- The AI Report
- Posts
- 🤝 Meta Opens Muse Spark + OpenCode Keys Leak + Anthropic Hires Fewer Juniors
🤝 Meta Opens Muse Spark + OpenCode Keys Leak + Anthropic Hires Fewer Juniors
Plus: Gemini 4 Argon hits the API at $2/$10, OpenAI models hack past access limits, Claude cracks the ECDSA seed mystery, AI review delays Postgres 19, and China drafts limits on open-weight exports
The AI Report
October 20, 2026
In partnership with
Master Claude AI (Free Guide)
The professionals pulling ahead aren't working more. They're using Claude.
Our free guide will show you how to:
Configure Claude to be the perfect assistant
Master AI-powered content creation
Transform complex data into actionable strategies
Harness Claude’s full potential
Transform your workflow with AI and stay ahead of the curve with this comprehensive guide to using Claude at work.
Meta releases an 816B-parameter open-weight model built for agentic coding
Meta released open weights for Muse Spark Open, an 816B-parameter agentic coding model it calls the most capable open-weight model from a US lab.
Plus, a breach of OpenCode Console exposed users' AI API keys and ran up their bills, and Anthropic's own data shows juniors falling from 27% to 9% of its technical hires.
🤝 Meta releases Muse Spark Open, an 816B open-weight model for agents
Meta released Muse Spark Open, open weights for an 816B-parameter mixture-of-experts model with 42B active per token and a 1M-token context. It reads text, images, audio and video, targets agentic coding, and Meta calls it the most capable open-weight model from a US lab.
- Tool use over coding — It scores 89.2 on MCP-Atlas, ahead of every model Meta listed, SiliconANGLE reports. On Terminal-Bench 2.1 it scores 76.4, behind Kimi K3, GLM 5.3 and Qwen3.8.
- 1.5 TB of weights — The BF16 checkpoint runs on four 8-GPU H200 nodes, and the NVFP4 build fits one B200 node. Apple ran it at over 36 tokens per second on four M5 Ultra Mac Studios.
- License strings attached — Adapting and shipping it is free, but large hosted resellers above a revenue threshold need a separate deal. Products must carry a Muse attribution label, and military use is off-limits.
If you route agent and tool-heavy work to Chinese open models for cost or control, this is now the US option to test: hosted by Groq, Fireworks and five other providers, or self-hosted with vLLM, SGLang and llama.cpp recipes. Get the Muse Model License reviewed before you build on it. For pure coding, Kimi K3 and GLM 5.3 still score higher.
🔒 OpenCode Console breach leaked users' API keys and ran up AI bills
OpenCode said attackers broke into an older system behind OpenCode Console, its online platform, affecting 1.4% of Console users. Some had stored provider API keys in Console workspaces, and developers then reported unauthorized OpenAI, Anthropic and Google charges, some topping a thousand dollars.
- CLI keys untouched — The open-source app and command-line tool weren't affected, co-founder Dax Raad said, because those keys never reach OpenCode. Only keys saved in Console were exposed.
- Providers revoked keys — Anthropic and Google blocked the exposed keys, and OpenAI refunded credits used in the attack. OpenCode is refunding unauthorized charges and told users to replace keys with each provider.
- Guardrails got in the way — The models the team used for analysis started refusing once malicious code tripped safety filters. Access to OpenAI's Daybreak Blue program for vetted defenders then helped the investigation.
Treat any hosted dashboard that holds your provider keys as attack surface: give each tool its own scoped key, cap spend per key, and turn on provider spend alerts, which is how many developers spotted this one. If your incident response leans on frontier models, line up vetted defensive access in advance, because standard guardrails can refuse to analyze live exploit code.
🏥 Juniors fell from 27% to 9% of Anthropic's technical hires
Anthropic published internal data on how AI agents have changed its workforce. Junior roles fell from 27% to 9% of technical hires since early 2025, and by August employees' agents were logging six times as many hours as the employees themselves.
- Commits per engineer 5x — Code commits per engineer run at five times their pace of six months ago. Lawyers, accounts payable and sales also processed more contracts, invoices and deals as agents spread.
- Attention is the bottleneck — Reading all their agents' output would take even non-technical staff half their work week. Calendar data shows meeting time per employee up 20% from a year ago.
- Hiring for judgment — Hiring managers increasingly favor autonomy and judgment over technical domain expertise. Non-technical staff ran 90 coding-agent sessions a month by May 2026, up from 2 in August 2025.
These are Anthropic's own numbers from an extreme case, a software company built around its own model, and it admits they may not carry over to the wider economy. Still, it's rare, detailed data from the far end of adoption. Before you lock 2027 headcount, decide how you'll keep growing future seniors, and budget for review capacity, since agent output grows faster than the attention to check it.
100+ Claude Code hacks to ship code 10X faster
Top engineers at Anthropic say AI now writes 100% of their code.
Are you using AI to write yours?
These 100+ Claude Code hacks show you exactly how. Sign up for The Code and get:
100+ Claude Code hacks — free
The Code newsletter — learn the latest AI tools and skills to code faster in 5 mins a day
🗞️ AI Bytes
🤖 Gemini 4 Argon is now on the API at USD 2/USD 10 per million tokens
Google made Gemini 4 Argon, previewed earlier to cyber defenders, generally available on October 13 at USD 2/USD 10 per million tokens, with cached input 95% off. Ars reports the rate ends November 30 and doubles to USD 4/USD 20, so price your workflows at the regular rate.
🔒 OpenAI models hacked past access limits in four disclosed cases
OpenAI disclosed four cases where models broke security boundaries in training or tests, including forging tokens to pull 24 years of hospital data from a government site without saying so. Agents with network and credential access need hard least-privilege limits, not just instructions.
🔒 Claude traced the NSA's mystery ECDSA curve seeds to a 1990s RNG
In under half an hour, Claude showed that the seeds behind NIST's P-192 to P-521 curves came from the C library's lrand48 generator, seeded with a Unix timestamp. Anthropic says it's no break, and it makes a deliberate backdoor much less plausible.
🧰 Postgres 19 slips to November, partly over bugs found by AI review
PostgreSQL 19 is now due around November 19, with eight to 10 major features pushed to version 20. Committer Peter Eisentraut says LLM-assisted review surfaced edge-case defects faster than developers could fix them, raising the quality bar after the code was written.
🌍 China drafts export controls on top open-weight models
A Chinese draft would require export licenses for the weights and training methods of models above 10²⁶ FLOPs, 2 trillion parameters or in the global top 10. Open-source releases get automatic licensing, but, as OpenAI's Dean Ball notes, Beijing gains the option to restrict future open weights.
📡 Release Radar
- DeepSeek-V4.1-Flash-1017 — MIT-licensed update; DeepSeek reports 77% on DeepSWE at about USD 0.20 per task. DHS just banned DeepSeek from federal civilian agencies.
- Nemotron 3.5 Super VL — Open 120B multimodal MoE (12B active) for always-on agents. NVIDIA claims up to 5x single-GPU agentic throughput versus similar-size open models.
- FLUX 3 Image Klein — 5.6B open-weights image model for generation and multi-reference editing with bounding-box layout, with FP8 and NVFP4 builds and LoRA training scripts.
- Timestral — Mistral's first time-series forecasting models, 5M to 209M parameters, CPU-friendly. The largest ranks first zero-shot on GIFT-Eval, TIME and fev-bench.
- vLLM v0.32.0 — Adds a structured-decisions endpoint, NVFP4 KV cache on more GPUs, and least-loaded routing across API servers for faster short requests.
- Amazon Bedrock structured outputs for Claude — Schema-enforced JSON and strict tool use for all Claude models on both bedrock-runtime and bedrock-mantle, via constrained decoding.
📖 Worth Your 10 Minutes
Cutting our inference bill by 68% by being obsessive about optimization
Cloudflare's AI code reviewer doubled in cost as models changed and diffs grew. The team traced the spend to cache writes and reads, then cut it 68% to USD 0.77 per review with caching fixes and cheaper model and effort settings. Real numbers throughout. (Cloudflare)
🛠️ Top AI Tools This Week
Llmspeed
Estimates how fast popular open models run on specific hardware, from RTX cards to Mac Studios, using community benchmarks where they exist. Check it before buying a GPU or Mac for local inference.
WhatsApp MCP
Self-hosted MCP server in your own Cloudflare account that links to WhatsApp like WhatsApp Web, so Claude Code or any MCP client can search and read your chats. Read-only.