I have a product I've been building for 2 years that works with Sign in with ChatGPT and is a polished OSS cross-platform cowork-style desktop app.

I analyzed ~$10K of my own Claude Code usage over 6 weeks. Auto mode cost me 12.7 hours of wall clock time. 27,179 bash commands each sat ~1.7s behind a permission gate. That's an average of 25 minutes a day. I took the data and wrote an allow list. ~50 rules covering read-only git, gh, head, jq, wc, sort, etc. That only would have saved 2.2 of the 12.7 hours. So for ordinary work, I now run bypass mode instead. Frontier models are safe enough and the blast radius is contained to my versioned codebase.

Vibes for Gemini 3.1 Pro Preview in @quests_dev ⚾ elegantly contained 🍄 Mario is broken, but has some nice animations 📐 Linear is spot on 🔮 @theo's image gen MUST be purple 🦩 @simonw pelicans steal the show. motion, detail, and anatomy Wait, isn't Gemini 3 still preview?

Vibes for Sonnet 4.6 in @quests_dev 🏀 Claude has balls on lock 🎮 Mario has power-ups, animations, and game completion state - Linear is Linear 🐥 Pelicans average 💜 theo gen is purple, but at least it's subtle Thx to @OpenRouter for Claude tokens (while it's still legal)

Another model drop, more @quests_dev vibes: MiniMax M2.5 ⚽ contained 🏰 Mario buggy, but cute 〰️ Linear is minimal 🟪 image generation UIs must be purple 🪿 average, but speedy Impressed with tool call accuracy and speed (that might have something to do with @openrouter)

Vibes for GLM 5 🏀 contained, but wiggly 🎮 Mario has animations, hit detection, and crashes when you die ― Linear even has a board view 🟣 Theo's image gen bench is always purple 🦤 Regular pelican is great, california brown is scary @Zai_org cooked! Evals via @quests_dev

Vibe evals for Opus 4.6 in @quests_dev 🏀 mostly contained 🎮 BEST Mario, with animations, state, power-ups, and shell bouncing?! ⬛ BEST clone of Linear 🤷‍♂️ Theo's image bench is purple but masonry 🦩 Pelicans are biking thx to @openrouter for tokens during launch rush!

Vibe evals of Qwen3 Coder Next in @quests_dev 🚫 balls gone wild 🪴 Goombas animate when squished ⚪ Linear aka Redmine 👿 theo's image gen bench is purple hell 🦤 ugly pelicans Impressed with tool calling accuracy on day one from @openrouter

Kimi K2.5 vibe tested in @quests_dev 🍡 balls are _mostly_ contained 🏰 Mario falls to his death 📸 Theo's image gen studio minimal and purple 📐 nailed the aesthetic of the Linear clone Big upgrade for a .5 release. Thx to @openrouter for reliable launch day tokens!

Minimax M2.1 @quests_dev evals ❌ Balls not contained 🪠 Ugly Mario w/ buggy hit detection 🤔 Kanban Linear ≠ Theo's image bench is SoTA... in 2024 🦩 RLed pelicans or best SVG generation ever? Thanks @openrouter for 3rd party inference with capacity to make this possible

GLM 4.7 🫨 evals in @quests_dev! 🏀 balls are contained. 🏰 SUPER PLUMBER has a working mushroom?! 🥵🥶 Linear looks great, but Theo's image gen UI is basic 🦃 Those are not pelicans Does @Zai_org not have Christmas off? 🤔

When users BYOK in @quests_dev, what inference provider or gateway do they prefer? Over 50% choose @openrouter.

Gemini 3 Flash Preview just dropped on @openrouter! Early vibes in @quests_dev 💥 Balls in heptagon slightly off ⊙ Linear clone design shockingly good for a fast model 🛠️ Mario functional, but unusual 🥁 BEST PELICANS EVER?! Did Google just optimize for better pelicans?

🔥off the token-press: @quests_dev evals of GPT 5.2 😈 Purple gradients... remain (but more subtle) 🍡 Best ball physics I've seen 🍄 Mario took 5 minutes, but has functioning power ups, animations, and particles?! 🦩 Chunky AF 🎩 @openrouter unlimited API on launch day

Zero-shotting iOS app UIs from world knowledge? 🎯 Opus 4.5 can do that.

Lil' zero shot black Friday shopping simulation in @quests_dev with Opus 4.5 for fun

🧵 Opus 4.5 zero-shot recreating popular web app UIs by name alone

Early vibe evals of 4.5 Opus in @quests_dev ✅ Balls are contained 🏆 BEST Linear clone to date. How did it do this from training data alone?! 🍄 The CUTEST and most complete version of Mario ever! Thanks to @openrouter for allowing for this much usage minutes after launch

🧰 THIS IS NOT A DRILL: early Gemini 3 vibe coding evals using @quests_dev - Short and precise trajectories, no compile or runtime errors - Best Linear clone I've seen yet (animations too!) - Bouncing falls escaped - Mario has audio So far it's iterative, not revolutionary