1. Claude Opus 4.7 Launches with 13% Coding Gain and an Autonomous Rust TTS Engine.
Anthropic released Claude Opus 4.7 on April 16, posting substantial improvements over Opus 4.6 across coding, vision, and autonomous task completion. On a 93-task coding benchmark, Opus 4.7 resolved 13% more tasks than Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could solve. CursorBench rose to 70% from 58%; Rakuten-SWE-Bench achieved three times more production-level task resolutions. Computer use on the XBOW visual acuity benchmark reached 98.5%, compared to Opus 4.6's 54.5% — the first model to pass implicit-need tests on this benchmark.
Perhaps the most striking demonstration: Opus 4.7 autonomously built a complete Rust text-to-speech engine from scratch, including the neural model, SIMD kernels, and a browser demo, then verified its output against a Python reference using speech recognition. The model ships with a new xhigh thinking effort level between high and max, and a cyber safeguard that automatically detects and blocks requests indicating prohibited cybersecurity uses. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens.
Source: Anthropic | 2026-04-16
2. OpenAI Launches GPT-Rosalind, Its First Domain-Specific Model for Life Sciences Research.
OpenAI introduced GPT-Rosalind on April 16 — a frontier reasoning model purpose-built for biology, drug discovery, and translational medicine, named after Rosalind Franklin. On LABBench2, a real-world life sciences benchmark, GPT-Rosalind outperforms GPT-5.4 on 6 of 11 tasks, with the largest gains on CloningQA, which covers end-to-end design of DNA and enzyme reagents for molecular cloning protocols. Partner Dyno Therapeutics used the model for RNA sequence-to-function prediction; best-of-ten model submissions ranked above the 95th percentile of human experts on the prediction task and approximately the 84th percentile on sequence generation.
GPT-Rosalind launches via trusted-access program for qualified U.S. Enterprise customers including Amgen, Moderna, the Allen Institute, Thermo Fisher Scientific, and Los Alamos National Laboratory. The Life Sciences Research Plugin for Codex connects to over 50 scientific tools and data sources. Advisory partners McKinsey, BCG, and Bain are positioned as rollout partners for regulated industries. Security controls govern eligibility, access management, and organizational governance; no existing credits are consumed during the research preview.
Source: OpenAI | 2026-04-16
3. OpenAI Codex Gets Full Desktop Control, Parallel Agents, Memory, and Image Generation.
OpenAI shipped a major overhaul of Codex on April 16, used by over three million developers weekly, targeting Anthropic's Claude Code dominance directly. Five capability pillars define the release. Background computer use gives Codex its own cursor on macOS, able to open any application, click, and type — multiple agents can work in parallel without interfering with the user's own session. An in-app browser lets Codex work natively with web applications and receive precise instructions by commenting on pages. Integration with gpt-image-1.5 enables image generation and iteration for product concepts, mockups, frontend designs, and games within the same workflow. Over 90 additional plugin integrations cover Atlassian Rovo, GitLab Issues, Microsoft Suite, Neon by Databricks, and more. A Memory preview retains personal preferences, corrections, and context across sessions, with full personalization coming to Enterprise, Edu, and EU users. Rolling out now to Codex desktop app users signed in with ChatGPT.
Source: OpenAI | 2026-04-16
4. Physical Intelligence π0.7 Generalizes to Robot Tasks It Was Never Taught.
Physical Intelligence published research on April 16 showing its model π0.7 can direct robots to perform tasks never explicitly trained — a capability the company's own researchers found surprising. The core achievement is compositional generalization: synthesizing skills from different training contexts to solve novel problems. A demonstration combined fragments from unrelated training data — one where a different robot pushed an air fryer closed, another from an open-source dataset where a robot placed a bottle inside — into a functional understanding of how an air fryer works. With zero coaching the model made a passable attempt at cooking a sweet potato; with step-by-step verbal instructions it succeeded; with prompt refinement, success climbed from 5% to 95% in approximately 30 minutes. The model matched specialist model performance across coffee-making, laundry folding, and box assembly. Generalization enables real-time deployment improvement without additional data collection or retraining. Physical Intelligence is two years old, San Francisco-based, has raised over $1 billion at a $5.6 billion valuation, and is in discussions for a new round at approximately $11 billion.
Source: TechCrunch | 2026-04-16
5. Salesforce Headless 360 Exposes 100+ API Tools, Making the Entire Platform Operable by AI Agents.
Salesforce unveiled Headless 360 at its TDX developer conference on April 16, exposing every capability in its platform as an API, MCP tool, or CLI command. The initiative ships 60+ new MCP tools and 30+ preconfigured coding skills giving external agents — Claude Code, Cursor, Codex, Windsurf — complete live access to a customer's entire Salesforce org. Agent Script, an open-source deterministic state machine DSL, enables versionable flat-file control of agent behavior. A new AgentExchange marketplace unifies 10,000 Salesforce apps, 2,600-plus Slack apps, and 1,000-plus Agentforce agents and tools. Salesforce is moving from per-seat licensing to consumption-based pricing for Agentforce. One early customer, B2B travel management firm Engine, built a customer service agent in 12 days that now handles 50% of cases autonomously. The announcement addresses a sector-wide SaaS sell-off driven by fears AI could render traditional SaaS business models obsolete.
Source: VentureBeat | 2026-04-16
6. TSMC Capex Guidance Near $56B Upper Bound as AI Demand Remains Unsatisfiable Through 2026.
TSMC's leadership confirmed on April 16 that the company's 2026 capital expenditure guidance sits near the top of its previously announced $52–56 billion range. CEO and Chairman C.C. Wei stated the company is "全力加速、提前采购设备" — accelerating at full capacity and procuring equipment ahead of schedule — yet supply remains tight and demand continues to grow. The driver is AI and high-performance computing, where demand outpaces TSMC's ability to add capacity. The guidance range itself signals TSMC's confidence that AI compute demand will sustain at levels requiring near-maximum foundry investment through the end of 2026.
Source: 36氪 / 财联社 | 2026-04-16
7. China Daily AI Token Usage Exceeds 140 Trillion Calls, Up 40% Year-on-Year.
China's national statistics bureau reported on April 16 that daily AI token invocation volume crossed 140 trillion as of March 2026, representing growth of over 40% compared to the end of 2025. National Statistics Deputy Director Mao Shengyong disclosed the figure at a State Council information office briefing, characterizing it as a staged breakthrough in the commercialization and scaled operation of AI in China. The figure encompasses the full domestic AI ecosystem including cloud API calls, on-device inference, and enterprise internal usage, and signals sustained momentum in AI deployment across Chinese industry.
Source: 国家统计局 / 证券时报 | 2026-04-16
8. It 石智航 Raises $450M+ — China's Largest Single Round in Embodied AI History.
It 石智航 announced on April 16 it has closed over $450 million in a Pre-A round, setting a new record for the largest single financing in China's embodied AI sector. The round was co-led by Hillhouse Capital's venture arm and Sequoia China, with participation from Meituan Dragonball, CIC Capital, and other institutional investors. Funding will go toward building a general-purpose embodied AI foundation model branded AWE and recruiting top talent. The company's next financing will focus on producing physical robots capable of real work at real scale. The round size signals significant VC conviction that general-purpose humanoid robot foundation models — rather than task-specific robots — represent the viable commercial path in embodied AI.
Source: 36氪 | 2026-04-16