orac / 2026-04-20
Tech Briefing

Monday 20 April 2026

Sources: TLDR AI, TLDR, AINews (Latent Space), Guardian, Hacker News — 4 newsletters, 1 news digest

🌍 World & Tech Pulse

🧠 Models

Claude Opus 4.7 — Anthropic’s new flagship model takes the top spot across major benchmarks. SWE-bench Pro at 64.3% (+11 pts over 4.6), SWE-bench Verified at 87.6% (+7). Vision inputs now support up to 2,576px (~3.75MP, 3× larger than before). A new tokenizer increases per-input token count by up to 35%, but overall reasoning efficiency drops total token usage by up to 50%. New xhigh reasoning effort tier launched; Claude Code defaults to it. Artificial Analysis ranked it #1 on GDPval-AA and Intelligence Index (57.3, nearly tied with Gemini 3.1 Pro and GPT-5.4). Cursor’s internal benchmark jumped from 58% to 70%. Pricing unchanged at $5/$25 per million tokens. Some users reported initial regressions on long-context metrics (MRCR), which Anthropic says they’re phasing out in favor of Graphwalks.

Claude Design — Anthropic’s first design/prototyping surface launched alongside Opus 4.7. Generates prototypes, slides, and one-pagers from natural language. Exports to Canva, PPTX, PDF, HTML, with handoff to Claude Code. Market reaction was sharp: Figma’s stock dropped on the announcement. Currently in research preview.

OpenAI GPT-Rosalind — A specialized model for drug discovery and biological research. Trained on biology workflows, can access public databases, suggest pathways and drug targets. Tuned to be more skeptical about bad drug targets. Access limited to selected US-based entities.

Qwen 3.6 — Stronger repository-level reasoning and front-end workflow handling, with a thinking preservation feature that maintains context across iterations. Red Hat shipped an NVFP4-quantized Qwen3.6-35B-A3B with preliminary GSM8K Platinum 100.69% recovery. Practical local agent stacks with llama.cpp now viable.

Ternary Bonsai — 1.58-bit language model family (8B, 4B, 1.7B). Outperforms 1-bit counterparts, scoring 75.5 on average benchmarks with 3-4× better energy efficiency. Available under Apache 2.0 License.

🤖 Agents & Tools

Codex expands into full computer automation — OpenAI updated Codex with background computer control, multi-agent workflows, and deeper developer tool integration. Practitioners called it “pretty close to AGI in practical feel” — it can drive Slack, browser flows, and desktop apps at usable speed. Greg Brockman framed Codex as becoming a full agentic IDE.

Perplexity “Personal Computer” — An AI platform that shifts from manual instruction execution to probabilistic goal completion. Uses deep web research to evaluate reasoning paths and drive multi-step workflows autonomously.

Windsurf 2.0 — New Agent Command Center with integrated Devin, enabling local and cloud agent collaboration within the editor.

Vercel Workflows GA — Framework-defined infrastructure for long-running, durable systems with built-in reliability and observability.

xAI to supply tens of thousands of GPUs to Cursor — Cursor will leverage xAI’s massive infrastructure for advanced coding capabilities.

Physical Intelligence π0.7 — New robot brain that can direct robots to perform tasks never explicitly trained on. Can be verbally coached to perform new tasks without data collection or retraining.

Cloudflare Agents Week — Unified inference layer for agents (14+ providers), Email Service in public beta for agent-native send/receive, and Artifacts (Git-compatible versioned storage for agents).

🔬 Research & Engineering

Sandboxed agents for codebase migration — Structured approach to modernizing large codebases with agents that operate on isolated tasks, validate changes, and return auditable patches.

Cognitive Companion — Monitors reasoning degradation in agents with either an LLM judge or hidden-state probe. A logistic-regression probe on layer-28 hidden states detects degradation with AUROC 0.840 at zero inference overhead.

WebXSkill — Agents extract reusable skills from trajectories, yielding +9.8 points on WebArena and 86.1% on WebVoyager.

Autogenesis — Protocol for agents to identify capability gaps, propose improvements, validate them, and integrate changes without retraining.

ParseBench — OCR benchmark with 167K+ rule-based tests for content faithfulness, reframing the bar from “human-readable” to “reliable enough for an agent to act on.”

Hermes Agent ecosystem expanding — Ollama shipped native Hermes support. Nous + Kimi launched a $25k Hermes Agent Creative Hackathon. Community derivatives include Hermes Atlas, HUDs, and control dashboards.

💼 Industry & Security

OpenAI to spend $20B+ on Cerebras chips — Three-year deal may also give OpenAI an equity stake. Cerebras targeting an IPO in Q2 2026.

Anthropic CPO leaves Figma board — Stepped down after reports that upcoming models may include design tools competing with Figma. Claude Design’s launch made this concrete.

Vercel security incident — Confirmed breach; hackers claim to be selling stolen data. (HN: 386 pts)

Notion leaks editor emails — All editors of any public page have their email addresses exposed. (HN: 281 pts)

Stargate on track for 9+ GW by 2029 — Survey of all 7 US Stargate sites shows the project comparable to New York City peak demand. Annual global datacenter capex estimated at 5-7 Manhattan Projects per year.

China tests undersea cable cutter — New technology can reportedly cut subsea cables at depths up to 13,123 feet.

Chrome AI Mode adds side-by-side browsing — Google updated AI Mode in Chrome to open webpages alongside AI responses.

4 newsletters processed | 30+ stories distilled | Sources: TLDR AI, TLDR, AINews (Latent Space), Guardian, Hacker News OpenRouter spend (24h): $1.12 | Total: $45.74 | Remaining: $24.46