What moved in AI today

New flagship model improves on software engineering, long-running coding tasks, and high-resolution image understanding. Leads SWE-bench Verified at 87.6% on agentic coding tasks — 13 points ahead of GPT-5.4.

New product for creating visuals, prototypes, slides, and one-pagers directly inside Claude conversations. Mobile app now renders fully interactive apps and live charts in-conversation.

Anthropic silently removed Claude Code from the $20/month Pro tier for ~2% of new users in an A/B test. Existing subscribers were unaffected. Reversed publicly after backlash, but signals ongoing pricing pressure.

New general-purpose model with unusually strong computer security capabilities. Available in preview at red.anthropic.com.

67% faster /resume on sessions over 40MB. Faster MCP startup with multiple stdio servers. Improved session resume, model persistence, safer auth handling.

GPT-4o removed from all plans effective April 3, 2026. GPT-5.4 is now the flagship, with GPT-5.4 mini rolling out to Free and Go users.

New reasoning model optimized for biology, drug discovery, protein engineering, and genomics. Targets research institutions and pharma. Deeper tool use for scientific workflows.

GPT-5.4 variant tuned for defensive cybersecurity — threat analysis, incident response, and detection. Expanded API access for security teams.

Computer use (macOS apps: see, click, type) · In-app browser with inline commenting · 90+ new plugins (Atlassian, CircleCI, GitLab, MS Suite, Neon, Remotion, Render) · Memory preview (preferences persist across sessions) · Scheduled tasks (Codex can wake itself up to continue multi-day work) · Proactive suggestions at session start.

Teams can build shared agents for complex, persistent tasks — described as "an evolution of GPTs."

Launched April 22. Reads across all Workspace apps by default — Docs, Gmail, Drive, Meet, Calendar. Admin controls let organizations restrict which data sources Gemini accesses.

Google Cloud committed $750M in resources and incentives to partners building agentic AI with Gemini Enterprise. Now integrates agents from Adobe, Atlassian, Deloitte, Oracle, Salesforce, ServiceNow, Workday, and more.

New agent for extended research across the web or custom sources. Gemini 3.1 Pro leads independent benchmarks as the strongest all-around model right now: 78.80% SWE-bench Verified, 94.3% GPQA Diamond.

Shipped quietly April 17. Adds native video understanding, AI slide creation, Speech-to-Text and Text-to-Speech APIs. Available to SuperGrok Heavy ($300/mo); standard SuperGrok can see but not yet access.

xAI's autonomous computer-use agent can operate apps, click, type, fill forms, and execute multi-step workflows without human intervention across a full desktop.

Rust-built messaging app with E2E encryption, voice/video calls, disappearing messages, and Grok integrated natively.

Despite prior pledges, Grok continues to generate non-consensual sexual deepfakes, triggering renewed regulatory scrutiny and public backlash.

Fundamental architecture shift: the Agents Window is now primary. Run many agents in parallel across local worktrees, cloud, and remote SSH — switch to IDE anytime. Design Mode lets you annotate browser UI elements for precise agent feedback. Agent Tabs allow side-by-side or grid view of multiple chats. New commands: /worktree (isolated git worktree) · /best-of-n (parallel multi-model runs). Real-time RL deploys improved checkpoints every 5 hours from real usage. Bugbot resolution rate nearing 80% — 15pp ahead of closest AI code review competitor.

Copilot Researcher now uses OpenAI GPT to draft, then Anthropic Claude to verify accuracy and citations. Microsoft plans to stop promoting specific model names — routing will be invisible and task-driven.

Bundles M365 Copilot + Work IQ + Agent 365 + enterprise security and identity governance. Wave 3 of Microsoft 365 Copilot generally available May 1, 2026.

Microsoft actively testing persistent background agents that run continuously without user triggers, responding to enterprise demand for autonomous, always-on workflows.

Major leadership transition. Ternus inherits the task of rebuilding Apple's AI strategy, which has lagged significantly behind Google, OpenAI, and Anthropic.

Long-awaited Siri rewrite expected in spring 2026: more conversational, capable of multi-step tasks. Apple is relying on Google Gemini as the backend for the major upgrade.

Apple accelerating three AI wearables — all built around Siri. Smart glasses confirmed with unique camera design and multiple frame styles.

Unrevealed Apple Intelligence features spotted in iOS 27 beta code — details not yet public.

Scout (April 2): 109B total / 17B active, MoE 16 experts, 10M token context window.
Maverick (April 5): 400B MoE, 128 experts — open weights, optimized for diverse task types.

Launched April 8. Natively multimodal reasoning model with visual chain-of-thought, tool use, and multi-agent orchestration built in. Meta's most capable model to date.

Reports: Meta developing two closed-source frontier models — Avocado (LLM) and Mango (multimedia). Developer community skeptical; some see it as "closing the gates once there's something worth protecting."

Agentic AI has crossed the chasm. 52% of employees are already using agents. Top use cases: code development, legal, financial, supply chain, R&D, and cybersecurity. But 79% of organizations struggle to scale — top barriers: data security (36%), lack of AI talent (25%).

AI's workforce impact has moved from prediction to reality — hitting early-career developers first. Senior developer headcount is growing, while junior roles evaporate.

All AI in hiring, evaluation, and performance monitoring is now "high-risk" under EU law. Mandatory: risk assessments, bias testing, human oversight, transparency disclosures. Fines: up to €15M or 3% of global annual revenue.

Berkeley researchers built an automated agent that achieved near-perfect scores on SWE-bench, WebArena, OSWorld, GAIA, and four others by exploiting structural flaws. Meanwhile, Princeton's HAL scaffold adds +30 absolute points to GAIA — larger than the gap between most frontier model releases. Raises fundamental questions about what benchmarks measure.