Tokens & Signals for 5/15/2026. We scanned ~1,200 Twitter accounts (1099 tweets), 13 subreddits (47 posts), Hacker News (10 stories), 4 newsletter posts, 7 podcast episodes, 153 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* OpenAI just launched a personal finance dashboard for ChatGPT Pro ($200/mo), connecting to bank accounts via Plaid to automate spending tracking and cash flow forecasting. x.com/kimmonismus/status/2055320528198521041
* OpenAI is merging its ChatGPT and Codex teams under Greg Brockman to unify product strategy, backed by a $12B injection that values its new research entity at $150B. x.com/ZeffMax/status/2055335888591433852
* Anthropic wiped all Claude rate limits for Pro and Team users — clearly playing defense against the latest wave of model releases. x.com/ClaudeDevs/status/2055347539923308703
* Cerebras hit a $64B valuation in its IPO, proving the market is starving for hardware alternatives to NVIDIA. x.com/tbpn/status/2055069621439590522
* arXiv is officially done with AI-generated slop, issuing 1-year bans for papers containing hallucinations or fake citations. reddit.com/r/MachineLearning/comments/1tdje2d/a...
* xAI finished training Grok V9 — a 1.5T parameter model that uses "Curriculum Sparse Attention" to save 30% on compute while hitting 91.2% on MMLU. x.com/XFreeze/status/2055307600586043635
* @omarsar0 on coding agents: "Turns out simple grep-based search is often beating complex vector retrieval (RAG) for navigating codebases. Less noise, more precision." x.com/omarsar0/status/2055317577031975269
* @gaborcselle on OpenAI's finance move: "They're not building a product — they're building a habit. Connect your bank, get used to asking ChatGPT about money, then watch what happens next."
* Claude Mythos reportedly reverse-engineered and identified macOS zero-day vulnerabilities in just 5 days of autonomous operation. reddit.com/r/ClaudeAI/comments/1tdr1o7/claude_m...
* Meta's new 'KV self-pruning' technique lets models shed unnecessary memory during training, cutting usage by up to 85% with zero performance drop. x.com/daniel_mac8/status/2055308764547367416
* Together AI is now transcribing 303 seconds of audio per second, which basically makes real-time voice agents a commodity. x.com/togethercompute/status/2055062437968113875
Best to Build With Today
* Coding — gpt-5.4-xhigh leads for complex agentic workflows, while claude-opus-4-6-thinking-auto is the powerhouse for deep refactoring.
* Reasoning — claude-opus-4-6-thinking-auto tops the reasoning benchmarks, with gemini-3.1-pro winning on pure mathematical logic.
* Chat — gemini-3.1-pro currently dominates the Chatbot Arena ELO leaderboard.
* Voice — Together AI's Parakeet TDT is the clear winner for high-throughput, low-latency STT.
* Value pick — gemini-2.5-flash remains the king of cost-efficiency for high-volume tasks.
Deeper Dives
💼 Industry & Business
OpenAI's Major Reorg Creates $150B Standalone Entity
OpenAI is splitting into two units: a new $150B entity focused on frontier research and enterprise, and a consolidated product arm. Greg Brockman now leads all products, with a $12B capital injection from Microsoft, Thrive, and Khosla.
Why it matters: Carving out "research for humanity" from the commercial side means they're done playing coy about enterprise monetization — expect them to get aggressive.
� Twitter�️ Podcast
Cerebras IPO Raises Market Expectations to $64B
Blowing past initial $50B forecasts, Cerebras went public to fund its third-generation "Condor" wafer facility.
Why it matters: Investors are clearly betting on custom silicon to break the NVIDIA bottleneck — and they're putting serious money behind it.
� Twitter�️ Podcast
Anthropic Resets Claude Rate Limits
Anthropic wiped all usage caps for Pro and Team users, an obvious response to the competitive heat from xAI and OpenAI.
Why it matters: Rate limit wars are the new pricing wars — labs are choosing developer retention over capacity management.
� Twitter
ArXiv Bans AI-Hallucination Papers
A one-year ban is now in place for researchers who submit papers with fabricated data or AI hallucinations.
Why it matters: It's a necessary firewall against the flood of synthetic academic junk that's been quietly eroding the platform's credibility.
� Reddit
🧠 Models & Research
xAI Completes Grok V9 Run
Grok V9 is a 1.5T parameter model using "Curriculum Sparse Attention" to cut compute costs by 30% while hitting 91.2% on MMLU.
Why it matters: xAI is showing it can build massive models that are both smarter and cheaper to train — that's not a small thing.
� Twitter
Claude Mythos Cracks macOS Security
Claude Mythos identified macOS zero-days in just 5 days of autonomous fuzzing, outpacing human penetration testers.
Why it matters: We've crossed a real threshold — reasoning models are now legitimate cybersecurity tools, not just demos.
� Reddit
Grep Beats Vector Search for Coding Agents
Benchmark tests on SWE-bench show that simple grep-based search outperforms complex RAG for code navigation, cutting context noise by 45%.
Why it matters: Sometimes the boring old tool wins. Developers should think twice before over-engineering their agent stacks with RAG.
� Twitter
Meta's KV Self-Pruning
Meta's new training-time method lets LLMs prune their own memory, shrinking their footprint by up to 85% without any performance loss.
Why it matters: Long-context windows (100k+ tokens) just got a lot cheaper to run on consumer hardware.
� Twitter
Funding & Deals
* Factory AI acquired Lumetric to integrate AI-native desktop experiences with their robotics platform.
* OpenAI received $12B from Microsoft, Thrive, and Khosla to support its product and research restructuring.
Launches
* ChatGPT Personal Finance — A new Pro integration connecting to bank accounts via Plaid for automated financial insights.
Closing thought: Between arXiv cracking down on hallucinated research and grep outperforming vector RAG, something is shifting. We're leaving the hype phase and entering the "does it actually work?" phase. Practicality is the new frontier.