Claude Cowork Doubles Rate Limits for the Next 30 Days

🧠 LAUNCH

Claude Cowork Doubles Rate Limits for the Next 30 Days

Anthropic just doubled the 5-hour rate limits on Claude Cowork for a full month. If you've been hitting walls on complex multi-agent refactors or large migrations, this is your window — the bigger ceiling means longer uninterrupted runs before throttling kicks in. Queue up your most ambitious project now. (4,934 likes | 297 RTs) Read more →

Gemma 4 QAT Checkpoints Land on Hugging Face Across All Sizes

Google ships quantization-aware trained checkpoints for every Gemma 4 model size. Unlike post-hoc quantization methods that compress a finished model and hope for the best, these checkpoints were trained with quantization baked in from the start — meaning INT4 and INT8 inference should hold significantly more quality than GPTQ or AWQ alternatives. If you're running quantized Gemma in production, swap these in and benchmark. (1,931 likes | 168 RTs) Read more →

Google Ships Nano Banana 2, Co-Scientist, and Dreambeans in One Week: Three GA releases — Nano Banana 2 via the Gemini API, Co-Scientist for structured multi-agent scientific hypotheses, and Dreambeans for personalized image generation. Google's shipping cadence is accelerating faster than anyone can keep up. (151 likes | 16 RTs) Read more →

Higgs Audio v3 — A 4B TTS Model Small Enough to Self-Host: BosonAI drops a 4-billion-parameter text-to-speech model that's trending hard on Hugging Face. At 4B parameters, it sits in the sweet spot — compact enough to run on a single GPU while delivering quality the community is already benchmarking against commercial APIs. (117 likes | 408 downloads) Read more →

🔬 RESEARCH

Opus 4.7 Matches Human Experts at NMR Spectroscopy — Not Textbook Questions, Real Spectra

Anthropic's new science blog shows Claude Opus 4.7 performing at expert level on NMR spectral interpretation — the bread-and-butter analytical technique for identifying molecular structures in chemistry and drug discovery. This isn't pattern-matching on training data; the model is reading raw spectra and making structural assignments that match or exceed experienced human chemists. The first convincing evidence that frontier models can do real analytical chemistry, not just ace exams. (1,655 likes | 169 RTs) Read more →

Sakana AI Opens the First Lab Dedicated to Recursive Self-Improvement: Sakana AI officially launches its RSI Lab in Tokyo — the first research group explicitly focused on open-ended self-improving AI systems. Led by the team behind evolutionary model merging, they're bringing Japan's manufacturing-inspired constraint philosophy to AI research. Worth watching closely. (389 likes | 48 RTs) Read more →

NVIDIA's NitroGen Wins CVPR Honorable Mention for Physics-General Embodied Agents: NitroGen earns Best Paper Honorable Mention at CVPR 2026 for general-purpose embodied agents that master multiple physics simulations. Four years from MineDojo to physics-general agents marks real, measurable progress in embodied AI — not just demo-ware. (190 likes | 23 RTs) Read more →

Cog Ships 100-Hour Agent Evals With Financial Guarantees on Accuracy: Cog publishes the first enterprise-grade agent evaluations that run 6x longer than METR's 16-hour ceiling — and backs them with financial guarantees on accuracy. The methodology uses GPT-5 to estimate human-equivalent time from Claude Code transcripts. First company confident enough to put money behind eval results. (216 likes | 12 RTs) Read more →

🔧 TOOL

Opus 4.1 Gets a Hard Retirement Date: August 5, 2026

The clock is ticking. Anthropic sets August 5, 2026 as the hard retirement date for Claude Opus 4.1. If you're still pinned to claude-opus-4-1-20250805 in production, you have exactly two months to migrate to Opus 4.8. Grep your codebase now — every hardcoded model ID is a ticking time bomb. Read more →

TanStack AI Adds First-Class MCP Support With CLI Typegen: TanStack AI now ships MCP integration with CLI-generated types, managed server lifecycle, and pooled connections. If you're building MCP-powered apps in React, Vue, or Solid, this is the integration layer you'd otherwise spend a week building yourself. (161 likes | 11 RTs) Read more →

The Official Claude Cowork Product Guide Drops: Anthropic publishes the definitive playbook for multi-agent delegation patterns in Claude Cowork. Covers workflow design, task decomposition, and when to use parallel vs. sequential agents. Pairs nicely with today's doubled rate limits. Read more →

Anthropic TypeScript SDK v0.101.0 Adds Middleware Support: The Anthropic TypeScript SDK now lets you intercept, transform, or log requests and responses without wrapping the client. Clean extensibility for observability and retry logic — npm update and explore the middleware hooks. Read more →

📝 TECHNIQUE

Your RL Environment Might Be Sabotaging Your Model: Latent Space publishes a practical guide to RL environment bugs that actively make models worse — with specific anti-patterns from real training trajectories. If you're fine-tuning with RL and your results plateau for no obvious reason, the harness itself might be the problem. Audit yours against their checklist. Read more →

How an Anthropic Sales Lead Rebuilt GTM Workflows With Claude Code: A non-engineer on Anthropic's sales team shares how he automated go-to-market workflows using Claude Code — real internal case study with patterns any ops or sales team can steal. The most interesting detail: the workflows that stuck weren't the clever ones, they were the boring repetitive ones. Read more →

💡 INSIGHT

Reports Surface of Claude API Returning Another User's Output During Outage: Unconfirmed reports of cross-tenant data leakage during today's Claude API outage — users claiming they received inference output intended for other accounts. Anthropic's status page confirmed elevated errors but hasn't addressed the data leak claims. If you send sensitive data through the API, monitor for an official post-mortem. (91 likes | 5 RTs) Read more →

Mythos Pricing Leak: $70/M Output Tokens and the Coming Frontier Tax: A pricing leak suggests Claude Mythos will cost $70 per million output tokens — 3.5x current Opus pricing, timed to launch alongside GPT 5.6 and Gemini 3.5. The frontier is getting expensive, and the market is about to split hard between tasks worth $70/M and tasks that aren't. Model your workload costs now. (712 likes | 16 RTs) Read more →

Did Claude Actually Increase Bugs in rsync? An Independent Analysis: A detailed independent analysis digs into whether AI-assisted contributions introduced more bugs into rsync's codebase. High-engagement HN discussion with 270 likes and 259 RTs — but read the methodology before citing the conclusions, because the answer is more nuanced than either side wants it to be. (270 likes | 259 RTs) Read more →

🏗️ BUILD

NVIDIA Nemotron 3 Ultra 550B BF16 Weights Now Downloadable: Full BF16 weights for NVIDIA's 550B MoE model are now on Hugging Face — 9,125 downloads already as the community stress-tests NVIDIA's open agent backbone at scale. You'll need serious multi-GPU infrastructure, but if you have it, this is the largest open-weights model available for agent workloads. (116 likes | 9.1K downloads) Read more →

🎓 MODEL LITERACY

Quantization-Aware Training (QAT): When you quantize a model — reducing its weights from 32-bit floats to 4-bit or 8-bit integers — you lose precision, and with it, quality. Standard post-hoc methods like GPTQ and AWQ take a fully trained model and compress it after the fact, hoping the model can tolerate the information loss. QAT flips this: the model trains with simulated quantization noise from the start, so it learns to be robust to lower precision during training itself. Today's Gemma 4 QAT release is a clean example — these checkpoints should hold significantly more quality at INT4/INT8 than the same architecture quantized after training. If you're deploying models on constrained hardware, QAT checkpoints are almost always the better starting point.

⚡ QUICK LINKS

OpenAI API Platform: Redesigned navigation makes finding endpoints and docs faster. (554 likes | 20 RTs) Link
Opus 4.7/4.8 Fix: Anthropic engineer confirms active work on performance issues — patch incoming. (114 likes | 5 RTs) Link
Python SDK v0.106.0: Opus 4.1 officially deprecated, Foundry client copy()/with_options() fixed. Link
Vercel AI SDK Canary: Unified voice API across OpenAI, Google, and xAI in one interface. Link
Lowfat: CLI filter claims 91.8% LLM token savings by stripping irrelevant content before sending. (102 likes | 54 RTs) Link

🎯 PICK OF THE DAY

When a model matches domain experts at interpreting raw NMR spectra, we've crossed from AI-as-tutor to AI-as-colleague. Anthropic's science blog showing Opus 4.7 performing at expert level on NMR spectroscopy isn't another "AI passes exam" story — NMR interpretation is messy, ambiguous, experience-dependent work that chemists spend years mastering. The model isn't answering multiple-choice questions about spectroscopy; it's reading real spectra and making structural assignments that stand up to expert review. The implications for drug discovery and materials science pipelines are immediate: NMR is a bottleneck in compound identification, and a model that can reliably assist at this level changes the throughput equation for any lab running analytical chemistry workflows. This is the kind of capability that makes AI genuinely useful to domain experts rather than just impressive to generalists. Read more →

Until next time ✌️