Get live statistics and analysis of Sumeet 🎒's profile on X / Twitter

Today’s Cursor meetup Mumbai. Was awesome experience.

Been using Gemini 3 Pro since few hours, one thing fs it excels at is FrontEnd Was Gemini 3 Pro worth the hype? Which area does it excel at and what can be improved? How’s Antigravity compared to cursor/windsurf? Drop your experiences and Treat this as Google feedback 👇

Heads up, fellow builders 👀 Anthropic just dropped bonus Claude Code credits: 💳 Pro: $250 💳 Max: $1,000 Valid Nov 4–18, on web, separate from your usual limits. Let's build without burning our main quota. Love seeing these boosts for builders. Don’t sleep on it 🔥

Adaptive reasoning in 5.1 Instant is pretty smart - it figures out on the fly when it needs to think harder. Honestly, the instruction-following improvements are what I’m most excited about since it means less time fighting with prompts in production. That said, there are a few things I’m keeping an eye on: The speed variance is real - 5.1 Thinking is twice as fast on easy stuff but twice as slow on hard problems. If you’ve got strict SLA requirements, that unpredictability can be rough. How to handle it: Wrap your calls with timeouts and fall back to Instant mode if things are taking too long. For stuff that needs to be snappy, explicitly use gpt-5.1-chat-latest in the API. For complex tasks where speed doesn’t matter as much, queue them async and let the model take its time. Context windows are still limited - We’re stuck with the same 8K/32K/128K limits in ChatGPT (though the API gets 400K). If you’re doing RAG or analyzing multiple docs, you’ll hit the ceiling fast. How to handle it: Just use the API directly for anything heavy. Break large docs into chunks with some overlap, or go full RAG with vector embeddings so you’re only loading what’s actually relevant. If you’re enterprise, bug Azure for higher quotas. The router is still a mystery - The autoswitcher that picks between Instant and Thinking is a black box. When it breaks (and it has), everything suddenly feels way dumber. How to handle it: Don’t trust Auto mode for anything important. Pick your model explicitly in the API - use gpt-5.1-chat-latest for speed or gpt-5.1 when you need reasoning. Log everything so you can spot when things go sideways, and A/B test your routing logic. It’s rolling out slowly - Paid users first, free tier later. If you’re shipping enterprise features, you’re dealing with users on different model versions for weeks. How to handle it: Feature-flag everything. Build in graceful fallbacks to GPT-5 or 4o. Check what model is actually available before each request and route accordingly. Be upfront with users about what requires which access level. The tech is legitimately impressive, but the real skill is building around these limitations instead of just hoping they get fixed.

Do you think Gemini 3.0 will beat Claude & Chatgpt?

Today’s Cursor meetup Mumbai. Was awesome experience.

Heads up, fellow builders 👀 Anthropic just dropped bonus Claude Code credits: 💳 Pro: $250 💳 Max: $1,000 Valid Nov 4–18, on web, separate from your usual limits. Let's build without burning our main quota. Love seeing these boosts for builders. Don’t sleep on it 🔥

Adaptive reasoning in 5.1 Instant is pretty smart - it figures out on the fly when it needs to think harder. Honestly, the instruction-following improvements are what I’m most excited about since it means less time fighting with prompts in production. That said, there are a few things I’m keeping an eye on: The speed variance is real - 5.1 Thinking is twice as fast on easy stuff but twice as slow on hard problems. If you’ve got strict SLA requirements, that unpredictability can be rough. How to handle it: Wrap your calls with timeouts and fall back to Instant mode if things are taking too long. For stuff that needs to be snappy, explicitly use gpt-5.1-chat-latest in the API. For complex tasks where speed doesn’t matter as much, queue them async and let the model take its time. Context windows are still limited - We’re stuck with the same 8K/32K/128K limits in ChatGPT (though the API gets 400K). If you’re doing RAG or analyzing multiple docs, you’ll hit the ceiling fast. How to handle it: Just use the API directly for anything heavy. Break large docs into chunks with some overlap, or go full RAG with vector embeddings so you’re only loading what’s actually relevant. If you’re enterprise, bug Azure for higher quotas. The router is still a mystery - The autoswitcher that picks between Instant and Thinking is a black box. When it breaks (and it has), everything suddenly feels way dumber. How to handle it: Don’t trust Auto mode for anything important. Pick your model explicitly in the API - use gpt-5.1-chat-latest for speed or gpt-5.1 when you need reasoning. Log everything so you can spot when things go sideways, and A/B test your routing logic. It’s rolling out slowly - Paid users first, free tier later. If you’re shipping enterprise features, you’re dealing with users on different model versions for weeks. How to handle it: Feature-flag everything. Build in graceful fallbacks to GPT-5 or 4o. Check what model is actually available before each request and route accordingly. Be upfront with users about what requires which access level. The tech is legitimately impressive, but the real skill is building around these limitations instead of just hoping they get fixed.

Do you think Gemini 3.0 will beat Claude & Chatgpt?

Been using Gemini 3 Pro since few hours, one thing fs it excels at is FrontEnd Was Gemini 3 Pro worth the hype? Which area does it excel at and what can be improved? How’s Antigravity compared to cursor/windsurf? Drop your experiences and Treat this as Google feedback 👇