Tokens & Signals for 7/30/2026. We scanned ~1,200 Twitter accounts (1246 tweets), 13 subreddits (56 posts), Hacker News (10 stories), 6 newsletter posts, 4 podcast episodes, 113 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* OpenAI just nuked their prices — Luna is 80% cheaper and Terra is down 20%. They also dropped a 'Fast' mode for Sol that's 2.5x quicker. x.com/zephyr_z9/status/2082876360499007544
* GPT-5.6 Sol is now rewriting its own GPU kernels to shave 20% off serving costs. The model is literally optimizing its own infrastructure. x.com/zephyr_z9/status/2082882939365519701
* ChatGPT just hit 1 billion weekly active users — the fastest any app has ever reached that scale. x.com/MTSlive/status/2082619769417760803
* Leopold Aschenbrenner's AI hedge fund is reportedly collapsing after leveraged bets on compute futures went badly wrong. x.com/tbpn/status/2082914062745399383
* Thinking Machines dropped Inkling-Small, a 276B parameter model that's punchy and built for edge use. x.com/miramurati/status/2082886925988598138
* Google DeepMind's Gemini Robotics 2 brings "whole-body" movement to robots, letting them handle complex physical tasks with something close to human dexterity. x.com/GoogleAI/status/2082844740446253125
* @nicdunz on OpenAI's ARC-AGI-3 scores: "Turns out you don't always need more training; sometimes you just need to fix your testing harness and use recursive reflection." x.com/nicdunz/status/2082884002201878824
* @karpathy on self-improving models: "First it was humans writing code. Then AI assistants writing code. Now models rewriting their own infrastructure. Timeline compressing fast."
* Samsung just posted a wild 1,813% profit surge, proving the AI infrastructure boom is paying off massively for hardware suppliers. x.com/jukan05/status/2082615141355712819
* DeepSeek is building a 1GW datacenter in Mongolia and using the freezing 4°C climate to keep their massive compute clusters cool. x.com/ns123abc/status/2082912664762618239
Best to Build With Today
* Coding — claude-opus-4-6-thinking-32k (Still the king of the arena for complex codebase navigation).
* Reasoning — gemini-3.1-pro (Currently topping the charts for high-level math and complex logic).
* Chat — gemini-3.1-pro (The overall leader for day-to-day assistance and creative tasks).
* Open-source — Inkling-Small (The new efficiency champion for high-performance edge deployment).
* Value pick — GPT-5.6 Luna (An 80% price cut makes it the obvious budget-friendly choice right now).
Deeper Dives
💼 Industry & Business
OpenAI Slashes GPT-5.6 Model Prices
OpenAI has overhauled its API pricing — standard input token costs are down 65% and output tokens are down 80%. They also launched a "fast" mode for GPT-5.6 Sol with a 2.5x speed boost, aimed squarely at making frontier intelligence cheaper for enterprise workflows.
Why it matters: This is the latest push in the race to make AI "too cheap to meter" while locking in enterprise customers.
� Twitter� Reddit� Hacker News
Leopold Aschenbrenner Hedge Fund Faces Collapse
The "Situational Awareness" fund is facing forced liquidation after aggressive, highly leveraged bets on compute futures went south. Citadel and other firms are reportedly circling the assets as the fund scrambles to manage the fallout from its volatility-heavy strategy.
Why it matters: A harsh reminder that the "AI infrastructure trade" is incredibly risky and prone to massive swings.
� Twitter� Reddit
Scale AI Appoints Francis deSouza as CEO
Alexandr Wang is shifting to a stewardship role to focus on core infrastructure, bringing in the former Illumina CEO to lead Scale into its next phase of enterprise and government scaling.
Why it matters: Scale sits at the center of the AI data pipeline, and this move signals a shift from scrappy startup to serious infrastructure provider.
� Twitter
Samsung Electronics Reports 1,813% Profit Surge
Samsung's Q2 earnings show a 1,813% profit jump driven by insatiable demand for AI memory chips (HBM).
Why it matters: The AI buildout is printing money for hardware suppliers, even as the model builders keep burning cash.
� Twitter
🧠 Models & Research
GPT-5.6 Sol Optimizes Its Own Infrastructure
OpenAI is using GPT-5.6 Sol to autonomously rewrite its own production GPU kernels and refine speculative decoding. The result: a 20% reduction in serving costs and a 40% drop in latency.
Why it matters: We're entering an era where AI agents improve their own underlying efficiency — a compounding cycle that's going to be hard to slow down.
� Twitter� Reddit
OpenAI ARC-AGI-3 Score Triples via Settings
OpenAI showed that GPT-5.6 Sol can triple its ARC-AGI-3 benchmark score — from 22% to 66% — just by applying "recursive reflection" and "iterative visual reasoning." No retraining, just smarter prompting.
Why it matters: How you talk to a model and how you structure its thinking can matter just as much as how it was trained.
� Twitter� Reddit
🚀 Products & Launches
Google DeepMind Announces Gemini Robotics 2
Gemini Robotics 2 introduces "Whole-Body Intelligence," letting humanoids coordinate high-level reasoning with physical movement. It processes live video to handle failures in real-time and has shown a 45% improvement in task completion across unstructured environments.
Why it matters: This is a meaningful step toward AI that can actually operate reliably in the physical world.
� Twitter� Hacker News� Reddit
Funding & Deals
* Anthropic and Google formed a $15B partnership to build out major datacenter capacity in Texas, further cementing their compute alliance. x.com/AndrewCurran_/status/2082910104316493873
* DeepSeek is backed by a $4 billion consortium of private equity firms for its new 1-gigawatt facility in Mongolia. x.com/ns123abc/status/2082912664762618239
Launches
* Inkling-Small — Thinking Machines' new high-performance, parameter-efficient MoE model. huggingface.co/blog/thinkingmachines-inkling
* Agent-Manager — A new Tmux TUI tool to help developers juggle multiple coding agent instances. github.com/YoanWai/agent-manager
Closing thought: Between AI models optimizing their own hardware and hedge funds blowing up over compute bets, it's pretty clear the "infrastructure" layer of AI has quietly become the most volatile and important battlefield in tech.