Tokens & Signals for 7/1/2026. We scanned ~1,200 Twitter accounts (1187 tweets), 13 subreddits (61 posts), Hacker News (9 stories), 5 newsletter posts, 4 podcast episodes, 128 Discord messages, and leaderboard data for you. Estimated reading time saved: ~11 hours.
* Anthropic redeployed Claude Fable 5 and Mythos 5 globally after a two-week export control suspension; users get 50% weekly usage through July 7. x.com/AnthropicAI/status/2072163884430229756
* Together AI secured an $800M Series C at an $8.3B valuation to scale its open-source infrastructure. x.com/vipulved/status/2072321276094673083
* Devin for Security is live; Fortune 500 pilots successfully remediated over 1,000 production vulnerabilities. x.com/walden_yan/status/2072377406267273248
* arXiv spun out from Cornell to become an independent nonprofit, securing long-term research access. x.com/jxmnop/status/2072083035995164733
* Ollama updated to v0.31, boosting Gemma 4 performance by 90% on Apple Silicon via MLX optimizations. x.com/ollama/status/2072121580201848926
* Etched came out of stealth with $800M and $1B+ in customer contracts for specialized transformer inference ASICs. x.com/tbpn/status/2072106011050266960
* Kling AI grabbed Silver and Bronze Lions at Cannes for "The Last Real Man" — AI film craft has officially arrived. x.com/Kling_ai/status/2072296943666016418
* vLLM now supports Qwen3.6-27B-NVFP4 on Blackwell GPUs, cutting VRAM usage by 2.5x. x.com/vllm_project/status/2072413378669134306
* @simonw on Sonnet 5 costs: "The tokenizer makes it ~1.4x more expensive for English, meaning it's sometimes pricier than Opus 4.8 per task." x.com/simonw/status/2072068898648949184
* EleutherAI released a benchmark for "engineering taste" to evaluate whether agents write maintainable code, not just correct code.
Best to Build With Today
* Coding: claude-opus-4-8-thinking-32k (Leads LiveBench/Arena coding).
* Reasoning: claude-opus-4-8-xhigh-effort (Current math/logic leader).
* Chat: gemini-3.1-pro (Top-ranked for general-purpose conversation).
* Open-source: Qwen3.6-27B-NVFP4 (Optimized for production on NVIDIA Blackwell).
* Value: claude-sonnet-4-7-thinking-32k (High-tier reasoning at a lower cost).
Deeper Dives
💼 Industry & Business
Together AI Raises $800M Series C
Together AI just closed $800 million at an $8.3 billion valuation. They build specialized infrastructure for training and inference, and their pitch is simple: stop being locked into proprietary API ecosystems.
* Why it matters: That valuation signals serious investor conviction that hardware-optimized, open-weights infrastructure isn't a niche bet — it's the next big one.
� Twitter
arXiv Transitions to Independent Nonprofit
After 35 years under Cornell's roof, arXiv has spun out into an independent nonprofit backed by the Simons Foundation and Schmidt Sciences. The move gives it more organizational flexibility and puts it on steadier financial footing.
* Why it matters: arXiv is basically the heartbeat of AI research. If it's healthy, science moves faster.
� Twitter� Hacker News� Reddit
Etched Announces Specialized Inference Hardware
Etched just came out of stealth with $800 million in funding and over $1 billion in customer contracts for custom inference ASICs — chips built to sidestep the thermal and bandwidth walls that general-purpose GPUs keep hitting.
* Why it matters: Silicon designed specifically for transformers could be the unlock for serving large models at scale without bleeding money on compute.
� Twitter
🧠 Models & Research
Claude Sonnet 5 Tokenizer Efficiency Debate
Turns out the new Sonnet 5 tokenizer bumps real-world costs by about 1.4x for English tasks — enough that some users are finding it pricier per completed task than the more powerful Opus 4.8.
* Why it matters: It's a good reminder that sticker price per token and what you actually pay to get something done can be very different numbers.
� Twitter� Reddit
New Benchmark for AI Agent Engineering Taste
EleutherAI dropped a benchmark aimed at "engineering taste" — testing whether an agent makes sensible design choices, not just whether its code runs without crashing.
* Why it matters: Passing tests is table stakes. This is about whether the AI writes code a human would actually want to maintain.
� Twitter� Discord
🚀 Products & Launches
Anthropic Redeploys Claude Fable 5
After working things out with the U.S. government, Anthropic has lifted export controls on Fable 5 and Mythos 5. Both models ship with updated cybersecurity classifiers and are available globally as of July 1.
* Why it matters: Clears a geopolitical compliance headache for anyone who needs their most advanced reasoning models — and couldn't use them for the past two weeks.
� Twitter� Hacker News
Devin for Security Released
Devin's Security Swarm is now open for enterprise use. Early pilots with Fortune 500 companies found and patched over 1,000 production vulnerabilities — not in a sandbox, in prod.
* Why it matters: This isn't a coding demo anymore. Autonomous security remediation at enterprise scale is a genuinely different category.
� Twitter
Funding & Deals
* Together AI: Raised $800M at an $8.3B valuation to expand generative AI infrastructure.
* Etched: Raised $800M total with $1B+ in contracts to build custom inference ASICs.
Launches
* Devin Security Swarm: Agentic vulnerability detection and patch generation for enterprise.
* Qwen3.6-27B-NVFP4: Quantized model variant optimized for Blackwell and Hopper GPUs.
Closing thought: With Fable 5 back online, Together AI's massive round, and Etched bringing purpose-built silicon into the mix, the industry is clearly moving past "can it do this?" and into "can it do this reliably, at scale, without costing a fortune?"