This Week in Claude Code News

Okay, so I'll be going over the Claude code  updates of last week. Firstly, you want to   make sure you're on the right version, so you need  to be on 0. 88 or greater. So you can write claude   --version, and you can see it over here, I'm on  0. 89. And you can write claude update if you're   not on the same version as me. I'm not sure what  they changed between 0. 88 and 0. 89 because they   haven't added it to changelog, but we'll see. The  main thing they did in this update is they fixed

some error tracking, usage tracking, and slash  cost. So if you're using the API, then you can   write slash cost, and then you can see how much  cost you've taken up or used so far. I'm using   the Claude Max plan, so I don't have to worry  about that. But I can see my own cost here using   the ccusage status line, which if you don't have  installed, you can download by writing ccusage   statusline. And then finding on GitHub, it should  be linked down below as well. And it should look   something like this. Something else you can do is  set environment variables for these two over here.

So if you write slash model, for example, then you  should see it says default for up to 50% of usage,   depending on your plan, or Opus or Sonnet, or Opus  plan mode, which we'll use Opus in that particular   plan mode. If for some reason you want to switch  out the model for an older model or a model from   a different provider, if you change your base URL  as well, then you can do that over here. So I can   go back to Claude Opus 4. 0 for the Opus model,  and then go back to Sonnet 3. 7, for example,   for the Sonnet model over here. And then I can  press Enter, and then I can run Claude, and now   I'm using some of the older models. And you can  see that over here on the status line, because it   now says Opus 4 instead of Opus 4. 1, which I said  earlier. I think this is probably more useful if   you're using models from a different provider, but  in that case you would have to change your base   URL as well. I'm not sure what they added in,  8. 7 over here, because there's a bit of a gap,

so maybe it's nothing important. The next thing  you can do is you can write slash context,   so you can actually see the context that's being  used so far. I have a Claude session that I just   ran over here, and I can write slash context  over here. And then I should be able to see   it's used this amount of messages in the  message context, this amount of memory files,   this amount of system tools, and the system prompt  is about 3,000 tokens, and this is how much free   space I have left. So that can be useful if you're  like debugging context-related issues. It can   especially be useful in case like you've filled  your entire memory or put different MCP tools or   something, and you can know the effects of that  because some people like to do that for some   reason. They have like 50,000 tokens of MCP tools,  for example. But yeah, I think this design over

here is pretty cool. In 8. 5 here, they included  session cost info in the status line. So I usually   prefer to use ccusage's status line, but if you  prefer making your own status line, then you can   do status line. And then you can say include  how much I've used this session, for example,   the cost information. And then it will make your  custom status line based on that. They also fix

some bugs over here. With the at-mention, they now  include the. claude folder in your home directory.    So if you write. claude over here, at. claude/,  then you can see users/ray/. claude. And the other  . claude is the one in the project folder. So you  can see the output styles that I added before and   so forth. They made it so that ripgrep is used by  default instead of grep. So ripgrep is a faster

version of grep, which basically means that if  you have a really large codebase, then you should   notice some performance improvements when Claude  is searching through the codebase to find relevant   things. And you can opt out of that behavior by  using this command before running Claude Code or   putting that in your environment variables. They  also made it so that @-mention supports files

with spaces in the path. And I don't think that  should matter for a lot of programmers because   programmers usually don't have spaces in the path.   But for other people, it may matter more if you're   using Claude Code for something else. So you can  write at, and then you can write speech mark over   here. something with  a space in the name. So you can write test 1.    And you can see I have test12, which is a folder,  and test123, which is a file. Both have spaces in

them. They also added a new shimmering spinner.   And that basically means the star over here,   you can see it's moving left and right and getting  bigger and smaller. I don't know if you remember,   but the previous star basically only got bigger  and smaller rather than moving left and right as

well. And that's basically everything that they  explicitly mentioned. Also, Claude Sonnet 4 did   introduce a 1 million token context window. So  some people are wondering when it's going to   be available in Claude Code. And it seems that  it's rolling out slowly. So for example, this   user now sees it in their slash model. So if they  run slash model, they can see it says default,   Sonnet 4 with 1M context. And they have a video  proof of it actually like appearing when they   run slash model over here. I still don't see the  option, but I don't think this will be a big issue   for many people. It's still about ensuring  your context window does not get polluted,   even though you can have longer chats. If you're  working on many different features within the   same chat, for example, then you can have many  different distractors in your context window   that can reduce the model's performance.   And learning more about what I just said,   then you can watch my previous video about  Claude's new context window. And that should   be linked down below. But if you are really keen  to try the new context window, then Cursor does   support it with Claude 4 Sonnet. And you can also  use a cursor CLI if you want a more CLI experience

for it. Another thing related to Claude Code is  that DeepSeek drops another model, DeepSeek-V3. 1.    And they say it's better at multi-step agent  tasks, which means that it's going to be better   at coding. And they actually made a Anthropic  API endpoint, which means that you can now use   DeepSeek in Claude Code and benefit from all the  features that Claude Code has available but using   DeepSeek. And I go through how to do this in  my previous video. It should be linked above   somewhere right now and also down below in the  description. Another relevant thing for Claude

Code is that Salesforce AI Research came out with  a new paper two days ago introducing MCP Universe,   which is a brand new benchmark to evaluate LLMs  and realistic and hard tasks for interactions   with real world MCP servers. And you can scroll  down and see the results and read through the   paper as well. But basically GPT-5 comes out on  top across all these different categories. And   the scores are pretty low because these are  pretty hard tasks. And of course, they don't   have coding as a benchmark over here. But you  can see that Claude-4. 0-Sonnet performs pretty   badly when it comes to web searching. And I found  that to be the case myself when I get Claude Code   to do any web searching through the web tool that  it has available or to use a Tavily MCP server to   find information online. Then it performs pretty  badly compared to when I use GPT-5 in ChatGPT. So   often what I do is I use ChatGPT with thinking  mode on and also the high setting and also like   web search enabled to find information online  or documentation or find an implementation for   something. And then I pass that information  back to Claude Code because GPT-5 is just   better at web searching when you have the right  settings. And you can use something like RefTools,   which provides context. It's like context 7 as  well. And that is an MCP server and it can find   good documentation for you. But if you're looking  at more like nuanced things or things that may   not be in the documentation when it comes to like  finding solutions to bugs that you're facing, then   I think using GPT-5 will be a better approach than  using the default search capabilities of Claude

Code. A trend that has been happening in the last  few weeks is that people have been running Claude   Code in a VPS and having it go crazy and skipping  permissions and basically just vibing with it. So   you can do this yourself in a VPS like Hetzner. So  if you sign up to Hetzner, you can get a VPS for   $4 a month, not including tax. And you can use my  referral code in the description down below to get   $20 in cloud credits. That should be linked down  below. So it's basically free if you sign up   using that code. And basically you can make your  own VPS over here. You can choose a region. You   can set your SSH key. And then once it's set  up, you can go inside your VPS, install NPM,   install Claude Code. And then after SSHing into  your server, installing Node. js, installing NPM,   and then installing Claude Code as well, you can  run this command over here. And then press enter.    And then you can just set up Claude Code again.   And because they can't really open the link in   the browser, you have to click on this URL and  then copy the code and then paste it back in over   here. And then when it says Claude Code is running  in bypass permissions mode, you can press yes, I   accept over here. Then you can say something like  install Docker and Docker Compose and set up a   LiveKit server for me with a Next. js video calling  chat application attached to it. And then you can   see how it performs on this like bigger task. And  I think this is more for like playing around and   having fun with it. And then basically seeing what  the capabilities of different models are. Because   you can also swap out the Anthropic models with  models from other providers. And then you can have   like two different VPSs. And you can see how they  both like execute on these longer horizon tasks,   which are more difficult as well. And you can  get a feeling for like the raw behavior of it.    When you're not having to approve it every single  time when it wants to do something. There have   also been people who have been connecting Claude  Code, which is running in a VPS or on their local   laptop at home to their phones as well. So they  can use Claude Code on the go and then like keep   track of anything that's happening. I may make a  video about that as well. And yeah, it seems that

Claude Code is underway. It installed Docker,  Docker Compose. It's now setting up a front end   as well for the application. But yeah, that's  basically it for the video. I do go through the   changelog every single week and talk about any  new updates and also news surrounding the Claude   Code community as well. So if you do enjoy that  kind of stuff and you do want to see more of it,   then do subscribe for that. Because I  will continue to do this until Claude   Code is no longer relevant or it's replaced by  something better or just anything else like that.