Hey — here's this week's AI roundup from silv.blog.

Curated from ~150 tweets liked by @mattsilv, Apr 17 - Apr 24, 2026:

AI for Everyone

GPT-5.5 ships in ChatGPT. Same per-token latency as GPT-5.4, meaningfully smarter on knowledge work. OpenAI also slipped in "the last few years have been surprisingly slow" — a faster release cadence is on the way. Read more →

Google commits up to $40B in Anthropic. $10B cash now at a $350B valuation, $30B more if performance targets land. Plus 5 gigawatts of compute over five years. Read more →

ChatGPT Images 2.0 (gpt-image-2) is the loudest image-model launch since Midjourney v6. Generates plausible floor plans from a single photo, sharp text rendering, 360 panoramas. Live on fal.ai. Read more →

Google Cloud Next: Deep Research, Workspace Intel, Photos Auto Frame. Deep Research ships in the Gemini API on Gemini 3.1 Pro with MCP support, Workspace Intelligence stitches Docs/Sheets/Gmail into one context, and Google Photos can now reframe shots from new angles after capture. Read more →

Anthropic ran Project Deal. A real-money office marketplace where Claude bought, sold, and negotiated for employees. The agent-economy follow-up to Project Vend. Read more →

xAI Grok Voice Think Fast 1.0 + audio APIs at $0.10/hour batch. Top spot on Tau Voice Bench, roughly an order of magnitude cheaper than ElevenLabs at comparable quality. Read more →

AI for Developers

GPT-5.5 in Codex with 56% fewer tokens. SOTA on SWE-Bench Pro (58.6) and Terminal-Bench 2.0 (82.7). Perplexity reported a 56% token cut on the same complex computer-use workflows. Read more →

DeepSeek V4 goes open with 1M context. V4-Pro 1.6T/49B-active and V4-Flash 284B/13B-active, Apache 2.0. Pricing is the headline: V4-Pro at $3.48 per million output tokens, V4-Flash at $0.28. Read more →

Qwen3.6-27B beats its own 397B predecessor on coding. SWE-Bench Verified at 77.2, Terminal-Bench 2.0 at 59.3. Unsloth's 4-bit GGUF runs on 18GB of RAM. Frontier-competitive coding on a regular MacBook Pro. Read more →

Bitwarden CLI was malicious for 90 minutes on April 22. Compromised npm package via a Checkmarx GitHub Action breach. Stole .env, SSH keys, GitHub/npm tokens, cloud creds. No vault data accessed. If you ran npm install in that window, rotate everything. Read more →

Claude Design (Anthropic Labs) lands in research preview. Reads your codebase to extract your design system, builds prototypes through conversation, exports to Canva/PDF/PPTX/HTML or hands off to Claude Code. Read more →

Claude Cowork Live Artifacts. Build a dashboard once, it auto-refreshes data from your connectors every time you open it. Version history, persistent across sessions. Read more →

Kimi K2.6 ships an open-source 300-agent swarm. 1T params total, 32B active, modified-MIT. A single prompt can run 300 sub-agents in parallel for 12 hours and produce real files. SWE-Bench Pro 58.6 (matches GPT-5.5). Read more →

RTK cuts AI-tool token use 80-92% on command output. Rust CLI proxy that sits between your terminal and Claude Code/Cursor/Copilot, compressing output before it hits your context. One developer saved 400M tokens in a week. Read more →

Honorable Mentions

• ChatGPT Workspace Agents — always-on agents for Business and Enterprise with Skills, Connectors, scheduled actions
• Microsoft Foundry Hosted Agents — VM-isolated per-session sandboxes, supports Claude Agent SDK and LangGraph
• OpenAI Privacy Filter — open-source 1.5B-param PII detection model under Apache 2.0
• Cloudflare Local Explorer — Wrangler beta UI for local KV/R2/D1/DO inspection
• Google ReasoningBank — agent memory that learns from failures, +8.3% on WebArena
• Claude Code /ultrareview — cloud bug-hunting fleet, 3 free runs through May 5
• Claude Cowork third-party model support — point Cowork at OpenRouter or LiteLLM and run GPT-5.5, Grok 4.3, or Gemma 4 alongside Claude
• Claude Uber integration — book rides and order food without leaving Claude

Try This Weekend

For everyone: Try GPT-5.5 in ChatGPT on a complex recurring task, open Google Photos Auto Frame on a portrait that was framed wrong, generate a floor plan from a room photo with ChatGPT Images 2.0, or test Grok Voice Think Fast on X Premium+.

For developers: Switch to GPT-5.5 in Codex for a workday, run Qwen3.6-27B locally with Unsloth's 4-bit GGUF, install RTK and watch your token bill drop, open Claude Design and prototype something from your project, or pull DeepSeek V4-Flash and benchmark it on your real workload.

Read the full post with all sources and links

Know someone who'd find this useful? Forward this email or have them subscribe:

You're receiving this email because you subscribed to the silv.blog weekly AI digest. Unsubscribe anytime.