Week of April 13, 2026 — Sia Reads What Matters

RELEASE2026-04-16

Claude Opus 4.7 ships with 87.6% SWE-bench Verified, hi-res vision, task budgets, and a new tokenizer — Anthropic retakes the GA crown

Anthropic released Claude Opus 4.7, jumping SWE-bench Verified from 80.8% to 87.6% and CursorBench from 58% to 70%. It's the first Claude model with high-res image support (up to 2576px), introduces task budgets that let the model manage long agentic loops gracefully, and ships with a new tokenizer. Pricing stays at $5/$25 per million tokens.

Anthropic Releases Claude Opus 4.7 — Strongest Generally Available Model With 87.6% SWE-bench

Anthropic Announcement What's New in Opus 4.7 VentureBeat Analysis

RELEASE2026-04-16

Codex can now see your screen, click through apps, generate images, schedule future work, and run 90+ plugins — coding assistant is underselling it

OpenAI shipped a massive Codex desktop update that adds background computer use (multiple parallel agents controlling your Mac), an in-app browser for annotating web pages, image generation via gpt-image-1.5, persistent memory, and the ability to schedule and wake up for long-running tasks. Over 90 new plugins including Atlassian Rovo, GitLab Issues, and CircleCI expand its reach well beyond code.

OpenAI Blog gHacks Breakdown Help Net Security

RELEASE2026-04-17

OpenAI built a reasoning model just for drug discovery and genomics — GPT-Rosalind scores above the 95th percentile of human experts on prediction tasks

GPT-Rosalind is OpenAI's first purpose-built domain-specific model, optimized for biochemistry, genomics, and protein engineering. It supports evidence synthesis, hypothesis generation, and multi-step scientific workflows. In Codex evaluations it ranked above the 95th percentile of human experts on prediction tasks. Access is restricted to a trusted-access program with partners including Amgen, Moderna, and the Allen Institute.

OpenAI Announcement Fierce Biotech Axios Coverage

HOT TAKE2026-04-16

Google DeepMind's CEO says the commercial race triggered by ChatGPT shifted the field away from what matters — and he's betting on scientific AI instead

In a Fortune interview tied to a new biography, Hassabis argued that ChatGPT's viral success in 2022 shifted AI into a fast, product-driven dynamic that deprioritized curiosity-driven research. He says the chatbot wars are a sideshow compared to getting AGI safely over the line, and that education needs to be rebuilt from the ground up. Coming from the CEO of the lab that just shipped Gemini Robotics-ER 1.6 and Gemma 4, the critique of product-first AI carries real weight.

Fortune Interview Fortune London Profile

NEWS2026-04-14

A week after Anthropic locked Mythos behind Project Glasswing, OpenAI ships its own cyber-focused model to a select group of defenders

OpenAI released GPT-5.4-Cyber, a fine-tuned variant of GPT-5.4 designed for spotting software vulnerabilities, to a limited group of vetted security professionals. The timing is pointed — it lands a week after Anthropic's Project Glasswing restricted Mythos Preview to defensive cybersecurity. Two frontier labs now have dedicated cyber models, and neither is publicly available. The defensive AI arms race is real.

Bloomberg Report The Hacker News

RELEASE2026-04-15

DeepMind's new robotics model reads gauges, estimates proportions, and reasons about physical space — Boston Dynamics is already integrating it into Spot

Gemini Robotics-ER 1.6 combines visual reasoning with code execution to achieve precise instrument readings and spatial understanding in physical environments. The model uses an 'agentic vision' pipeline: snapshot, resolve fine details, run code to estimate proportions, then reason about the reading. Boston Dynamics is integrating it into Spot's AI Visual Inspection system. Available via the Gemini API and Google AI Studio.

SiliconANGLE Report MarkTechPost Analysis