Opus 5.5: Anthropic Just Built the Best Coding Model

Author: Prompt Engineering · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Opus 5.5: Anthropic Just Built the Best Coding Model · Channel: Prompt Engineering · 09:13 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI content generation, Markets and economy, AI models

Summary

The video presents Opus 5.5 as a strong coding model that performs well across agentic benchmarks, including Terminal Bench, while costing less than its predecessor. The presenter highlights its verification loops for coding tasks and a more natural, readable communication style. Demos show visual projects ranging from a black hole simulation to an ISS tracker and a product-launch website, with mixed but generally positive results. The video also notes lower API pricing and rate-limit resets as benefits of competition among model providers.

Key points

Opus 5.5 is described as competitive with or better than other models on many agentic benchmarks.
The presenter singles out Terminal Bench 4.0 as a useful, relatively unsaturated benchmark.
The model is said to perform well at agentic coding, including verifying its outputs in a browser.
Benchmark results vary by reasoning setting, and the presenter suggests data distributions may explain unusual results.
The presenter praises Opus 5.5’s more natural writing and clearer communication compared with earlier Opus models.
The model is available through Claude Code, the Claude desktop app, and the API.
Demo projects include a black hole simulation, real-time ISS tracking, and a product-launch website.
The presenter also highlights lower API prices and the ability to reset rate limits.

Chapters

Start	Chapter	Summary
00:00	Opus 5.5 and benchmark performance	The presenter introduces Opus 5.5, discusses its pricing, and reviews its results on coding and agentic benchmarks.
02:08	Reasoning results and communication style	The video considers unusual benchmark results, then turns to the model’s writing style and availability.
05:28	Coding and visual demos	The presenter demonstrates a black hole simulation, an ISS tracker, a product-launch website, and 3D voxel art.
08:43	Rate limits and closing thoughts	The video notes rate-limit resets and concludes that Opus 5.5 combines improved communication and coding with lower costs.

Quotes

This is probably one of the benchmarks that is not saturated yet — 01:05

the model is not only writing the code, but actually verifying its outputs usually within the browser — 02:08

the main thing I’m personally excited about is the communication style. — 03:51

it also put the sun on the right side, which is actually a really neat touch. — 06:34

better communication, better code quality, lower pricing, plus lower usage or tokens. — 09:13

Concepts and entities

Opus 5.5, Anthropic, Fable 5.1, Terminal-Bench 4.0, Frontier Code, Claude Code, International Space Station

Full transcript (original language)

00:00 Okay, so Opus 5.5 is here and it beats Fable on almost every benchmark. But the thing that I’m personally excited about is it can write like human, not the AI slop that all the other previous Opus models have been producing. We’ll look at a few demos, but the main headline is this is the first model in their Claude 5.5 family. And it seems to be on par or actually better than Fable 5.1, but costs percent less compared to Opus 5, which is really good news.

00:32 And another good news is that they are actually reducing the API pricing. There’s a minor reduction in both input, output, as well as the hash token pricing, which is going to be very beneficial. And I feel like Anthropic is really feeling the heat because uh as to performance ratio is one of the main things right now for users when they’re selecting models. All right, so how is the performance? It pretty much beats every other model on almost all the agentic benchmark. The most important one for me is the

01:05 terminal bench 4.0. This is probably one of the benchmarks that is not saturated yet and it’s a new iteration of the terminal 2.1. And this model performs really strongly. Even it beats Astra and Fable 5.1. Now, I am putting it to the test because benchmarks usually don’t really represent the real world coding capabilities, but they do give you an indication of what to expect. Now, on the rest of the benchmarks, you can basically see that it is state-of-the-art.

01:37 So far, Anthropic has really been pushing for model intelligence, but now they are also pushing for making that intelligence cheaper. And if you look at this model on terminal bench on different reasoning settings, it not only out proof the rest of the models, but at much lower cost, which is I think really great for us as customers. And you see this throughout all of the agentic coding benchmarks. Now, when it comes to agentic coding tasks, it’s actually really good, especially its verification loops. So, for example,

02:08 here’s my go-to prompt, but right now the model is not only writing the code, but actually verifying its outputs usually within the browser, which actually is a really important step if you want to be running these models unattended for long horizon tasks. I’ll come back to these in a minute. This one is weird. It seems like the medium on uh frontier code and reasoning performs much better compared to high, extra high, and then it performs way similar to the max.

02:45 This whole plot looks really weird, but it’s also true for Fable for some reason. And even for Opus, now the GPT models are doing or behaving much differently on this specific benchmark, right? So, the thing is that the performance of these models I think is really dependent on of what type of data they are trained on. This benchmark may have some other data distribution or it’s going to be probably out of distribution for how they are training the model, and that’s why we are seeing a very weird behavior here. Something to note, I think it

03:19 would be good to look at what exactly is in this frontier code benchmark. Okay, apart from coding, uh its work is another area where these frontier companies are pushing a lot. And again, this seems to be state of the art on the GPT wall. When it comes to intelligence, which is measured in Elo score, and then cost per task. Now, hopefully it’s going to produce much more concise answers. Uh so, they’re not only going to perform better on these benchmarks,

03:51 but hopefully this is to all be also reflected in the real world testing, which I’m going to show you in a minute. Okay, but the main thing I’m personally excited about is the communication style. So, they say that we have made major improvements to the way Opus 4 5.5 writes and communicates. One of the most common areas of feedback we have heard about Opus 5, right? Uh messages Its messages are much easier to understand at a glance, which testers said helped during long running sessions. And if you have used Opus, uh it’s this very peculiar style

04:25 of writing which feels unnatural. So, for example, they’re asking here explain a certain issue. Fable kind of starts going off by the extra drop is not free tier. Any regression in other specific issue, right? Now, on the other side, it’s actually a lot more natural for Opus 5.5. It says the free tier change accounts for only 1.1 of Acme’s August drop, right? So, this is a lot more natural compared to what style is.

04:57 Here are a couple of other examples from the blog post. We’re going to also look at how good it is its explanation is when I am testing the model. This gives you an indication that I hope it’s going to be much more pleasant to read its outputs. Okay, so the model is available on both Claude code and the Claude desktop app. You’ll have to update them, and you can also use it through the API with Opus 5.5 tag. Now, there are a whole bunch of things

05:28 regarding data retention, distillation, and safety. I’ll highly recommend to read through them if you’re interested in that, but let me show you a couple of demos. Okay, so this first demo is black hole simulation. I provided a very detailed prompt, and I’m going to say I’m pretty impressed with the visual output quality that it’s able to produce. I think it’s pretty neat. Now, I think it probably is a lot brighter compared to what I was hoping for. But, we could adjust things like density. It gives us all of the

06:02 different controls that you would need, right? Now, this is running on my CPU. So, since there are a lot of different processes competing for resources, so the smoothness may not be as good as I would hope for. But, overall, I’m actually pretty impressed with what it’s able to produce. Okay, the second one is a real-time tracking of International Space Station. It’s one of my go-to prompts. This not only has the model capability to produce visually stunning outputs, but also

06:34 uh stability to make real-time API calls. And I think it’s one of the best outputs that I have seen so far. Especially, not only it’s tracking the real-time location of the ISS, but it also put the sun on the right side, which is actually a really neat touch. Okay, here’s my another go-to prompt where I asked the model to do research on its own and then put together a website for product launch. Now, in this case, I would say the visual quality can definitely use some work. It

07:06 doesn’t look like AI slop, but I think I would probably was expecting a lot more. However, it added these really neat animations when it comes to the benchmarks. And benchmark numbers actually seems to be correct, right? So, throughout the website as you go down, you see these really interesting prompts. These are some of the example that they have highlighted. It put a comparison between Opus 5 and Opus 5.5, which is I think it’s pretty great. Now, by default, they

07:41 set the reasoning level to medium, which it captured. And they actually recommended probably it’s better to run it on high for more complex tasks. But overall, I’m pretty happy with what it’s able to produce. Especially these small visual elements that it’s adding. Now, on the first go, I thought I probably didn’t really like the visual quality. However, as I’m going through the website, I think it’s actually pretty neatly put. Oh, okay, this is pretty

08:12 awesome. And keep in mind, it’s just Opus 5.5 running in Claude code coming up with its own visuals without using any assets, which is pretty incredible. Now, as expected, it’s actually really good at 3js, and it can produce these voxel arts. This is a very detailed version that I think probably I have seen from any model. And it followed the prompt really accurately, right? So, a really good job with the visual assets that it can produce. Now, I’m going to

08:43 show you a couple of other really interesting things that I found while recording this video. And it’s not only they are increasing the 5-hour limit. I was hoping that they would also increase the weekly limits. Now, they’re giving the users the ability to reset their rate limits, which is great. Now, this is probably inspired from OpenAI cuz OpenAI has started doing this, and this is why competition is really good for us customers. Model companies not only have

09:13 to compete on the model quality, but also on pricing and then incentives like banked resets. Okay, so this was a quick update on Opus 5.5. It seems like a really good model, actually with better communication, better code quality, lower pricing, plus lower usage or tokens.