AI Tech digest — April 17, 2026

The AI Tech Digest is evolving. We’re shifting from industry news to focusing on what matters to builders: new tools, trending open-source projects, and the best from the AI developer community. If you want CEO drama, you’re in the wrong place.

Model Releases

Claude Opus 4.7: The Model That Runs While You Grab Coffee

Anthropic released Claude Opus 4.7 on April 16, and the headline feature isn’t a benchmark number. The model can now execute complex, multi-step projects for hours without losing context or needing constant supervision. Hand it a campaign audit, a competitive research sweep, or a slide deck, and come back to a finished first draft.

87.6% on SWE-bench Verified, 94.2% on GPQA, and a 14% improvement in multi-step agentic reasoning with a third fewer tool errors compared to Opus 4.6. Vision resolution jumped to 3.75 megapixels (3x higher), enough to reliably read dense charts, diagrams, and screenshots that earlier models mangled.

New developer-facing features:

xhigh effort level: a new tier above high for when you need maximum reasoning depth
/ultrareview command in Claude Code: for deep code review passes
Task budgets in public beta: set token/time budgets for agentic loops so runaway agents don’t drain your wallet
Auto mode for Max subscribers: the model picks its own effort level based on task complexity

Pricing stays the same at $5/$25 per million tokens (input/output), with a promotional 7.5x premium multiplier until April 30th. Replit reports it pushes back on bad instructions instead of blindly agreeing, a qualitative improvement that doesn’t show up in benchmarks but matters enormously in practice.

Anthropic announcement → | Benchmark breakdown →

OpenAI GPT-Rosalind: A Frontier Model for Drug Discovery

OpenAI introduced GPT-Rosalind on April 16, a domain-specific reasoning model built for biology, drug discovery, and translational medicine. Named after Rosalind Franklin, it’s optimized for the specific workflows scientists use: literature synthesis, hypothesis generation, sequence-to-function interpretation, experimental planning, and multi-step data analysis.

On benchmarks, it outperforms GPT-5.4 on 6 of 11 LABBench2 tasks, with the most notable improvement on CloningQA, which requires end-to-end design of DNA and enzyme reagents for molecular cloning. In a blind evaluation by Dyno Therapeutics, best-of-ten GPT-Rosalind submissions ranked above the 95th percentile of human experts on RNA sequence-to-function prediction.

The companion Life Sciences research plugin for Codex (free, available on GitHub) connects to 50+ public multi-omics databases, literature sources, and biology tools. Access to GPT-Rosalind itself is gated behind a trusted-access program for qualified enterprise customers, with no token consumption during the research preview.

This is a clear signal that frontier AI is moving beyond general-purpose models into domain-specific verticals. Life sciences is the first target; expect similar specialized models for law, finance, and engineering.

OpenAI announcement → | Codex plugin →

OpenAI GPT-5.3 Instant Mini Ships as New Fallback

On April 9, OpenAI quietly released GPT-5.3 Instant Mini as the new fallback model in ChatGPT, the model you hit after exhausting your GPT-5.3 Instant rate limits. It won’t appear in the model picker, but it’s there when you need it.

Compared to the previous GPT-5 Instant Mini fallback, it delivers more natural conversation, stronger writing, and better contextual awareness across chats. Separately, OpenAI added a new $100/month Pro tier with 10x more Codex usage than Plus, and rebalanced the Plus tier toward more sessions per week rather than longer single sessions.

ChatGPT release notes →

Tools & Platforms

OpenAI Agents SDK: The Next Evolution

OpenAI released a major update to the Agents SDK on April 15 that addresses one of the biggest pain points in agent development: infrastructure. The new version provides:

Model-native harness with configurable memory, sandbox-aware orchestration, Codex-like filesystem tools, and standardized integrations with MCP, skills, shell execution, and file editing
Native sandbox execution: agents run in controlled environments with built-in support for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel sandboxes
Manifest abstraction: portable workspace definitions that work across local prototyping and production deployment, with mounts for S3, GCS, Azure Blob, and Cloudflare R2
Built-in snapshotting: if a sandbox container dies, the agent’s state is restored in a fresh container from the last checkpoint

The key architectural decision is that harness and compute are separated. Credentials stay out of the sandbox where model-generated code runs. This is the right security model for production agents: prompt injection is assumed, not prevented.

Available now in Python via the API with standard token-based pricing. TypeScript support is planned.

OpenAI announcement →

Codex Gets an In-App Browser and Can Click Your Mac

The Codex desktop app (v0.121.0, April 16) is expanding beyond “coding agent” into “general-purpose work agent.” Major additions:

In-app Atlas browser: open local or public pages, comment directly on rendered output, and ask Codex to address page-level feedback
macOS computer use: Codex can now see, click, and type in native Mac apps. Think automated testing, GUI-only bug reproduction, and simulator flows
gpt-image-1.5: image generation baked into the workspace
111 new plugin integrations
Thread automations: schedule follow-ups that wake up the same conversation on a timer, preserving full context
Richer PR review: inspect GitHub PRs in the sidebar, review diffs, ask Codex to address feedback inline
SSH remote connections (alpha), multiple terminals, multi-window support, Intel Mac support

The computer use feature is explicitly chasing Anthropic’s Claude Code, and the breadth of new integrations (Outlook shared mailboxes, calendar management, sidebar file previews) positions Codex as a full desktop productivity agent.

Codex changelog →

Google Chrome Skills: Save Your Best Prompts as One-Click Tools

Google rolled out Skills in Chrome on April 14, a feature that solves a real workflow annoyance: reusing good prompts. Save any Gemini prompt as a Skill once, then trigger it with a forward slash on any page you’re viewing. Skills can operate across multiple tabs simultaneously.

Google is also shipping a starter library of prebuilt Skills: protein macro calculators, side-by-side product comparisons, ingredient breakdowns, and PDF scanners. Confirmation prompts protect sensitive actions like calendar events and emails.

Available on Mac, Windows, and ChromeOS, English-US only for now. This is Chrome’s moat play: the browser as an AI operating system, with prompts as installable software.

Google blog → | Ars Technica →

Perplexity Personal Computer: A Desktop Agent That Never Sleeps

Perplexity launched Personal Computer on April 16, a Mac desktop app that puts their AI agent directly on your machine, running full-time. Unlike the cloud-based Perplexity chat, this agent works across your local files, native apps, connectors, and the web in a single orchestrated system.

Drop it on a Mac mini and it stays available 24/7, ready to handle tasks that are too messy for a chatbot and too repetitive to keep doing by hand. Currently available for Max subscribers on Mac, with waitlist users getting early access as the rollout expands.

The desktop agent space is getting crowded fast (Codex, Claude Code, Perplexity, Google’s Project Astra). The differentiator will be which agent can actually complete multi-hour workflows without human intervention, and Opus 4.7’s stamina improvements suggest Anthropic currently leads on that front.

Perplexity announcement →

Open Source & GitHub

April 2026 has been strong for new open-source AI projects gaining real traction (from Fazm’s tracking):

Repository	Stars (2 weeks)	Language	What It Does
google/adk-python	8,200+	Python	Google’s Agent Development Kit for building multi-agent systems
meta-llama/llama-stack	6,400+	Python	Unified deployment stack for Llama 4 family models
openai/codex-cli	5,800+	TypeScript	Terminal-native coding agent with sandboxed execution
block/goose	4,900+	Rust	Local-first AI agent framework with extensible toolkit

Google’s ADK is particularly notable. It provides a structured framework for building multi-agent systems with built-in tool use, memory, and orchestration. Block’s Goose (in Rust) takes the local-first approach, keeping everything on your machine without cloud dependencies.

OpenClaw Still Dominating: 346K+ Stars

OpenClaw remains the fastest-growing open-source project in GitHub history at 346K+ stars, with 162 production-ready agent templates now listed in the awesome-openclaw-agents directory across 19 categories. The latest release (v2026.4.14) focused on reliability, fixing browser automation classification issues, Matrix streaming, and Telegram bot verification.

Quick Hits

Adobe Firefly AI Assistant: A conversational agent that orchestrates multi-step workflows across Photoshop, Illustrator, and Premiere Pro. Claude integration confirmed. Public beta coming soon. Adobe →
Codex CLI 0.121.0: Marketplace plugin installs, TUI reverse search (Ctrl+R), memory controls, secure devcontainer profiles with bubblewrap, and expanded MCP APIs. GitHub →
ChatGPT Outlook integration: Shared mailbox and shared calendar actions for teams, including reading, moving, and sending from shared mailboxes. Release notes →
Canva AI 2.0 (research preview): Turns Canva into a conversational agent running on proprietary Proteus and Lucid Origin models. Generates multi-channel ad campaigns from a sentence, pulls context from Slack/Gmail, and maintains persistent brand memory.

What to Watch

TypeScript support for the new Agents SDK: Python shipped first, but the JS/TS community is the larger market for agent tooling. This will be the real test of adoption.
Claude Mythos public access timeline: Still gated behind Project Glasswing with ~50 organizations. At 93.9% SWE-bench and 94.6% GPQA Diamond, it’s the most capable model that most developers can’t touch.
Grok 5 from xAI: Expected in the Q2 window. xAI has been quiet since the Grok 4 release cycle, which usually means something big is coming.
Google I/O (May 2026): Expect Gemini 3.x updates, deeper Chrome Skills integration, and likely new Agent Development Kit features.
Desktop agent convergence: Codex, Claude Code, Perplexity, and Google are all racing to build the “always-on AI assistant that lives on your computer.” The winner will be determined by reliability on multi-hour tasks, not benchmarks.