HUGE Claude Code Update: New Model + Mobile & Web App?

Okay, so I'll be going over all the brand new  Claude Code news from the last week. There   will be timestamps down below, so you can skip  ahead to any point you find useful. Firstly,   you want to make sure on the latest version of  Claude by doing claude update. And then as of the   time of recording, it's 2. 0. 19. They added a brand  new model to model selector, which is Haiku 4. 5.

And that is basically the like least powerful, but  fastest version of all the Claude models. And you   can see if you do /model right over here. And it's  actually just as powerful as Sonnet 4 was when it   comes to software engineering, according to this  benchmark over here. So you can see Haiku is 73. 3   and Sonnet 4 is 72. 7 on the software engineering  benchmark verified. And with the same performance   of Sonnet 4, it's one-third of the cost, but more  than twice the speed. But some people have found   that it's faster than that because this person  says that's 3. 5x faster in their experience. I did   a tiny bit of testing of the model earlier today,  and I have found it to be much faster and just   as accurate as Sonnet 4 was when it comes to a  small add-on that I'm making for Anki. And that's   basically like an AI language conversation like  add-on when it comes to Japanese to basically help   me learn words better. And I think that for really  small codebases and really well-defined tasks,   it is performing pretty well and I will be using  more for that purpose. One thing that I have found   quite annoying recently is because often I use  Claude Code for a task and then Claude Code can   take up to seven to 10 minutes for the exact  same task. That is like relatively easy and   I would have just given to Haiku instead.   But of course the previous Haiku model was   not that good at coding and it seems they've  made it much better and faster. So honestly,   I think a pretty viable strategy is that you can  use Sonnet 4. 5 to come up with changes that should   be made and then use Haiku 4. 5 to actually execute  on those changes because of its performance when   it comes to software engineering and its speed as  well. And then you can probably use Claude Code   if you still like to use that for any of the much  harder tasks or any of the like persistent bugs   that Claude Code is not able to solve. So doing  something like that may mean that you won't hit   your Claude Code rate limits quite as much because  of course Haiku 4. 5 is three times cheaper than   Sonnet 4. 5. The only smaller disadvantage seems  to be that it still has a context window of around   200,000 tokens, whereas there's a variation of a  Sonnet 4. 5 which has a context window of a million   tokens. So they'll probably release a million  token version soon enough, we'll just have to   wait and see. And I think that Opus 4. 5 will be  released soon enough because on the Claude Code

website, the source code, there were two brand  new models, which are Rainbow Butterfly Stream   and Crystal Mountain Breeze. That was like in the  JavaScript source code or something like that.    And then two days later, one of them turns out to  be 4. 5 Haiku, and I think the remaining one would   be 4. 5 Opus unless they come out with a brand  new, completely different model instead. They

also introduced a brand new Explore subagent,  which is powered by the Haiku model, and it   will search through your codebase to efficiently  save context. So you can do @ and then do explore,   and that will trigger the default subagent,  which is explore. Then if I press enter within my   HyperWhisper codebase, you can see that's going  through the codebase pretty quickly. And it's   going to provide a comprehensive overview of these  components over here. So the project structure,   key technologies, main components, architecture  patterns, entry points, dependencies, and some   notable design decisions as well. And after  about 72 seconds, it came up with this pretty   comprehensive list of everything. So the project  structure, technology stack, and so forth. And   yeah, it's like pretty long. This could probably  be used to make a claude. md file or a agents. md   file or something along those lines and probably  just getting people onboarded with a brand new   codebase as well. Now, one interesting thing  about this brand new Explore subagent is that   it's baked into the Claude Code system prompt.   You don't have to manually call it every time   because I used Proxyman to intercept the request  to the Anthropic API endpoint and extracted the   brand new Claude Code system prompt. And in the  prompt it says, when exploring the codebase to   gather context or to answer a question that is not  a needle query for a specific file/class/function,   it is critical that you use the Task tool with  subagent_type=Explore instead of running search   commands directly. And then you can see right here  that it gives some examples. So I think this will   probably mean that Claude Code is better at  finding things within your codebase because   of the specific subagent when it comes to making  changes in your codebase. Another improvement is

that Claude Code will auto-background long-running  bash commands instead of killing them. And you can   customize it by setting this environment variable.   And that basically means that like when you have   your bashes, if you do /bashes, then anytime  Claude Code runs a command and realizes that   it's taking quite long, then it will move it  into background so you can continue coding. I   have found in some cases Claude Code decided  to run a command like npm install whatever,   and then depending on my internet as well, it may  take like three or four minutes to do that. And if   I press escape, then it will kill the command. But  I think in this case it will now put it back into   background. So if I say something like run npm  run dev in the nextjs folder as a bash command,   then it will start running that and I can press  control B to run it in the background. But I think   after a couple of seconds it will automatically  move into background instead. And it seems that   that happens after 30 seconds. If the bash command  has not completed running, then it will be moved   into background. And you should see if you do  /bashes as a background task. So if you want   to change that, you can either go to environment  variables or you can write export this and then   do 15000 for example, and that will be 15 seconds.   They fixed a bug where you can toggle MCP servers

by doing enter after writing them. So you can do  @, for example, shadcn, and that's my MCP server.    Press enter, that turns it off. And then doing  that again turns it back on. So that can make sure   that you're not using too much of your context  window with your MCP servers because you can   toggle them on and off when needs be. They also  wrote document --system-prompt. And I think it

basically means that when you go to claude --help,  then you can see it's now on the list here where   it says system-prompt to use for the session. And  that can basically help you override the default   system prompt. But if you want to add something  to the end of the system prompt for that session,   then you can do append-system-prompt instead. And  it seems that Claude Code will also be integrated

into the mobile version of Claude. You can see  some examples of how it will look because you   can select your repository, connect it to your  GitHub, and it will appear on the left-hand side   in the Claude application, kind of like this.   So there will be like a Code section, but it   has not been released to the general public. So  it's kind of like how on the ChatGPT mobile app,   you can trigger sessions of Codex on the go and  then review on your computer or review on GitHub   as a PR or something like that. And the only time  I really find myself using it is when it comes to   like, I get an email, a support email saying  that there's a certain bug. And I'm not on my   computer. And then I quickly like tell Codex to  find the bug and fix it. And then I review the PR   when I get back to my computer and then publish  that update. That is the only time I really find   myself using it on the go, but I'm sure many  other people have different use cases as well.    And Claude Code will be releasing a web version  of the application soon as well, as you can see

here. And that's basically it for the video. I do  make these videos every time there's a big Claude   Code update. So do subscribe to the channel if  you do like that kind of stuff. And if you do   have a Twitter account, then do consider following  me on Twitter. There will be a link down below.