Tokens & Signals for 4/30/2026. We scanned ~1,200 Twitter accounts (1299 tweets), 13 subreddits (54 posts), Hacker News (12 stories), 8 newsletter posts, 4 podcast episodes, 157 Discord messages, and leaderboard data for you. Estimated reading time saved: ~12 hours.
* OpenAI's GPT-5.5-Cyber is here: A specialized model for defensive security, hitting a 71.4% pass rate on red-team benchmarks and solving multi-step attacks in under 11 minutes. x.com/sama/status/2049712078836170843
* The $130B Capex Shock: Amazon, Google, Microsoft, and Meta dropped over $130B on AI infrastructure in Q1 2026 — a 71% YoY jump that's basically funding a private Manhattan Project every single quarter. x.com/GaryMarcus/status/2049612941419565456
* Speeding up agent loops: OpenAI added WebSocket support to the Responses API, slashing interaction latency from ~300ms to under 80ms. Real-time agents just got a lot more real. x.com/OpenAIDevs/status/2049595890395152728
* Claude-Jupiter incoming: Spotted in Workbench, this upcoming model is showing a 14% boost on GPQA diamond and a 20% drop in hallucinations. Worth watching. x.com/testingcatalog/status/2049896926997369194
* DeepSeek's visual thinking: A new paper introduces "Visual Primitives" — geometric tokens that improve spatial reasoning by 28% without needing a mountain of compute. x.com/testingcatalog/status/2049896926997369194 k-a.in/Thinking_with_Visual_Primitives.pdf
* Stripe's agent-first payments: A new billing protocol lets autonomous agents manage their own subscriptions and payments. No human approval required. x.com/stripe/status/2049889788585795895
* Samsung's HBM profit surge: The memory giant posted a record ~$27B operating profit, almost entirely on the back of gold-rush demand for AI training memory. x.com/jukan05/status/2049637296136491214
* Warp goes open source: The terminal-turned-ADE (Agent-native Development Environment) released its full stack and hit 37k stars overnight. x.com/warpdotdev/status/2049611416219627898
* @theo on the "OpenClaw" drama: "Anthropic is actively punishing users for mentioning specific protocols in their commit logs, which is a wild look for a dev tool." x.com/theo/status/2049645973350363168
@karpathy on the new age of engineering: "We are moving from outsourcing understanding to outsourcing thinking* — and the agents are getting scarily good at 'alien science' that humans can't even predict." x.com/aerockrose/status/2049839995452534820
Best to Build With Today
* Coding — claude-opus-4-7 — Still sitting at the top of the ELO charts for complex, multi-step repo engineering.
* Reasoning — gemini-3.1-pro — The current leader for math and hard logical deduction.
* Chat — gemini-3.1-pro — Most reliable for general day-to-day tasks. Boring answer, but it's the right one.
* Open-source — Qwen-3.6-27B — Punches well above its weight for cost-effective, local agent deployments.
* Value pick — gpt-5.5-high — The sweet spot for performance vs. cost in agentic workflows.
Deeper Dives
💼 Industry & Business
Big Tech Q1 capex hits $130B+
Amazon, Google, Microsoft, and Meta collectively spent $130.65B on infrastructure in Q1 2026 — a 71% jump year-over-year. The money is flowing into hyperscale data centers, custom AI silicon, and energy capacity.
Why it matters: That's 3x the cost of the Manhattan Project, every quarter. The AI arms race isn't slowing down — it's compounding.
� Twitter
Google Cloud reports 63% revenue growth
Alphabet's cloud division hit $20B in revenue, with generative AI products growing 800% YoY. Healthcare and finance are seeing record deal sizes.
Why it matters: It shows the massive infrastructure spend is actually converting into real enterprise revenue — not just hype.
� Twitter
🧠 Models & Research
OpenAI rolls out GPT-5.5-Cyber
Tuned specifically for defensive security, this model hit a 71.4% pass rate in AISI red-teaming tests and achieved privilege escalation in 78% of simulated containerized environments.
Why it matters: It's the first model optimized for autonomous penetration testing — which fundamentally changes how we think about defending against AI-augmented threats.
� Twitter� Reddit
Anthropic tests 'Claude-Jupiter'
Leaked via Workbench, this model is targeting ultra-long context windows and tighter formal logic. Early benchmarks show a 14% improvement over Opus on GPQA diamond.
Why it matters: Anthropic is keeping the pressure on OpenAI by doubling down on reasoning reliability rather than just raw capability.
� Twitter
DeepSeek releases 'Thinking with Visual Primitives'
By breaking visual input into geometric tokens, DeepSeek achieved a 28% performance lift on complex spatial tasks — an efficient way to make models "point" while they "think."
Why it matters: This tackles the "Reference Gap" head-on, letting models actually understand UI and charts instead of hallucinating their way through them.
� Twitter� Reddit
🚀 Products & Launches
OpenAI speeds up agentic workflows with WebSockets
Bidirectional communication in the Responses API cut latency from 300ms to 80ms — a seemingly small number that makes a huge practical difference.
Why it matters: Latency has been the biggest thing holding agents back from production. This removes a real blocker.
� Twitter
Stripe launches AI Agent Billing Protocol
Stripe's new suite lets AI agents independently handle payments and identity verification, with granular authorization limits for developers to set.
Why it matters: Agents can't truly work for us if they can't pay for their own resources. This is the missing financial rail for machine-to-machine commerce.
� Twitter
🔥 Takes & Drama
Anthropic's 'OpenClaw' moderation fallout
Prompts containing the keyword "OpenClaw" are apparently triggering automated rate-limits or session terminations. The community is split on whether this is a legitimate security measure or Anthropic trying to gatekeep the agent ecosystem.
Why it matters: It surfaces a real and growing tension between platforms protecting themselves and developers trying to build independent agent tooling on top of them.
� Twitter� Hacker News
Launches
* Qwen-Scope — Sparse autoencoders from Alibaba that let developers steer internal model features directly. x.com/Alibaba_Qwen/status/2049861145574690992
* Warp Terminal — Now fully open source and open for community-contributed agent skills. x.com/warpdotdev/status/2049611416219627898
Closing thought: When agents are paying their own bills and doing "alien science" that humans can barely follow, the line between building a tool and building an autonomous system isn't blurring — it's gone.