn't a Fable scenario where if you mention anything about cybersecurity biology,

00:03:45it's going to get shut down. Those guardrails are much less strict with Opus 5 versus Fable 5.

00:03:50Now, while there are still some classifiers in there, they say it's 85% less likely that they actually

00:03:56trigger versus Fable. Now, lastly, the big thing is cost. Like I said, this is half the price of Fable.

00:04:01We're looking at $5 per million input and 25 per million output. And like we saw from the benchmarks

00:04:07in the charts, this is a relatively token efficient model. So I'm really excited to try this one out

00:04:12because if the numbers are true, we kind of almost have a better Fable for half the cost.

Key Takeaway

Claude Opus 5 matches or outperforms Fable 5 across most agentic coding and knowledge benchmarks at half the cost, priced at $5 per million input and $25 per million output tokens.

Highlights

Claude Opus 5 delivers output performance within 0.5% of Fable 5 on CursorBench 3.2 at maximum effort while cutting task costs in half.
Anthropic sets the price of Opus 5 at $5.00 per million input tokens and $25.00 per million output tokens.
On high and extra-high effort levels, Opus 5 matches or exceeds Fable 5 performance across coding, agentic search, and terminal agent benchmarks.
The model includes guardrail classifiers that trigger 85% less frequently than Fable 5 during cybersecurity and biology tasks.
When tasked with reconstructing a 3D CAD model from an unseen drawing, Opus 5 autonomously created its own computer vision pipeline to parse pixel geometry.

Timeline

Benchmark Performance and Pricing Structure

Opus 5 outperforms Fable 5 and Opus 4.8 in agentic terminal coding, knowledge work, and agentic search.
On CursorBench 3.2 at maximum effort, Opus 5 performs within 0.5% of Fable 5 while cutting task costs in half.
Fable 5 leads slightly in multidisciplinary reasoning, legal, and health benchmarks when tools are not used.

Comparative benchmark testing shows Opus 5 surpassing previous frontier models like GPT-5.6 and Opus 4.8 across multiple core operational metrics. While Fable 5 retains a 0.5% edge at absolute peak computation on CursorBench 3.2, Opus 5 produces equivalent results at standard high and extra-high effort settings. On the Artificial Analysis Coding Agent Index and FrontierBench, Opus 5 outperforms Fable 5 without requiring premium token pricing.

Token Efficiency and Visual Output Capabilities

Opus 5 achieves high token efficiency relative to Sonnet 5 and Opus 4.8.
The model generates interactive 3D SVG and HTML graphics directly to compensate for native image generation limits.
Visual outputs include rendered wind tunnel simulations and interactive 3D animal cell models.

Unlike Sonnet 5, which incurs high per-task costs, Opus 5 uses fewer tokens to complete complex queries. Because the Claude suite lacks native raster image generation, Opus 5 relies on programmatic SVG and HTML code rendering to generate precise visual artifacts. Demonstrated outputs include aerodynamic flow visualizers and manipulable 3D biology diagrams.

Long-Horizon Agentic Execution and Safeguard Adjustments

Opus 5 autonomously validates and iterates on long-horizon loop tasks until target criteria are met.
Without explicit instructions, the model wrote a custom vision pipeline to extract pixel geometry from an image and build a 3D FreeCAD model.
Safety classifiers in Opus 5 are 85% less likely to false-trigger compared to Fable 5 on technical domain prompts.

Self-verification and error correction allow Opus 5 to handle multi-step, goal-oriented engineering workflows without human intervention. During testing, the model successfully reconstructed a mechanical component in 3D CAD by writing code to analyze image pixels directly. Additionally, Anthropic reduced guardrail friction for technical topics such as cybersecurity and biology, lowering false-positive interventions by 85% relative to Fable 5 while improving open-source vulnerability detection.

Community Posts

Write about this video