Opus 5 Is Fable 5 for Half the Money?

Opus 5 Release and Benchmark Performance [00:00]

• Opus 5 is Anthropic's latest release, noted for being on a Friday and outperforming Fable 5 on major benchmarks.
• It achieves this performance at half the price of Fable 5.
• The model shows exceptional scores on third-party benchmarks like Arc GAIT, reaching 30%.

  You can say that they really cooked with this model.

Agentic Coding Capabilities [01:01]

• Opus 5 excels in agentic coding, demonstrated by a task to rebuild a machine part from a drawing into a 3D model.
• The model developed its own computer vision pipeline to interpret the drawing without direct visual input.
• This showcases the model's agency and ability to handle complex tasks.

  So, the model has agency.

Third-Party Benchmark Comparisons [01:31]

• Cursor Benchmark shows Opus 5 on max settings is close to Fable 5 on max settings at a lower price.
• On high settings, Opus 5 surpasses Fable 5.
• It also surpasses GPT-5 6 solve on extra high settings.

  And if you look at high, it actually surpasses Fable 5.

Frontier Code and Thinking Budget [02:17]

• Frontier Code benchmark places Opus 5 on par with Claude 3 5.
• Increasing the thinking budget improves performance, though it increases cost.
• Opus 5 surpasses GPT-5 6 solve across different thinking settings, especially in agentic coding.

  Specifically for agentic coding.

Availability and Pricing [03:20]

• Opus 5 is available on claude.ai and Claude console.
• Pricing remains the same as Opus 4 8, with performance improvements at the same cost.
• Reduced cybersecurity capabilities may have aided timely release.

  They're keeping the same price level, but with every we are getting performance improvements and intelligence increase.

Token Usage and Performance [04:19]

• Reports of Opus 5 being token-hungry are not strongly supported by performance plots.
• For similar performance levels, it does not appear significantly more costly than Opus 4 8.
• Higher efforts and performance do increase cost, but hopefully not substantially for real-world tasks.

  For the same level of performance, and actually or even higher level of performance, you are getting better pricing.

Creative Web Design Tests [05:33]

• The first prompt asked for an encyclopedia of the first 25 legendary Pokémon, resulting in a well-designed website.
• The output showed significant creative freedom and a nice aesthetic, surpassing typical card-like LLM outputs.
• Opus models are noted for their web design capabilities.

  And it's a really, really nice-looking website. I'm kind of impressed.

Animation and Spatial Understanding Test [06:23]

• A second prompt involved creating an animation of a crowd forming text, requiring spatial tracking.
• The model successfully created the animation, forming "Hello World, I'm Opus."
• Later, it was updated to include camera controls, allowing zoom and angle adjustments.

  Oh, wow. All right, awesome. This is pretty neat, right?

ISS Tracker Task and Performance [08:06]

• The ISS tracker task took over 50% of the time and was still working on claude.ai, indicating slowness.
• The final output was impressive, showing status levels, tracking, and a potential path.
• The detail and accuracy of the ISS tracker were highly praised.

  This thing is really, really good.

Final Thoughts and Future Plans [11:00]

• The speaker concludes that Anthropic "really cooked" with Opus 5.
• A more detailed video with production code testing is planned.
• The current first impression suggests Opus 5 is a very impressive model.

  So far, I am going to say that on topic really cooked with this model.