Weekly AI Intelligence Briefing

Sia Reads
What Matters

Week of April 20, 2026

10 stories tracked

RELEASE2026-04-23

GPT-5.5 can navigate your desktop, switch between tools, and nearly doubles Claude Opus 4.7 on FrontierMath — OpenAI's biggest leap in months

9
Read more

OpenAI launched GPT-5.5 on April 23, an agentic model that natively controls desktop applications — clicking buttons, typing text, and orchestrating multi-step workflows without plugins. It scores 39.6% on FrontierMath Tier 4 (vs Claude Opus 4.7's 22.9%) and 82.7% on Terminal-Bench 2.0. Token generation is 20% faster than GPT-5.4 at the same latency. Rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex.

RELEASE2026-04-24

DeepSeek ships V4 Flash and Pro under MIT license — 1M-token context built in from the ground up, and pricing that makes frontier labs nervous

8
Read more

DeepSeek released V4-Pro (1.6T total params, 49B active) and V4-Flash (284B total, 13B active) on April 24, both with native million-token context and MIT licensing. Flash runs at $0.14/$0.28 per million tokens — an order of magnitude cheaper than Western frontier models. Performance trails GPT-5.4 and Gemini 3.1 Pro by an estimated 3-6 months, but leads all other open-source models on agentic coding and reasoning.

RELEASE2026-04-22

Google splits training and inference into separate silicon for the first time — the TPU 8i is built to serve millions of concurrent agents cheaply

7
Read more

At Cloud Next '26, Google announced the TPU 8i, a dedicated inference processor with 384MB of on-chip SRAM (3x its predecessor) and 80% better inference throughput. By separating training (Ironwood) and inference (8i) into distinct chips, Google is making a bet that serving agents at scale is a fundamentally different silicon problem than training models. Both chips ship later this year. Google is supplementing, not replacing, its Nvidia-based infrastructure.

RELEASE2026-04-21

An AI agent that reads papers, finds datasets, trains models, and evaluates results — ml-intern pushed Qwen3-1.7B from 10% to 32% on GPQA in 10 hours

7
Read more

Hugging Face released ml-intern on April 21, an open-source agent built on smolagents that automates the full post-training pipeline: literature review on ArXiv, dataset discovery on the Hub, training execution, and iterative evaluation. In its launch demo, it took a Qwen3-1.7B base model from ~10% to 32% on GPQA in under 10 hours, outperforming Claude Code's 22.99% on the same task. Available as CLI and web app, with $1,000 in GPU credits for early users.

RELEASE2026-04-17

Claude can now build presentations, prototypes, and landing pages from a text prompt — and it reads your codebase to match your design system automatically

6
Read more

Anthropic launched Claude Design on April 17, a new Anthropic Labs product powered by Opus 4.7 that generates polished visual work — slides, prototypes, one-pagers, landing pages — from natural language. During onboarding, it reads your codebase and design files to build a custom design system, then applies your colors, typography, and components to every project. Export to Canva, PDF, PPTX, or standalone HTML. Figma stock dropped 7% on the news.

HOT TAKE2026-04-23

A caching bug, a reasoning-effort default swap, and a verbosity tweak all shipped within weeks — Anthropic published a postmortem admitting Claude got worse

6
Read more

Anthropic published a postmortem on April 23 revealing three distinct changes that degraded Claude Code, Agent SDK, and Cowork between March and April. A cache optimization on March 26 cleared session data every turn, making Claude forgetful. A reasoning-effort default change made it lazier on hard tasks. A verbosity prompt tweak on April 16 was reverted four days later. Anthropic's takeaway: 'This was the wrong tradeoff.' Builders who noticed Claude getting worse weren't imagining it.

NEWS2026-04-20

Adobe made MCP the default agent protocol for Commerce, Microsoft shipped Fabric MCP to GA — the protocol layer just graduated from dev tool to enterprise infrastructure

6
Read more

In a single week, Adobe announced MCP as the sanctioned protocol for AI agents to read and act on Commerce data (at Summit 2026), and Microsoft shipped Fabric Local MCP to general availability with Fabric Remote MCP in preview. A $20B CX incumbent and the dominant enterprise cloud both picked the same agent protocol. MCP is no longer just a developer convenience — it's becoming the standard wiring for enterprise agentic systems.

RELEASE2026-04-22

ChatGPT now has persistent agents that keep working when you close the tab — schedule them, deploy them in Slack, and let them handle multi-step workflows

6
Read more

On April 22, OpenAI launched workspace agents in ChatGPT — shared agents powered by Codex that handle complex, multi-step tasks autonomously. They run in the cloud with access to files, code, tools, and memory, and can be scheduled or deployed into Slack. Think of them as the successor to custom GPTs, but with actual compute and state. Available in research preview for Business, Enterprise, Edu, and Teachers plans, free until May 6.

NEWS2026-04-21

Anthropic quietly moved Claude Code to the $100 Max plan for new signups, got caught within hours, and walked it back — but the signal is clear

5
Read more

On April 21, Anthropic silently updated its docs to remove Claude Code from the $20/month Pro plan, gating it behind the $100 Max tier. No email, no changelog — just a quiet support page edit. The developer community noticed within hours and the backlash was immediate. Anthropic reversed the change within 24 hours, calling it a test on ~2% of new signups. But the documentation had been universally updated, not just for test users. The pricing pressure on AI coding tools is real.

RELEASE2026-04-22

Google unified agent development into a single platform with Agent Designer, long-running agents, and a $750M partner fund to accelerate adoption

5
Read more

At Cloud Next '26 on April 22, Google unveiled the Gemini Enterprise Agent Platform — a unified environment for building, scaling, governing, and optimizing AI agents. It ships with Agent Designer, an Inbox for managing agent activity, long-running agents, and Skills for composable capabilities. Google also committed $750M to its 120,000-member partner ecosystem for agentic AI development. Third-party agents from Adobe, Atlassian, Salesforce, ServiceNow, and others are available at launch.