Claude Code vs Codex vs Cursor  🔗 **Source:** [Claude Code vs Codex vs Cursor](https://www.youtube.com/watch?v=JMYspR42HFM) This is not going to be my usual video.I want to try and break down the realdifferences between things like cloudcode, codecs, and cursor. Not in thelike how smart are they sense or whattasks are they capable of sense. I wantto break down the philosophy differencesbetween these solutions because I findthat they are way more different thanpeople seem to think. I know that we'veall used cloud code and codecs and maybeyou've even tried out the cursor CLI,maybe you use their IDE, maybe you usetheir cloud environment. Whatever it is, many of you guys have tried all of thesetools. Many of you haven't tried some ofthem for a while. Many of you have onethat you really like and you don'tbother with the others. And if that'sthe case, I get it. If you're alreadyusing Claude Code and you're happy, I'mnot here to try and change your mind. Iwant to try and explain philosophicallywhat all of these things are trying todo and how they differ. My main reasonFor doing this, my own philosophy haschanged around doing agent development. Since I moved from Claude over toCodeX, obviously, it also changed whenI moved from Cursor to Claude Code, butit's changed in a way that I like a lotmore this time. And I find that whenpeople are talking about these thingsabout Cloud Code versus Codec, they'reoften just talking about the model orhow the UI looks and not about howdifferent the core of these things are,and how differently the teams buildingthem are thinking about how we makesoftware. To be more direct, this videois happening because this reply onTwitter got way more attention than Iexpected as I broke down some of the philosophical differences between howAnthropic and OpenAI are thinking abouttheir coding tools. But I wanted to goway further than I did here. I want toreally break down how these toolsdiffer, how they can be appliedeffectively in your work, and where II think they might be going long term. I'mgoing to do my absolute best to presentas unbiased as possible because, again,my goal here isn't to try and convinceyou to switch to a different tool. It'sto try and help you understand what thedifferent tools are meant to be builtfor and who they might be best used by. I don't want to sell you anything, okay?I do want to sell you one thing. Today'ssponsor. Today's sponsor is an AI codereview app. Wait, no, don't go. Trustme, this one's actually different. Macroscope's been blowing me and my teamaway. It's quickly become Julius'sfavorite tool that I've added to ourcode bases. It's what we run on T3 Codeand Julius loves it there. But that'snot my favorite part. My favorite partis all the cool status stuff they do.Keeping track of what my team is doingis hard as small as the team is because,Julius just ships so much. Not only doesit estimate how much time people arespending coding, which gets broken byJulius's absurd output. 413 hours soundsinsane, but when you see his commits,you understand what it means there. Thething that's way cooler, that it gives meis the status updates that show what myteam's been up to. Over the last week,Julius has added multi-provider sourcecontrol for T3 Code, PI extensions, andVS Code themes, T3 chat, model routing,billing, and UX, desktop effect framework rewrite. All of these thingsthat I just wouldn't know about becauseit's hard keeping track of everythingmoving on my team that it gives me asthe lead. It's super useful. And beingfully transparent, I've made harddecisions about team structure stuffaround the results I got from this, 'causeit gave me so much better insight onwhat's going on. The review side's evencooler, though. The run agents aregenuinely awesome. You canprogrammatically set up specific, likelint rule style things you want in yourcodebase as a directory, and then it justRuns with macroscope. Just create thefile in macroscope, describe what youwant it to do and what files you want itto check, and it will now run this onevery single code review going forward. Ben's been using this a ton on hisprojects. In particular, he's been usingit for automatic PR labeling, surfacingreviews, correctness checks, and all theusual stuff that you would expect. ThePR labeler in particular is reallyuseful. I hadn't seen this till now.Just showed it to me and I think I'mgoing to go add this to T3 Code. Itwould be super helpful over there. It'scool when I figure out my sponsors areeven better than I thought they werewhen I'm filming the ad because theypaid me to talk about them. They didn'tpay me to do this. See why I like themso much at soy.cope. First, I want to start with Claude Codebecause it's the one that people seemthe most familiar with nowadays. It'swhat people think of now when they thinkAbout coding with AI, it has pretty muchfully taken the mental position in thespace that Cursor used to have. Like Iremember in the YC batches where 90% ofthe batch was using Cursor. Now it'slike 70% are using Cloud Code. The shifthappened very aggressively. So, howshould we be thinking about Cloud Code? The goal of Cloud Code was initiallyBoris's attempt to think into thefuture. What could code look like if themodels were way smarter? How would Iwant to use these things? In a worldwhere people are going to claw.ai,pasting some code, asking the model forthoughts, getting some changes, thencopy-pasting them back into thecodebase, it was obvious that there wasroom to improve. And Cursor had madesome improvements there, especially withlike the chat interface in Cursor whereyou could talk to the model about thecode and watch it go make those changes. Cloud Code was different for a handfulof reasons. The biggest by far beingThat it was built to work in yourterminal. Instead of forcing you toadopt a whole new workflow, instead offorcing you to switch over your IDE tosome other thing, instead of making youinstall a different app or write code ina different way or move everything tothe cloud, cloud code met developerswhere they were, their terminals, andthey wanted to make a great experienceediting code with AI in your terminal. This means they have a lot of inherentcompromises they have to deal with.Things like pasting images will neverquite be elegant. You can't see thething you pasted. Things like selectingtext and all of the weird spacing thatthings have because they kind of haveto. I'm using the new full-screen modeto prevent flicker because the flickerkills me. But the new full-screen modeis doing a whole different way ofrendering to the terminal, which meansthey have to fake all of the UI elementsthemselves instead of using theTraditional scroll buffer behaviors. Hasit benefits, has it negatives, is what Iprefer when I use Cloud Code. They'rehoping to make this the default in thefuture, but yeah, it's meant to just beused in your terminal. I really want toemphasize the adoption story that CloudCode is pushing. They want it to be aseasy as possible to get started usingCloud Code. You paste the command toinstall the CLI, and now you're using it. You do have to go off quickly with yourbrowser or paste an API key or somethingalong those lines. Initially, it was APIkey only, which is still crazy to thinkabout, but the whole idea is that it'seasy to set up and use without having tochange any of your other tools. Youwould spin up Cloud Code, make changes,go look at them in VS Code, commit them,do whatever you want. It would just beanother tab in your terminal. But overtime, as you can guess, people wantedmore. People wanted things likeautomatic git commits and automaticpushes to GitHub and noticing whenchanges happen in your PRs. Peoplewanted to connect Cloud Code to othersources of information through thingslike MCP. People wanted Claude code tobe able to be triggered remotely inorder to control it from their phone oruse it in a better UI, and all thesetypes of things. People wanted CloudCode to be able to do more work at onceand work on things in parallel. Anthropic tried to not overstep in thisway initially, but at some point, Idon't really know what changed. Theystarted to embrace this stuff much more.They had things that would let you buildinto Cloud Code, like their hook system,that kind of stopped being iterated on. Some of my fun demos I built with CloudCode back when I was using it more likethe app that would only let you use yourbrowser and things like Twitter if youhad something running, and whenever youragent finished, it would lock you out ofTwitter until you went and responded to. Get it running again. A lot of thosethings had rough edges because the hooksystem wasn't fully implemented. I wouldsay the moment of no return, like the oh[ \_\_ ] moment for cloud code, wasdefinitely when Opus 45 came out at theend of last year. That jump in modelcapability made it much more compellingto give a model access to everythingthrough the terminal. The idea that Icould GPT through your codebase to findthings, that I could run commands todeploy things, that it could use MCP toget data and do so much more, made it waymore pleasant to use. And I found myselfleaning into Cloud Code way more heavily,because I could trust the code more. SoI didn't need a UI where I was readingit actively. I did want a terminalbecause I wanted it to have access toall the things you can do in a terminal.And the biggest thing you might noticeat the bottom here, bypass permissionson the model got smart enough that Icould trust it to not run malicious. Commands very often, if at all. And now,to this day, I run almost all of thesetools in bypass mode or yolo mode orsomething along those lines. Opus45 wasgood enough to cause a huge industryshift where lots of people startedleaning into cloud code more because themodel caught up with this style of tool. And this was a huge moment for Anthropic,especially in places like Twitter. Theyalso noticed what types of things wouldperform well on Twitter. People loveterminals. People love doing cool thingswith agents. People love when theirtools get better without doing anything.And they love this feeling of the stuffthey rely on and love improving. And Iam going to drop a somewhatcontroversial take here that I’ve saidmany a time. Cloud code is as much amarketing tool as it is a developertool. Anthropic is largely using ClaudeCode to push the idea that Anthropic isthe best thing to use for building withAI. A lot of the features and the coolStuff they add to Cloud Code is reallyoptimized for the Twitter screenshot. Things like the pet mode that they addedwas definitely so people wouldscreenshot their numbers and share themon Twitter. Things like the sub agentmode burns your tokens super hard, butit looks and most importantly feelssuper productive. I remember the firsttime I spun up a bunch of sub agents in ateam. It felt incredible to have so muchstuff happening with so little work onmy part. Cloud code is really optimizedfor that feeling of productivity. Itreally feels almost gified. Littlethings like the cute guy in the cornerhere. The way stuff shows in the UI. Theway that the teams spin up and checkthings out. I'll test an example here.Spin up a sub agent for each exampleproject here and audit it foropportunities to streamline the cloud code experience.Even though I'm in the mode that issupposed to hide my email address, itStill docks my email there. Thanks,Anthropic. We'll try again. See how much is happening in the UIhere. You see the tokens going up as itgoes. It has these little updates. Ithas the loading states. It has the dotsflickering. It feels like it's doing alot. And now it's going to launch theaudit sub aents that it will spin up forthese different things. And we can seeat the bottom here, we can switch tothose sub aents and see what they'redoing. It feels really cool and reallyproductive. There's a lot going on. That's kind of the point. Claude Codewants to feel really productive, but italso, to be frank, they are trying tomake it better and they're trying tomake it more capable, but generallyspeaking, Anthropic's philosophy aroundhow to make the tool do smarter thingsis to let it burn more tokens. Andthings like all the sub aent stuff, andeverybody's already saying this in chat,five minutes later, you're down $200. Token, token, token. You spend two hourstrying to fix this. Yeah, you guys getthe idea. It's very willing to burntokens in the interest of doing thingsin a way that looks and feels reallyproductive, and to be frank, often can be. A lot of the improvements they're tryingto make to cla code are finding new waysto burn tokens to potentially solve workand do more in cooler ways, but also,let's be frank, to look really cool forscreenshots and demos on Twitter. This is also why, in my opinion, Anthropic ismuch less interested in letting youintegrate clatoken maxing so hard that anythingspinning it up programmatically risksabsurd compute usage that they don'twant. Imagine setting up a tool similarto what I just did there where I ask itto spin up sub-agents to check things,but you implement that in your CIprocess whenever a pull request isformed on your project. Then you'regoing to burn a shitload of tokens. might not mean to. And that's a thingAnthropic doesn't want you to do. Ifyou're burning tokens with cloud code,they want it to look and feel productivebecause that's what they're trying tosell. They are eating a bunch of tokenspend to trade off the feeling of beingreally, really productive. And if you'renot looking at the anthropic cloud codeUI, when that happens, it feels muchless productive. They're not gettingwhat they're trying to here. This isalso why things like their desktop appare heavily underinvested in. This is also why things like T3 Code aren't asinteresting to them and why they're notletting us programmatically call theCloud Code CLI anymore without chargingyou way more money. They want you intheir CLI using their features, havingtheir experience and they also want youbuilding integrations into Cloud Codeitself. Because if a tool like T3 codeprovided other skills or provided otherresources, MCPS, etc., and then I canSwitch over to Codex and not have those things get stuck in Cloud Code. Theyactually come over when I switch betweenproviders. Cloud isn't getting the lockin that Anthropic wants it to get. Sothey really optimize for that. CloudCode is meant to be the world's greatestshowcase of what you can do withAnthropic models at a lot of cost toAnthropic in order to win a bunch ofsentiment so people feel and arguablyare more productive using Cloud modelsin Anthropics tools. And you can betyour ass they went hard modifying the UIto make it feel this productive. A lotof the performance issues that used toexist in Claude Code were because theywere doing all these fancy loadingspinners and things, which resulted inthe performance being degraded, which wasa hit they were willing to take becausethey were optimizing so hard for thefeeling. Hey to those in chat saying I'mbeing too nice. I'm trying to be becauseI'm not trying to shame people who likeClaude Code and are using Cloud Code. They engineered it to be like a slotmachine. They did a great job at it. Itgot me addicted to doing agentic coding.I wouldn't have gone as hard as I did atthe end of last year and the start ofthis year if it wasn't for quad code,optimizing itself for this type ofthing. Let's contrast this with codeex. I have it on medium fast, which is wayfaster than quad code. So I'll actuallyturn off fast mode, actually. And now I'm going to paste that same prompt,spin up a sub agent for each example projecthere and audit it for opportunities tostreamline. There is so much less goingon here. You have a timer. You have theworking text that's slightly animated,and nothing else. It did spin up thesetwo subprocesses, but I don't seeanything going on in them. It's veryminimal. This is way less addicting,is the way I would put it. And we sit andwait and wait. It's not trying to feellike a slot machine. It's not trying toFeel crazy. It's just trying to get workdone. And when you look at the types offeatures that are being added acrossthese tools, you will really quickly seethis difference. Just as an example ofone of the recent things they shipped inClaude Code, they added SL radio, whichopens a 24/7 low-fi Claude music stream. Like I'm not going to sit here and saythis is cringe or cool or whatever. Ijust think it shows the philosophy ofwhat and how they are building. Likethey really are optimizing for thefeeling of productivity and the killerthing to post on Twitter. Like you, youknow that in the feature review processat the company, they are thinking aboutwhat does the tweet look like beforethey build the feature. Meanwhile, let'scompare that to the official OpenAIdevelopers Twitter, which is mostly justCodex updates. Now, this tweet did wayfewer numbers. 8.2K likes. Codexanywhere and everywhere all the time.Now, your Mac doesn't need to beUnlocked for Codex to use your computer. This feature is not very flashy. Thisfeature is inside of the settingscomputer use section in Codex. You haveto go hit this switch in there, and thennothing cool happens. You literallycan't screenshot this feature because ithappens when your Mac is locked. But itmakes you way more actually productive.Let's keep scrolling. Ship paper cut.New diff marker settings in appearance. If you prefer the classic plus minus,you can turn that on. Otherwise, it willjust be color-coded. This is not flashyat all. This got 900 likes, but it'sreal and practical. They added a featurewhere Codex can bind a double commandhotkey to copy what you're currentlylooking at in your app and throw it overto Codex to populate the context. Ilove this. I'm a big fan of usingscreenshots as context. And having abuilt-in thing like this is great. Like,it's practical stuff that doesn'tdemocratize because the point of theseThings aren't to be cool demos onTwitter. It's to be frank to solve realproblems that they have as they arebuilding software using these things. Even computer use itself being such acore part of codecs, it didn't make thatmuch sense to me. It seemed like it wasone of those things that sounded reallygood but just wasn't that practical. But similar to how cloud code wasn't usefuluntil the models got good enough,computer use really wasn't useful untilthe models got good enough as well. Andnow that GPT-5 is really smart at thesethings and I also have a dedicated MacMini in the other room that I use for alot of my code tasks, Codex being ableto use computer use to go verify itschanges is awesome. And it's way morelikely that by the time Codex tells meit's done with a task that it actuallydid the task correctly. This is alsowhere another big philosophicaldifference comes in, where they want touse these tools. The vast majority ofPeople at Anthropic using Cloud Code are using it in the CLI. The vast majorityof people using Codeex at OpenAI areusing it through the app. Funny enough,a lot of my friends at OpenAI arenon-technical. They're not devs. They’reon marketing and whatnot. They all useCodeex, the app, and they use it for allsorts of cool things because thisinterface is a big part of how theywork. And it still is optimized to justbe a good experience, not to be a flashyexperience. Like again, I'll do the samedemo I did before. I'm just going totell it to spin up a sub agent in thisproject to audit all of the examples. There is very little going on in thisUI. There is just the thinking here.Now, it’s going to do a search. It’sgoing to go check things. It’s nottrying to be flashy. It’s trying to beproductive. Let me update to the latestcloud app so it’s less likely to break.Oh, great. I have to sign in again.Wonderful. Who would have thought? WhyIs it doing it through [ \_\_ ] Safari? I[ \_\_ ] hate Anthropic, man. This is sobad. I have to manually enter all this[ \_\_ ] cuz I don't have one pass on thelike inapp safari and it doesn't workwith my pass key through one pass. Thisis horrible. Holy [ \_\_ ] You can tellthey don't use it. Like they juststraight up do not use the desktop app. Even better, I'm now in, but it says atthe top right, failed to log in. It mayhave been canceled. They don't usethis app. I don't know how else to sayit. They clearly do not use this app.I'll update to the latest and hopefullythe problems will go away, but it'sprobably just going to reveal my emailagain. There we go. We have chatco-working code. Also, the first thing Inoticed is that all the threads I did inthe CLI aren't here. It doesn't syncacross. The codeex one does. Just anannoying thing. How do I even add a newproject in here? I don't know how to adda new project, guys. Do I need to likeSelect here? Why does it start in the root, not whereI already was? So much more ishappening. You have the fancy animationsfor the text here. You have the tokennumber bouncing up all the time. Thecounter, the super flashy differentshaped star there that stays when it'sdone, by the way, too. So, it's hard toknow when it's actually completed. It'sjust a lot. And it took so much effortto even like get this project added. It’s they didn’t build this app for youto use it. They built this app so thatit would exist if people stop asking forit. Personally, I’m not that into theslot machine feel that Claude Code seemsto be going for. I much prefer simple,reliable solutions you can really trust,like today's sponsor. Got to be so realwith you guys. Hiring sucks right now.The quality of the résumés coming inlooks like it's gone up, but it's alljust slop. And the volume of RS comingin is absurd. Like 10 to 100 times moreThan I've ever seen. Finding goodengineers is really hard. And thenconvincing them to join even harder. That's why G2I has become such an easyrecommendation for me. These guys gethiring. They already have a network ofengineers that they know and trust, andthey know them so well. That's whycompanies like Batteround have moved toG2I for their hiring. They've alreadyhired nine senior engineers, and they'renever going back. The biggest problemthey had was that the leadership at thecompany had to spend so much time hiringbecause finding good people requiresleadership, not a bunch of money spenton recruiters that don't know whatthey're doing. As they said,Batteraround didn't need more resumes.They needed fewer, better candidatesthat they could evaluate quickly andconfidently. If this quote doesn't winyou over, then I don't think G2I is foryou. I think the best part is thenetwork, the vetting process, ratherThan just the pedalling people, so tospeak. It's the biggest difference forme. Before they use G2I, about 50% oftheir hires were success cases, according to them. But since moving toG2I, they'd say the numbers are around90%. Stop wasting time and start hiringquality at soy. G2I. This isn't justabout how it looks and feels to use thething. There's also a very practical howthey're thinking about the models andwhere they're going aspect here. A lotof the bets in the direction thatAnthropic is going in is as the modelsget smarter, they should have theability to do more. Like if the agent issmart enough and it can write its ownprompts for sub agents and just spin upa bunch and let more tokens solve theproblem. Generally speaking, anthropicstrategy is if more tokens solve theproblem, use more tokens. And they arevery quick to look for ways to solveproblems with more tokens. Cloud code ismemory leaks on Windows. Go rewrite bunAnd rust with a shitload of tokens. LikeIt's insane how far they push that ideathat more tokens can solve the problem. OpenAI does not go that way at all.Okay. Also, like, do you see how muchlike the way that this updates is likeline by line, word by word in a way thatfeels very like visually engaging? That's by design. They do all of this tofeel as exciting as possible. But theagain, to the philosophy thing here aboutthe token usage, Anthropic will gladlyburn tokens. Whether it's yours ortheirs or anything else, they are morethan happy to come up with solutionsthat require you burning way more tokensto solve the problem instead of findingbetter ways to solve and verify that theproblem is solved. OpenAI's solution ismuch different in this regard. You cantell just from how many tokens they usefor things like the artificial analysisbench. They are trying really, reallyhard to make their models moreefficient. 54 mini, not the case. 54 mini. Burns tokens. Sonnet 46 burns a similaramount to Mini. GBT54x high and Opus 47burned a similar amount too, but GBT55used half as many tokens to get a betterscore. OpenAI is trying way harder to betoken efficient, but they don't want tocompromise on the accuracy of thesolution. So, they're trying to findways to know if the code worked or notusing fewer tokens. And that's wherestuff like computer use is reallypowerful. If the model can make a changeand then look at what it changed andverify if it succeeded or failed, itends up being a lot less work thanspinning up 15 agents to go triple checkevery line of code. This also means youneed to have an environment that workswell. This is one where the biggestdifferentiator is. Anthropic veryexcitedly sets up tools to let youtransport your thread from your machineto the cloud to finish it up there. Butgood luck getting the cloud environmentfor cloud code to work properly. I tried. And gave up. And the advice I got frommy friends at Anthropic was, you don'treally need the agent to run the code.Just throw the tokens at it. It'llprobably get things done just fine. Butwhen it doesn't, it sucks. And I likeusing the cloud runners for big tasksthat I don't want to run on my machine.So the fact that it is not really ableto verify the results because it doesn'thave the environment configuredcorrectly just kind of sucks. Codeex wastrying harder to get a good cloudenvironment, but it seems like it'sbecome much lower priority. The easiest example to see this is in the Codeexsection of the ChatGPT mobile app. Theyhid the cloud option under like threemenus. They're really not optimizing forcloud because they know how hard it isto get most projects running properly onthe cloud, but you probably have itrunning fine on your machine. They knowthis because that's their own experiencebuilding. They struggled to get everyoneAt the company's projects workingproperly in Codeex on the cloud. So theyjust made it so you can control your ownmachine with your phone instead. Anthropic's bet is the models will getsmart enough they don't even need to runthe code. They'll just write itcorrectly. OpenAI's bet is it's soannoying to configure things properlythat we should probably just use theconfiguration you already have on yourmachine. And now we need to talk aboutthe third option because their play inthe cloud is very interesting. It's kindof crazy how quickly Cursor went fromfar and away first place to third. And Ihonestly think a lot of that's cuzpeople don't know where Cursor's realpower is. You might notice my cursorlooks different from yours. It's cuz mycursor is currently in the browser, notthe IDE. That's because Cursor's cloudstuff is so far ahead of any of thecompetition other than maybe Devon. Crappy headless Linux sandbox that theagent can use to try and execute yourcode. The Cursor sandbox spins up a fullgraphical interface Linux instance thatit can use to run your full app the wayyou would on a real computer, and thenuses computer use to test the changes.So here it ran the T3 code codebase, andthen launched the browser version,because it’s exact same as the versionthat you would run in the electron shellin order to test its changes as it madethem. But Theo, I already have acomputer. Why would I want to do that? Iunderstand. When I’m on my phone, Ididn’t set up a computer to controlremotely. It’s very nice to be able togo spin this off. If I want to work on two different things that need a computeruse at once, it’s incredible to go spinthis off. There are so many use caseswhere one computer is limiting becauseone computer can only do one thing whenyou’re using it. And if it’snot on or not connected to the rightNetwork or any of those other things, andyou can't use it at all, you're justkind of stuck. And there's somethinglegitimately magical about being inSlack. Someone brings up a problem withthe product. You at the cursor bot andsay, "Hey, can you go spin up an agentto fix this?" and it responds in thethread with a video of the fix itapplied, showing you that it worked. This is so far ahead of anything that CloudCode or Codeex can do; it's insane.Another way of thinking about this isthat Codeex is betting really hard onwhere things are today. What is it likeusing agents to code right now? How doyou make it more likely they succeed?Anthropic is betting on where the modelswill be in a few months. Also, where themodels they're using are, because they'reall using Mythos. They are betting onthe models getting so smart that thesetypes of things matter less. Cursor isbetting way out in the future, wherewe're not running these things on ourComputer at all. We are just triggeringthem through whatever other tool we wantand then getting back proof that thething worked. That does bring up anotherreally important difference, and probablythe most painful philosophicaldifference between all of these things. I mentioned this before, but I reallywant to emphasize it. OpenAI uses theCodeex app all of the time. There arethousands of employees at OpenAI thatuse this app every single day. The clouddesktop app does not get anywhere nearas much love, largely because they'reall using the CLI instead. But there's aproblem here. The version and state ofthings that Anthropic employees areusing is very different from what we aregetting as end users. The model that alot of those employees are using isMythos. We don't have that. They have acustom build of Cloud Code with a bunchof cool new hidden features in it thatwe don't have access to. This is why weend up with really embarrassing things. Like system prompt stuff leaking becausethe system prompt that they useinternally is different from the onethat we get in the apps that we use.That's why I ran into problems like thiswhen I tried out Opus 47 in the newcloud code desktop app, where I asked itwhat are some design improvements Icould make on this site, and it openswith a heads-up about the system reminder regardingmalware, which looks like a prompt injection. This is clearly your personal site, theT3G homepage, not malware. Ignoring it.We're using something entirely differentfrom what they're using, and I don'tknow how to properly emphasize this,because people just don't seem to getthat. You do not get the tool Anthropicsemployees get. You get what they feelgenerous enough to give you, and they arenot testing it reliably enough. Thisleaks through the whole clot experience. It is clear that this tool works greatfor them and the portions of the thingthey're using that they cut off. Throw out to us are not tested wellenough, and the result is that we getweird [ \_\_ ] like this. I guarantee youthat one of two, if not both of thesethings is true: either this systemprompt is not the one they were usinginternally, which is very likely thecase, or this problem that I have herejust did not happen with Mythos, whichis what they were all using internally. The model we're using, the system promptwe're using, the harness we're using,the UIs we're using, all of those thingsare not the same thing that they areusing at Anthropic. Meanwhile, the Codexapp that we're using, the Codex modelsthat we're using, and the Codex pluginsand features and all of that that we'reusing, is the exact same thing thatthey're using at OpenAI. I can't cite mysources here, but you need to believe meon this. I have friends at all of thesecompanies. It is very apparently thecase, and you can tell externally thatthis is also the case. It's obvious. To their credit, Cursor is very similarhere. They use the [ \_\_ ] out of thethings that they are doing. Theyobsessively test every detail beforetaking the exact thing they're usinginternally and throwing it at us.They even do crazy stufflike disabling every model other thanthe new composer model and not tellingpeople that they did that. So, everyonethinks they're using Opus or GBT55 orwhatever, but they are dogfooding theirown thing to see what issues people runinto. The effort that Cursor puts intomaking sure things are good before theyput them out is insane. OpenAI is closeto that same level. They just haveissues. If anything, the thing I'veexperienced, which is funny, is thatthey have [ \_\_ ] versions internally atOpenAI. Like I ran into problems withGBT55 during testing because theresearchers didn't have the search toolon because they don't want thingsleaking, which I totally understand. The point being, we get more than theyhave. Cursor gives us exactly what theyhave. Enthropic gives you less than theyhave. That said, Cursor is not reallythat focused on integrating with otherthings. They did just throw out theCursor agents SDK, which is like cooland took them long enough to do it. Anthropic is actively trying to push usaway from using our subs outside oftheir UIs. OpenAI can't stop inventingawesome standards we can build on topof. Like, to be really frank, my agenticcoding app T3 Code that I built withJulius, mostly Julius, this would notexist if it wasn't for the CodeEx appserver, which is part of the open sourceCodeEx CLI. They give us so much stuffto build on top of and around. We arebuilding on top of the CodeC CLI, theexact same way the Codex app team is. Iam admittedly disappointed that the Codex app itself isn't open source. Iwish it was. It's another big part ofwhy we built T3 Codes. It was a goodOpen-source option, but we built on top of the Codex app server, which wasawesome to build with. And the Codexemployees will just openly post onTwitter, like, if you want to use Codexfor something different, tell your agentto go to this repo and look at thisfolder, and it'll figure it out. Cursorseems interested in building these typesof things, but historically hasn't donethat great of a job. When we wanted to add Cursor to T3 code, we had a lot ofproblems because the recommended pathwas through their agent client protocolbindings that they had set up for theCLI, but the Cursor CLI was super behindand the ACP stuff even more so. Youcouldn't even pick modern models. Wetried setting this up a few months ago,not even. And the most modern composermodel was one when two had already beenout for a while. They didn't have Opus46 or 47. They only had 45. They didn'thave any of the new Sonets. They didn’thave any of the new GPT models. It wasSuper limited, and they were allhardcoded in, so we couldn't pick newerbetter models. We had to go back andforth with the team for a while to getthem to fix that. They did end up doingit, and they ended up also putting outthe SDK. So, it seems like they want tomake more of these integration things,but it's not a priority or focus forthem. OpenAI is heavily prioritizingand focusing on this. Claude wouldreally rather you don't do this. Anthropic is trying to keep you frombuilding integrations with cloud code,unless they're deep inside of cloud codeitself. OpenAI wants you to do whateverthe [ \_\_ ] you want. They just want to seepeople build the future using whatever,especially if they're using their stuff.Cursor wants that, but isn'tprioritizing it properly. One last piecethat I think is important to understand,model improvements. Enthropic has to adda bunch of features because their publicmodels haven't improved since December. I still feel like 46 and especially 47are more so regressions than they areimprovements. A lot of why Codex hasimproved is because the models got waybetter. 52 to 53 to 54 to 55 weremassive improvements. In the time wewent from 52 to 55, we went from Opus 45to 47. So, Opus got worse twice andOpenAI models got better three times.That compounds massively. If you tookthe same old version of the codec clithat was around at the time and switchedit over to 55, it would feel incredible. If you took the old version of CloudCode from the end of last year and triedto use Opus 47 in it, it would probablycrash because Cloud Code was getting sobuggy, especially then with theperformance issues. But if you could getit working, it would probably still feelworse because the model was dumber andless willing to do the types of work Iwant to do. So Anthropic has to make upfor this somehow. And they're doing itby doing a shitload of cool lookingFeatures and trying to make momentshappen with cloud code. They're tryingreally hard to seem like they're makinga lot of progress. And to be frank, they're getting away with it. It'sworking. People say things like this onTwitter. The Anthropic team is stillinnovating the most in the harness layer,and it's not even close. The only reasonyou think this is because they do better onTwitter and they keep adding things becausethey kind of have to. But since theiractual better model is Mythos, andthey're not releasing it for variousdifferent reasons that we can conspireabout all day, their tool has to lookbetter because it's not able to improvejust from models getting better. Separately, Codeex is meaningfullyimproving constantly. 55 itself isenough of a reason to make the jump.This is definitely very different fromother videos comparing these tools. Iknow most of them just like send thesame prompt and compare how good theAnswers are from all of them. Whatever. If you like the model, you like themodel. If you like the tool, you likethe tool. The things I care about moreare where is the team going with this?What problems are they trying to solvewith this? And can I trust this thing towork well tomorrow like it does today? Ihope this helps give you a good answerto those things. If you're stillprimarily coding using a sidebar in VSCode or Copilot or Cursor, please go tryany of these, especially the desktopapps. They've gotten a lot better andit's a much nicer way to code using AI. I'm so blown away with how much I likethings like the Codeex app or the T3code experience. It's just, it's reallygood. The models are just one piece ofthis puzzle. And as great as the modelsfor Anthropic might be, all these otherthings matter, too. So, give these toolsa shot. See which ones fit your desiresand your workflows the best and go withthat. If you hate coding, and you want itTo feel more fun as you do it, you wantto feel good and like you're productivethe whole time. Claude Code's actuallyvery good at that. If you need helpstaying motivated, go hop in Cloud Code. If coding is still scary to you, go tryout Cloud Code. It'll probably do agreat job at that. If you've been anengineer for a long time and you want atool that mostly stays out of your way,and just gets [ \_\_ ] done, Codex isphenomenal. It really feels like it'sbuy and for engineers, and I hope itstays that way. Believe me, if Codexregresses, you'll hear it first from me. And finally, with Cursor, I recommendusing it through the cloud side. TheCursor cloud is incredible. It's sopowerful. It does things I never would havepredicted that Cursor could do or thatAI could do as a whole. It's also one ofthe best solutions to set up for yourteam so nontechnical people can kickthings off in Slack. It's really theenterprise-ready end-to-end solution. Highly recommend trying it if you haven't. The cloud stuff is so cool.So, to simplify, cloud code for unmotivateddevs or bad devs that want to feel likethey're productive. Codeex for skepticaldevs that want to use these tools in away that is productive and leans intogood workflows that they're alreadyprobably doing. Cursor for people thatwant to set their teams up for successwith cloud runners that are able tosolve real problems in simple ways usingtools inside of their Slack or whatever. And to answer the final question fromchat, what about anti-gravity? It's agreat way to secure your future byconvincing your boss that AI is notcapable of doing anything at all. And itis not useful for anything beyond that.And if you're asking about PI, you don'tneed this video in the first place. Goplay with the cool fun things. Hopefully, this helps you understand thebenefits of all of these differentsolutions, the philosophy of how theyAre building these different things, andthe reasons why you might pick somethingover another thing. But in the end, yougot to just try all of them and see whatmaps to your workflow the best. And letthe tool shape your flow a bit too. WhenI tried to treat Codeex like Cloud Code,I didn't like it as much. When I let itdo its thing more, I ended up reallyliking it. Hopefully, this has beenhelpful. This is a very strange videofor me to make, but I hope that it is asuseful as I wanted it to be. Let me knowwhat you all think. And until next time,he snirts., Claude Code vs Codex vs Cursor, YouTube, 01ksj2jg159555gzdta8sqqwp3.