Darshan Tank

@hwchase17 Research direction: > Extracting useful signals from that data for eval/environment generation, harness engineering, and post-training

@trq212 This is really good approach to understand anything new. HTML gives better visualisation!

Now rate limit looks better @claudeai

@comet has better voice AI mode than @googlechrome

30$/M input tokens and $180/M output tokens!! On what basis? on the name of benchmark numbers! Many questions: - Need of further challenging benchmarks? - Dependence shift to OS models? - Normalisation of using model locally?

@theo Price is a relative parameter! The day Anthropic makes it cheaper, it will also go down, but this will never happen and it will rise from here onwards on the name of model hype or benchmark numbers.

$5 per mil in, $30 per mil out. GPT-5.5 is smart. I've been using it for a bit. It's also weird, hard to wrangle, and too expensive IMO. Double the price of GPT-5.4. 20% more expensive than Opus 4.7.

@OpenAI Are people still using chatgpt!

Introducing GPT-5.5 A new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion. It marks a new way of getting computer work done. Now available in ChatGPT and Codex.

Coverage, Complexity and Quality!

@SpaceX @cursor_ai Full banger 🔥

SpaceXAI and @cursor_ai are now working closely together to create the world’s best coding and knowledge work AI. The combination of Cursor’s leading product and distribution to expert software engineers with SpaceX’s million H100 equivalent Colossus training supercomputer will allow us to build the world’s most useful models. Cursor has also given SpaceX the right to acquire Cursor later this year for $60 billion or pay $10 billion for our work together.

@bcherny @stolinski It would be interesting to know, if you guys analyse each session by clustering users- you can kill any growing startup/company 👀

@stolinski That's weird. Could you run /feedback and share the id here? Would love to debug

@alexalbert__ It’s strange! You are showcasing what is realistically possible with Mythos and selling something low!

Some of my favorite things in Opus 4.7: - Very good at async work and following instructions - Effort levels are far more predictable for token control (+ new xhigh level) - No more downscaling of high-res images - Noticeably more taste in UIs, slides, docs

@trq212 Context rot is a real problem it hallucinates- miss or hit problem

@trq212 rewind and subagent are something that I've not used that much will try to adapt it!

@aakashgupta If they think of building one time infrastructure and use OS models that would be game changer! 1. It will save cost! 2. Your data stay within your system only!

Uber gave 5,000 engineers access to Claude Code in December. By February, usage had nearly doubled. By April, the CTO told the company they'd burned through the entire annual AI budget. The adoption curve tells you everything about what happened. In December 2024, 32% of Uber's engineers were using Claude Code. By February 2026, that number was 63%. That's not a gradual rollout. That's a product so useful that engineers pulled it into their workflow faster than finance could model the spend. Uber has about 34,000 employees. Engineering is roughly 15% of that headcount, somewhere around 5,100 people. At enterprise API pricing, Claude Code runs $100 to $200 per developer per month on Sonnet alone. But that's the subscription math. The real number is token consumption, and Uber's engineers aren't building hello-world apps. They're building rider-driver matching algorithms, dynamic pricing engines, and real-time logistics across 70+ countries. Every one of those tasks eats context windows for breakfast. The scale of what these engineers are actually doing with AI is wild. 92% of Uber's developers use AI agents monthly. 65 to 72% of code written inside IDEs is now AI-generated. 11% of all pull requests are opened by agents, not humans. The company's AI code review system, uReview, analyzes over 90% of the 65,000 diffs Uber ships per week. AI-related costs at Uber are up 6x since 2024. CTO Praveen Neppalli Naga's quote was "I'm back to the drawing board." That's the CTO of a $144 billion company admitting that the tools work so well his team can't afford to keep using them at this rate. Here's the part nobody is pricing in. Anthropic's Claude Code hit $2.5 billion in annualized revenue by February 2026. That's up from $1 billion in November 2025. The fastest enterprise software ramp in history, and a huge portion of that growth is coming from exactly this pattern: companies deploy Claude Code, engineers love it, usage explodes, budgets evaporate. Uber won't be the last company to have this conversation. The average Claude Code developer burns about $6 per day. Multiply that across thousands of engineers running complex agentic workflows, spawning sub-agents that each maintain their own context windows, and the math compounds fast. One engineering team running Claude Code in automated CI/CD loops can drain a monthly budget in days. The CFO problem is now the bottleneck for AI adoption at the enterprise level. The technology works. The productivity gains are real. Uber's own data says 75% of AI code review comments are marked helpful by engineers. The constraint is that traditional annual budgeting was designed for tools with predictable per-seat costs, and AI coding agents have usage curves that look like cloud compute bills from 2015: exponential until someone notices. Every enterprise CTO is about to have the same meeting Praveen just had. The tools are too good to pull back. The costs are too unpredictable to ignore. And the companies that figure out token cost optimization first will have a structural advantage over every competitor still running annual budget cycles against exponential adoption curves.

@aakashgupta We have best open source model like MiniMax M2.7, GLM 5.1 with fraction of cost to opus4.6 but still people are falling for claude code! It is not about LLM anymore it is about agent and it's harness

@BeingSalmanKhan @realpreityzinta Bhai finally replied🥹

Well done Zinta,Congratulations Zinta, Team is playing well …. @realpreityzinta

> use claude code for heavy lifting > cursor as reviewer and to make minor changes observation: cursor is quicker as they are have done indexing

@TheAhmadOsman Opencode is really cool

running Qwen3.5 397B MoE (17B active/token) on 4x DGX Sparks in FP8 (~400GB) > OpenCode driving > agent exploring its own config > probing all 4 Sparks (via ssh) + reporting thermals > inspecting how vLLM is serving it > collecting + analyzing its own stats local AI is awesome

@opencode is as good as claude code with lower cost!