Anthropic have finally done it - they've taken the crown for the best coding model in the world. Last week we had Gemini 3 and GPT-5.1, but they all lacked one massive crucial ability and that's coding.
Opus 4.5 is smashing it on all the coding benchmarks, especially the SWE benchmark where Gemini 3 fell short. And here's the best part: get it on it now because it's cheaper until December 5th on most coding platforms - it's the same price as Sonnet!
The token efficiency is a lot, lot better and the training cutoff date is August 2025 - a really close cutoff date that's good for up-to-date libraries and coding practices.
They've also released a really interesting new search tool that helps with the MCP problem by using defer loading to save context and money. This should help massively with the MCP problem I've been complaining about!
—
TIMESTAMPS:
0:00 - Opus 4.5 Takes the Crown
0:49 - The Price Drop Announcement
1:32 - Why the Cutoff Date Matters
2:57 - Claude Code in Desktop App 🖥️
3:38 - The New Search Tool Solution
4:51 - How Defer Loading Works
6:07 - Solving the MCP Problem
7:00 - The Cost-Saving Reality 💸
8:17 - What is "CMD + AI"?
—
Unlock the full potential of your online presence with Kabarza and Samuel—experts in web design and development (respectively), powered by cutting-edge AI solutions. We blend creative design with advanced tech to deliver smart, high-impact websites that stand out. Ready to elevate your business? Contact us today and see what AI-driven innovation can do for you!
LINKS & RESOURCES:
Check out Kabarza's amazing work: https://kabarza.com
Visit Samuel's website for more: https://samuelgregory.co.uk
📷 Follow on Instagram: https://www.instagram.com/cmdaishow
—
HASHTAGS:
#Opus45 #Anthropic #Claude #Coding #Programming #AICoding #Gemini3 #GPT5 #CodingModel #AIProgramming
Transcript
Anthropic have finally done it. They've taken the crown for the best coding model in the world. Last week we had Gemini 3, GBT 5.1, but they all lacked one massive crucial ability and that's coding. The benchmarks are just insane. We won't focus too much on this because I know benchmarks are a bit of a yfest, but these are all the ones that it's smashing it on. And the SWE benchmark was one that again Gemini 3 didn't was the one that it lost on even last week and now Opus 4.5 has taken the lead again mostly kind of developer stuff and whatever not so much on you know everyday tasks graduate level reasoning visual reasoning multilingual Q&A and things like that it's it's not it's not great on that but that's not why we love it right get it on it now because it's cheaper until December 5th on most of the coding platform. Most of the like plugins and things like that, it's the same price as Sonnet. So, it's $5 in, $25 out. It used to be, correct me now, it used to be $15 in, $75 out. So, it's already cheaper and now it's cheaper again. It's the same price as Sonic on most of the PL except claw code, but I'm guessing the um limits are lower uh higher. Sorry. This is amazing. I would suggest using open opus plan mode which I tend to give opus the planning the structure and all of that stuff with with coding tasks and then I have set actually do the coding tasks. I have haiku set on my agent so my agents run with haiku just so it can churn those out um quickly. I'd be tempted to sometimes use Haiku depending on the coding task using Haiku in the actual coding thing, but I always use Opus. Get I would strongly recommend just getting used to planning uh all of your coding tasks before actually doing them, even if they're kind of fairly short, fairly quick. So yeah, price very very welcome. Token efficiency is a lot lot better as well. And another thing is the cut off date for the information. So it is August I think I read training cutoff date. So Opus 4.5 is here is August 2025. So it's a really really close cut off date. And this is good for like you know up to-ate libraries just coding practices and things like that. I have some MCPs running that allow me to get the latest like articles and documentation and things like that. uh X MV MCP and ref MCP but it cannot hurt to have really up-to-date knowledge even the reliable knowledge cutoff date is March 2025 so this is this is really really good in the clawed desktop app they've added now clawed code in the desktop experience which is amazing kind of replaces clawed code for web a little bit I can't tell you how pricing works for because you last month they gave you $1,000 of credit using claude code for the web. Is the Is this claw code for the web? Technically, it's got to spin up a sandbox environment. Technically, there's a lot of import like uh crucial things. So, this is like what claw code now looks like. And then you've got code up here, which now looks like this. And even even like some of the stuff I've been doing on clawed code for the web appears in here. So, again, is this using claw code credits? I'm not too sure, but either way, I'm here for it. And you're doing all this inside of just one desktop app. It would be cool if this connects with my local instance of clawed code really because um it's just a nicer UI, isn't it really? We all like a bit of nice UI, but presumably again this has got to spin up environments actually. So any look at that. Um so yeah, claw code for the web. The other thing is that they've released a really interesting new uh search tool. They call it a tool search tool. MCPS take up a lot of context. Tools, not just MCPS. I keep using the term MCPS, but it's not. It's tools as well. They take up so much context even before you've written a line of code, even before you've written a prompt or anything like that. And so we've seen a few articles w with anthropic discussing tools and how much how many tokens they take and introducing things like skills and all of this in the name of making tool use more efficient. But with with Opus along with Opus they released a separate tool which searches other tools. So when you make a request to claude I want to try and here we go. Here's a here's a Let's go type script. When you make a normal request to to anthropic SDK, it's not available in claw code. I will say this. When you make a normal request, you pass in a beta flag to use to, you know, because this is a beta feature. And what you normally do is you pass it tools for it to be able to use in that request. And look, here we go. We've got a get weather one. We've got a search um search files tool here. and they've marked these as defer loading true and that enables this tool that you've given it the ability to search them. So I think it's just this that gets sent into the context which obviously is tiny um and this is able to then search that get a get an understanding whether we need to get the weather or or search files or something like that and then use it and then that's when it gets brought into context. So really really cool. Oh, you can you can also defer loading to false if you do want that tool call because presumably there's a bit of a performance hit here as well. And here's an example of MCP utilization as well. So what have they got? MCP. Oh, here we go. So they've referenced they've this is a tool set. So they've marked this as a MCP tool set and then referenced the database server as the MCP name which is an MCP server. Here they've named it database server. So you're doing that um default config defer loading true. So it's slightly different to you know just just saying defer loading true and same again there. So search events I'm guessing this is maybe a tool within this MCP. I'm not too sure. Regardless this is just how you would use it. This should help massively with the with the MCP problem. And I want this to come inside of chord code. That sounded weird, didn't it? It saves you money. Sorry. At the end of the day, which called Claude or Anthropic are not getting. And I'm so I'm thinking, is this a way to gather money because obviously claw code is a much cheaper version than using actual API calls. It's like, you [snorts] know, if you if you aren't subscribed to Claude Claude code or Claude's max plan and then you use their models inside of cursor, those tend to be quite expensive, right? This is this, you know, you might see this stuff used inside of kilo code, inside of augment, all of those extensions using these features, but still because they are using the API pricing, they're going to be more expensive. they'll figure it out and then I think they'll build this into claw code or they'll make it into like a skill like the skill would be the thing that um utilizes this. I I don't know how exactly they're going to do it but just to say that it's not inside of claw code yet. I'm waiting for it because I'm fed up of turning MCPs on and off and managing them and this and that. It's it's a pain in the bum because I'll forget an MCP is running and I'll be like I've just done all that work and all the while I've had task master and taskmaster MCP running which takes up so much you know context and whatever. So that's that's kind of Opus 4.5 in a nutshell. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter.