Elie Steinbock — oss/acc

OMG I got Qwen 3.5 4b running on my emails now And it's handling them correctly on just 5GB RAM 🤯 using @inboxzero_ai as the harness

@euboid gpt5.4 mini hallucinates less?

@elie2222 I find it hallucinates way too much in internal benchmarks. So much that I can't trust it for almost any use-case 😅

Gemini 3 Flash still a GOAT

@wesbos Need to use xhigh and high thinking on the next test 🫡

Composer 2 vs Opus 4.6 vs GPT 5.4 - a totally unscientific test > Create a Twitter clone. Use Better Auth, Vite, Sqlite, Drizzle, Typescript, and React with Tanstack Start. Each one took ~5 mins in plan mode. Each had access to a browser to test. Composer: 5 mins $6.04 1,250 LOC Opus: 19 mins $10.43 1,000 LOC GPT: 22 mins $14.15 2,000 LOC Opus seems to use the cache WAY more than composer, so it's not really 10x more expensive. Composer app ran first try. Other two needed a bit of CORS debugging but did work. Code between all was extremely similar All three done inside cursor - so Claude Code / Codex may have been different Models used: Composer 2.0 (regular, not fast) Opus 4.6 Medium Thinking GPT 5.4 Medium Thinking

@harriskennyx @inboxzero_ai ya. solid option to use in product. i wouldn't worry too much about exact model. best model for price changes the whole time. so you want to be adaptable. if you use openrouter/vercel ai gateway/ai sdk, it makes it easy to switch to the best option when things change

@elie2222 @inboxzero_ai this is exactly what i was thinking… i have some things i want to use AI for in our product but would be potentially very high volume… this might be it!

@wickedguro @euboid @getsentry Ah okay. So potentially look at Claude SDK. Why you thinking Codex SDK over it?

@elie2222 @euboid @getsentry Nah, I need the agent to find the problems

Has anybody figured out how to do this? - @getsentry issue reported - codex agent spun up with access to sentry + axiom logs & traces - Draft PR auto-created w/ root cause analysis + fix

@harriskennyx @inboxzero_ai You need good quality at a good price point. 3 Flash is very strong for the price. There is a new wave of models like Kimi/Qwen/Minimax that may even be stronger. But privacy concerns are the problem there.

@harriskennyx So for your day to day dev, I wouldn't recommend it. It's fine and cheap. But just use the frontier models for that. But if you have an AI product and you're spending thousands on tokens. eg. processing millions of emails as we do for @inboxzero_ai, then you don't need Opus.

This is massive! I'm yet to be convinced it's as strong as Opus 4.6. But it's strong, and price point is very good. I need to test more to see if it really competes with Opus. But why this is so big: #1 reason people have been moving away from Cursor is the price. Anthropic and OpenAI have been massively subsidizing tokens. Cursor had to resell someone else's model. With this upgrade, Cursor can finally sell their own model. But sell a version that's stronger and cheaper. Long term this is a huge advantage.

@wickedguro @euboid @getsentry Also I feel like if you actually just want to make API calls then you don't need the CLI or SDK either way.

@wickedguro @euboid @getsentry Ah, we weren't talking about customer support above. But simple approach that'll work across Codex/Claude Code is to write a skill, simple CLI it can call. That's about it. Also look at Claude SDK if you haven't already.

@wickedguro @euboid @getsentry Why not just use the CLI direct?

@elie2222 @euboid @getsentry This is interesting. I was thinking of using this @openai/codex-sdk" target="_blank" rel="nofollow noopener">npmjs.com/package/@opena… But, it's CLI-based, so I'm not sure how well I can run it with Docker on a server. This cursor automation stuff sounds good

@concaption thanks!

@elie2222 i watched this

1 of 10 video 😍

Inbox Zero CLI listed on Cursor Directory 😍

@ctatedev @nbaschez Would love to hear any insights there. What are approaches you’ve tested?

@elie2222 @nbaschez I've been thinking a lot about this lately

~100% of my dev is done in sandboxes in the cloud Highly recommend it: - Unlimited parallel agent sessions - My local machine stays safe - Can work from anywhere - Can close laptop - Lap stays cool Interesting idea to visualize with Kanban

@nbaschez @ctatedev Ya third party services has been the real challenge for us :(

@ctatedev Helpful! Thank you

Minimax 2.7 looking strong at $1.20m output 🤯🤯🤯

@janschultecom nice!

@elie2222 Gemini flash is just an amazing model and the only one that was actually useful for cuttlekit.com generative ui

@euboid oh wow 😍

GPT 5.4 mini is an incredible model It's on-par with GPT 5.4 on most of our evals while being 2-4x faster. And at a fraction of the cost. Insane.