Tokens & Signals for 4/29/2026. We scanned ~1,200 Twitter accounts (1213 tweets), 13 subreddits (65 posts), Hacker News (12 stories), 7 newsletter posts, 5 podcast episodes, 125 Discord messages, and leaderboard data for you. Estimated reading time saved: ~12 hours.
* GPT-5.5 is here: OpenAI's latest model, codenamed "Spud," is a full ground-up rebuild — 10T parameters, native omnimodality, and it's already topping benchmarks like Terminal-Bench 2.0 and GDPval while somehow matching GPT-5.4 on latency. developers.openai.com/api/docs/guides/prompt-gu...
* Mistral Medium 3.5 drops: A dense 128B model that rolls instruction-following, reasoning, and coding into one unified architecture with a 256k context window. No MoE tricks — just one big, capable set of weights. huggingface.co/mistralai/Mistral-Medium-3.5-128B
* Google's Cloud surge: Revenue hit $20B in Q1 2026, up 63% YoY, driven by enterprise AI infrastructure demand and Gemini adoption. The spending is finally paying off. x.com/sundarpichai/status/2049581838260461916
* Stripe's new agent tool: The Link CLI gives AI agents a secure way to make purchases — think of it as a credit card for your agent, with built-in authentication and automatic refunds. github.com/stripe/link-cli
* Figure AI's manufacturing leap: One humanoid robot per hour. That's a 24x throughput improvement in just 120 days. The hardware bottleneck is cracking. x.com/adcock_brett/status/2049514372264055116
* Alibaba's speed boost: FlashQLA kernels for linear attention are delivering 2–3x faster forward passes for edge-side agents.
* Claude Code instability: Users are hitting persistent 502 Bad Gateway errors during terminal tasks, with speculation pointing to load issues on Anthropic's proxy. Not great when you're mid-session on something important. news.ycombinator.com/item?id=47952722
* DeepSeek's price war: A 75% discount on the V4-Pro API, extended through May 31, 2026. Western frontier model pricing is looking increasingly awkward by comparison. x.com/deepseek_ai/status/2049312932014813344
* Musk v. OpenAI trial: Internal documents and diary entries are now in front of a judge, laying bare OpenAI's early governance debates and its drift from non-profit mission to commercial enterprise. x.com/ns123abc/status/2049527702076449244
* @karpathy on model scaling: "Parameter counts used to be the only thing people cared about, but now we're fighting over 'reasoning effort' and 'test-time compute'—basically giving models more time to think before they speak."
Best to Build With Today
* Coding — claude-opus-4-7-thinking-auto is still the top pick for complex reasoning and large codebases.
* Reasoning — gemini-3.1-pro leads Chatbot Arena for math and deep logical deduction.
* Chat — gemini-3.1-pro is also the highest-rated overall assistant right now.
* Open-source — Mistral Medium 3.5 (128B) is the strongest new option for dense, self-hosted deployment.
* Value pick — DeepSeek V4-Pro (75% off until May 31, 2026).
Deeper Dives
💼 Industry & Business
* Google Q1 2026 Earnings: Google Cloud jumped 63% year-over-year, fueled by enterprise AI infrastructure demand and Vertex AI. Sundar Pichai dropped a number worth sitting with: Gemini is now processing over 16 billion tokens per minute through direct API use.
* Why it matters: Big Tech has been pouring money into AI infrastructure for years. This is the moment that spending starts visibly turning into revenue.
� Twitter� Hacker News
* Figure AI Production Ramp: One humanoid robot per hour, up from essentially nothing — a 24x throughput increase in 120 days.
* Why it matters: The hard problem in robotics right now isn't software. It's manufacturing. Figure is actually solving it.
� Twitter� Reddit
* Musk v. OpenAI Trial: Court disclosures are surfacing internal documents and early governance debates, showing in detail how OpenAI shifted from a non-profit structure toward a commercial one. It's messy, public, and not going away.
* Why it matters: This is a rare, unfiltered look at how one of the most influential AI labs in the world actually got built — and the legal questions it's raising about AI governance are serious.
� Twitter
🧠 Models & Research
* OpenAI GPT-5.5: Natively omnimodal, built for long-context reasoning, with an estimated 10 trillion parameters under the hood. MoE routing keeps it efficient despite the scale.
* Why it matters: A truly unified omnimodal architecture raises the ceiling on what's possible for complex, long-horizon tasks.
� Twitter
* Mistral Medium 3.5: 128B dense model, one set of weights, 256k context window — and it handles instruction-following, reasoning, and coding without needing to switch modes.
* Why it matters: Dense models are more predictable than MoE for long-running agentic tasks. This is a serious option for engineers who need reliability over raw benchmark flash.
� Twitter� Reddit� Hacker News� Discord
🚀 Products & Launches
* Stripe Link CLI: Lets AI agents make purchases using one-time-use virtual credentials — so they can handle e-commerce operations without ever touching your real financial details.
* Why it matters: Autonomous agents that can spend money responsibly is a hard problem. This is a genuinely thoughtful first step toward solving it safely.
� Twitter
* Alibaba FlashQLA: A high-performance kernel library for linear attention, delivering 2–3x speedups on forward passes and cutting the overhead for long-context inference significantly.
* Why it matters: Faster kernels mean agents can do more thinking on cheaper hardware. That's a big deal for anyone running inference at the edge.
� Twitter� Reddit
🔥 Takes & Drama
* Claude Code Instability: Engineering teams are running into consistent 502 errors and session timeouts during heavy terminal use of Claude Code.
* Why it matters: Benchmark scores don't mean much if the service drops out mid-task. Reliability is quietly becoming the most important feature in agentic tooling.
� Twitter� Hacker News
Launches
* Mistral Medium 3.5 — Dense 128B flagship with a 256k context window and configurable reasoning effort.
* Stripe Link CLI — Secure, agent-friendly command line interface for autonomous payments.
* PyTorch AutoSP — New compiler tool for automatically optimizing LLM training for massive context windows.
* Zed Editor 1.0 — Production-ready, high-performance code editor built for agent-integrated workflows.
Closing thought: The conversation has shifted. It's no longer about who has the biggest model — it's about reasoning efficiency, infrastructure throughput, and whether your service actually stays up. Today's news makes it pretty clear: we're past the model release era and into the service delivery era, where reliability and cost matter just as much as whatever's topping the leaderboard this week.