OpenAI GPT-5.4 Thinking Model Launches

Published: March 22, 2026 06:00 (Asia/Shanghai)
Coverage: 2026-03-21 18:00 — 2026-03-22 06:00

📰 Top Stories

1. OpenAI GPT-5.4 Thinking Model Launches for Complex Professional Tasks

Source: OpenAI Release Notes
Time: March 21

OpenAI released GPT-5.4 Thinking, designed for difficult multi-step reasoning, long context, and tool-heavy workflows. The model achieves 83.0% on the GDPVal benchmark, placing it at or above human expert level on economically valuable tasks. GPT-5.4 Thinking is optimized for tasks requiring coherent multi-step workflows with lower latency, bringing stronger coding capabilities to the flagship frontier model.

• Key Features: Multi-step reasoning, long context handling, improved coding capabilities
• Performance: 83.0% on GDPVal benchmark, human expert-level performance
• Use Cases: Complex professional tasks, tool-heavy workflows, reduced back-and-forth

Read More

2. Anthropic Claude Code 2.1.76: MCP Elicitation and Enterprise Features

Source: Anthropic
Time: March 21 (15 hours ago)

Anthropic released Claude Code 2.1.76 with MCP elicitation support, allowing MCP servers to request structured input mid-task via interactive dialogs. New features include Elicitation and ElicitationResult hooks, sparse worktree paths for large monorepos, configurable session quality surveys for enterprise admins, and numerous stability fixes across tools, SLAs, and integrations.

• Key Features: MCP elicitation, Elicitation hooks, sparse worktree support, session quality surveys
• Enterprise: Configurable feedback survey rate for admins
• Fixes: Deferred tools, voice mode, Windows compatibility, auto-compaction circuit breaker

Read More

3. Anthropic Claude Persistent Agent Thread for Pro/Max Plans

Source: Anthropic
Time: March 17

Claude released a persistent agent thread for Pro and Max plans, enabling users to manage tasks in Cowork from mobile and desktop. The feature allows Claude Desktop or Claude for iOS/Android users to maintain persistent conversations for task management. Rolling out to Max plans first, then Pro plans over the following days.

• Availability: Pro and Max plans (research preview)
• Platforms: Claude Desktop, iOS, Android
• Integration: Cowork task management

Read More

4. Morgan Stanley: Major AI Breakthrough Expected in H1 2026

Source: Fortune
Time: March 13

Morgan Stanley warns that a transformative AI leap is imminent in the first half of 2026, driven by unprecedented compute accumulation at US AI labs. The bank cites scaling laws holding firm, with OpenAI’s GPT-5.4 Thinking already scoring 83.0% on GDPVal. Executives at major AI labs tell investors to brace for progress that will “shock” them.

• Prediction: Major AI breakthrough in H1 2026
• Driver: 10x compute scaling at AI labs
• Impact: Progress expected to surprise even Wall Street analysts

Read More

5. Microsoft Copilot Enhances Connectors and Governance Controls

Source: Microsoft 365
Time: March 21

Microsoft announced Copilot updates including larger connector support reaching Amazon S3 and more frequent ServiceNow refreshes. New agent lifecycle features allow administrators to block risky agents and remove inactive or ownerless agents via the Microsoft 365 admin center. Enhanced dashboards in Viva and Copilot Dashboard show adoption metrics and agent actions across groups.

• Connectors: Amazon S3 support, improved ServiceNow integration
• Governance: Agent lifecycle rules, risk blocking, inactive agent cleanup
• Analytics: Enhanced Viva and Copilot Dashboard metrics

Read More

6. US States Advance AI Legislation: Child Safety and Deepfake Protection

Source: State Legislatures
Time: March 17-21

Multiple US states advanced AI-related legislation this week. Hawaii and Idaho moved forward with addictive algorithm bans, chatbot safety bills, deepfake protection, and social media age verification. Virginia passed four AI bills including AI fraud/abuse frameworks and independent verification organizations. Georgia is considering chatbot disclosure and AI healthcare decision restrictions.

• Hawaii: Addictive algorithm, chatbot safety, deepfake, age verification bills
• Idaho: Similar child safety and deepfake protection measures
• Virginia: AI fraud/abuse, independent verification, minor protection bills (passed)
• Georgia: Chatbot disclosure, AI study committee, healthcare AI restrictions

Read More

7. Google Gemini API Updates: Built-in Tools + Function Calling Combination

Source: Google AI for Developers
Time: March 18

Google released new Gemini API features enabling built-in tools to be used with custom function calling tools in a single API call. Google Maps grounding now supports Gemini 3 models. The update streamlines complex workflows by combining Google’s native tools with developer-defined functions.

• New Feature: Built-in tools + function calling combination
• Maps Grounding: Now available on Gemini 3 models
• Benefit: Single API call for complex multi-tool workflows

Read More

📊 Today’s Stats

🔮 Next Briefing Watch

• GPT-5.4 Thinking user feedback and adoption rates
• Anthropic Claude Code MCP elicitation enterprise deployments
• US state AI legislation final passage outcomes
• Morgan Stanley AI breakthrough predictions validation

End of Briefing

AI News Briefing updates automatically every 12 hours (06:00 / 18:00)
Subscribe for updates at CinaGroup Blog