OpenAI Launches GPT-5.6, AWS Strengthens MCP Infrastructure, and Anthropic Expands Government AI
Welcome to the latest MCP Weekly. Nearly every major announcement this week focused on the infrastructure surrounding AI: lower operating costs, standardized connectivity, stronger governance, and enterprise deployment.
OpenAI, AWS, and Anthropic each addressed a different part of this bigger story. GPT-5.6 pushes down inference costs, AWS simplifies secure MCP connectivity, and Anthropic continues expanding into regulated environments. Taken together, they point toward an ecosystem focused on making AI practical inside enterprise infrastructure.
TL;DR
OpenAI launched GPT-5.6 and GPT-Live, introducing faster, more cost-efficient models with new multi-agent and real-time voice capabilities. AWS strengthened enterprise MCP adoption with OAuth authentication and an MCP server for the Registry of Open Data. Anthropic (https://claude.com/blog/cowork-web-mobile) expanded Claude across mobile, enterprise, and FedRAMP High government deployments while introducing Claude Reflection for AI usage analytics. Microsoft Research released Flint an open-source visualization language with MCP integration for AI agents.
Major Updates of the Week
OpenAI Launches GPT-5.6 and GPT-Live
OpenAI launched GPT-5.6 with three models: Sol (the full-power version), Terra (balanced middle ground), and Luna (optimized for cost). The new family includes Programmatic Tool Calling, which lets models run lightweight programs in memory to filter results before they leave the system. There’s also an ultra mode that coordinates four parallel agents simultaneously. The company hit 53.6 on the Agents’ Last Exam benchmark and 80 on coding tasks using max reasoning.
Separately, OpenAI shipped GPT-Live, a full-duplex voice model that listens and speaks at the same time. You can interrupt it without starting over. The system handles real-time conversation locally, then delegates the heavy reasoning to background models like GPT-5.5. It’s rolling out to Go, Plus, and Pro users now, with an API for developers coming soon.
AWS Adds OAuth Security to MCP Infrastructure
AWS added native OAuth support to the AWS MCP Server, so developers can connect agents using existing AWS credentials instead of long-lived API keys. The system supports both interactive flows for developers and headless authorization for automated agents. Everything logs to CloudTrail.
AWS also launched an MCP server for the Registry of Open Data, which hosts over 1,100 datasets across satellite imagery, climate science, and life sciences. Researchers can now search and sample data through natural language using compatible AI assistants. It’s open source under Apache 2.0.
Anthropic Expands Claude Across Enterprise and Government
Anthropic rolled out Claude Cowork to mobile and web alongside desktop. Tasks keep running in the background even when your device is offline. Scheduled actions execute continuously, and you control approvals. The company doubled usage limits through August 5.
Claude Reflection launched in beta to help users track and understand their interaction patterns. The dashboard breaks down usage across 1, 3, 6, or 12 months using the 4D AI Fluency Framework: Delegation, Description, Discernment, and Diligence. It includes quiet hours and break reminders built in with MIT Media Lab and Boston Children’s Hospital.
Anthropic also deployed Claude Code and Cowork in public beta for the government through FedRAMP High. The system runs in an authorized environment and stores all conversation histories locally on agency devices. Administrators get granular spending controls, hash-chained audit logs, and two-person approval on sensitive operations.
Microsoft Introduces Flint for AI Agent Visualization
Microsoft Research released Flint, an open-source language that helps agents generate reliable charts without verbose parameters. Instead of tweaking low-level settings, Flint acts as a semantic compiler. You describe what you want, and it automatically optimizes scales, spacing, and color schemes based on data type. The same specification compiles to Vega-Lite, ECharts, or Chart.js depending on what you need.
Flint includes an MCP server, so agents can generate and validate charts directly in chat. Benchmarks show it beats hand-written Vega-Lite output across multiple models.
Other Updates
| Vendor / Product | Key Action / Feature | Significance |
| Ollama | Secured $88M Series B led by Benchmark and Theory Ventures | Now serves 8.9M developers and 85% of Fortune 500. Cloud token volume doubles monthly. Capital focused on hybrid inference and day-one compatibility with new open models. |
| SpaceX AI Grok 4.5 | Launched flagship model optimized for coding and agentic workflows | Delivers 80 tokens per second with 4.2x token efficiency over Opus 4.8. Available via SpaceXAI API and Cursor. |
| Cognition and Claude | Integrated Claude Fable 5 into Devin autonomous software engineer | Scored 30% on Frontier Code benchmark, tripling prior Opus baseline. Enables 8-hour autonomous execution windows. |
| Anthropic and UST | Strategic partnership integrating Claude into hardware validation and enterprise systems | UST designated Global Premier Partner. Claude Code embedded in iDEC platform for silicon regression testing, care management, network operations, and banking. |
| AAIF Momentum Report | Linux Foundation published ecosystem analysis of 116 open-source projects | Frameworks and Infrastructure lead with 1.5M stars. Security debt jumped 2.6x over six months. Single-entity control remains a risk for LangChain and Milvus. |
| Wiz MCP | General availability of security platform integration via MCP | Exposes Wiz Security Graph and AI skills for vulnerability triage and compliance auditing. Real deployments at Cohere, Grammarly, Infosys. |
| Reuters MCP Server | Launched open-source MCP for AI-native journalism integration | Agents can search, retrieve, and download subscribed Reuters content. Enables automated multimedia assembly and content curation. |
| Anthropic Marketing Ops | Automated multi-platform workflows using Claude Cowork | Compressed two-day reporting loop to two hours. Uses modular skills for routing, proofreading, and auditing. |
My Thoughts: The Infrastructure Layer
Something interesting happened this week. Every major player released something focused on cost and integration instead of raw capability. OpenAI emphasized lower-cost inference and real-time orchestration. AWS focused on secure authentication and standardized access through MCP. Anthropic expanded into regulated government environments while giving enterprises better visibility into AI usage. Even Microsoft’s Flint is ultimately an infrastructure project, making it easier for agents to generate reliable visualizations without developers managing low-level rendering details.
These announcements reflect a broader shift in the market. Frontier models are becoming increasingly capable across the board, which means competitive advantage is moving up the stack. Organizations now care less about marginal benchmark gains and more about deployment, governance, interoperability, and operating costs. Those are the factors that determine whether AI becomes part of everyday business operations or remains an interesting prototype.
The MCP ecosystem reflects that same transition. AWS adding OAuth, Microsoft embedding MCP into Flint, and Anthropic continuing to expand enterprise integrations all point in the same direction. Successful standards eventually fade into the background because developers stop debating them and start building with them. MCP increasingly feels like it’s entering that phase. It is becoming part of the infrastructure layer that enterprise AI simply assumes is there.
