Bernardo Ferrari

@davidgomes @midego1 @cursor_ai on Codex it is 30% of the users, so not majority

@midego1 @cursor_ai That's good feedback! I personally think most people will prefer Fast. It's a great tradeoff.

Composer 2 is now available in Cursor.

@JorgeCastilloPr my feeling is that android tooling is awful and no one is doing it. Gradle outputs 1000 lines for every build, even simple errors get 300 lines. There is no way to use vscode instead of android studio. LLM will get imports wrong 99% of the time and there is no 'auto-fix' script.

This is my feeling too. I know how much work has been put into making Android Studio great, and I really appreciate it. But my reality is that I don’t find the need to open it almost for anything anymore.

@peteralexbizjak if you have complaints, invoke @sethladd if you have compliments, invoke @LoicSharma

I cannot find Dart (or Flutter) in the OpenTelemetry Registry. Somebody should look into this. In my opinion, we urgently need an official (or at least community) solution.

@mkurman88 Precisely, the credits based plan is not worth it, but the personal is.

@bernaferrari This week, I used 903 credits starting from March 17, and I have used 57% of my weekly quota :) BUSINESS PLAN sucks

GitHub Copilot offers the best value for the price. I upgraded to Copilot Pro+. I was skeptical at first, but now I can do much more for $10 than I could with the entire Codex business plan. > It includes 300 premium requests, which is almost a one-shot for most long-run tasks, and GPT 5.4 is just 1x (though I think it's 0.9); > tool calling is included - unlike Codex, which burns through business credits by the hour, you don't waste credits on tools. Yesterday, I ran two prompts and used nearly 225 credits lol.

@SkylerMiao7 @DavidSZD1 You can do 3x bigger, then a flash variant that still beats 2.7

@DavidSZD1 we’ll see. we’re definitely pushing it.

M3 is scaling up.

@niclas1fang @MiniMax_AI Haiku is still 5x more expensive than it

@bernaferrari @MiniMax_AI I'm not a hater. Actually I immediately subscribed token plan after release of m2.7. I just hope that Chinese models are not benchmaxed. Here are shopping websites m2.7 and haiku 4.5 built: …b69-7684-8c38-a76926a652b4.arena.site …b69-77a6-9e19-fe27b520692e.arena.site I mean, m2.7 isn't far better than haiku.

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents to 3-min on certain occasions. - Advanced Agentic Abilities: Trained for Agent Teams and tool search tool, with 97% skill adherence across 40+ complex skills. M2.7 is on par with Sonnet 4.6 in OpenClaw. - Professional Workspace: SOTA in professional knowledge, supports multi-turn, high-fidelity Office file editing. MiniMax Agent: agent.minimax.io API: platform.minimax.io Token Plan: platform.minimax.io/subscribe/toke…

@niclas1fang @MiniMax_AI 3.1-pro level at 10x cheaper

@MiniMax_AI Not even close totop models...

@ivanfioravanti yes, me too, but that's the problem, if there is any instability they handle super super poorly, where claude has a backoff retry policy which is much better.. Like any error droid will just stop, where claude will retry like 10 times.

@bernaferrari But I'm using it with BYOK without paying anything to them 👀

I have never used in deep Droid, but I'm really impressed by just first experiments 🤯 Are some heavy users out there that used other coding harnesses too? Is it really so much better? I will start using it more and more!

@ivanfioravanti then I just use codex on top to review/polish. But that's the thing, with Droid you have two options: pay their super premium or suffer. The output will be good but will fail half the time. With Claude it will just work, you can go sleep and ask it to work.

@bernaferrari But looking at end results it seems Claude tends to care less about details of the output generated.

@rohanvarma I use codex app for not work, for things that are going for weeks or months so it is easier to keep track. Things I don't even know where the directory is anymore. But codex app is so slow. Freezes all the time. So I still like the CLI.

Some interesting data we pulled today showed that ~40% of Codex users use multiple surfaces, between the App, CLI, and IDE extensions. Everyone seems to have a primary preference, but a bigger-than-expected chunk of users launch codex agents outside of their primary interface. If you use multiple surfaces, I'm curious why? And what could we do to improve the experience?

@albfresco @thsottiaux 5.3 because it is free, 5.4 mini is still expensive. 5.4 is 30% more expensive than 5.3 and 5.4-mini 1/3. So it is basically 50% cheaper than what we had 1 week ago, which is not enough for a mini model.

@thsottiaux with every good decision comes an equally difficult next decision am I supposed to spawn 5.3-sparks or 5.4-minis

Yesterday we launched subagents in Codex. Today we released GPT-5.4-Mini, which is SoTA in its category. Coincidence or genius move?

@_avdept @GregorySchier Pricing is mostly the same, and they were generous because M4 had a limit of 48gb I think. Now it is 64gb. But there is still a massive price jump to m5 max.

@bernaferrari @GregorySchier wow crazy i bought m3 in 2024 and havent checked on pricing for newer models

Should I upgrade my M1 Pro MacBook Pro (32GB) to M5 Pro (64 GB)? I write Rust and Go so the CPU/RAM upgrade *should* be meaningful, but looking to hear from people who actually made the switch.

@GregorySchier not necessarily because m6 will be a full redesign, you don't see people wanting to buy the last pre-m1 macbook

@bernaferrari Trade in will be pretty good on the M5 then. There's always one coming that will be better

@_avdept @GregorySchier the issue is that >64gb you need to go to m5 max which is $1000 more expensive, so not worth it. It is a jump from $3000 to $4500

@GregorySchier go for 96gig if possible this way you can also run some LLMs

@ivanfioravanti My favorite one is Kimi. I hope 2.7 matches it.

Testing MiniMax M2.7! 🔥🔥🔥 I have more time now!

@jaredpalmer you should add a preview image to github repos, it is too ugly not hot having images in my profile repos.