OpenAI Launches GPT-5.6, Claude Tag Goes Live, and Microsoft Expands Enterprise AI.
In this week’s MCP Weekly, we explore a landscape defined by significant model debuts, new enterprise tools, a major intellectual property dispute, and increasing state influence over high-end AI deployment.
TL;DR
OpenAI launched the GPT-5.6 family (Sol, Terra, Luna) under a government-restricted rollout, with all three models rated “High” risk in cybersecurity and biological domains. OpenAI also unveiled Jalapeño, its first custom inference chip, expanded Daybreak with GPT-5.5-Cyber across 30 million code commits, and published research showing Codex now handles 99.8% of internal output tokens with non-developer adoption up 137x since August 2025.
Anthropic formally accused Alibaba of running 28.8 million fraudulent exchanges to steal its model capabilities, its largest recorded distillation attack. The U.S. government partially resolved a two-week access freeze by clearing Claude Mythos 5 for roughly 100 trusted partners, while Fable 5 stays blocked. Anthropic launched Claude Tag, a multiplayer Slack agent built on Opus 4.8, and expanded Claude Desktop to AWS, Google Cloud, and Microsoft Foundry. Microsoft upgraded Copilot in Excel with institutional data connectors and audit-ready finance workflows. Cursor embedded its SDK inside Notion, and a new study found that 63% of top benchmark scores from frontier models came from data leakage rather than genuine reasoning.
OpenAI Launches GPT-5.6 Family
OpenAI has debuted its GPT-5.6 model suite, comprising the premier Sol, the mid-range Terra, and the high-speed, cost-effective Luna. Per U.S. government directives, initial access is restricted to vetted partners prior to a broader release. These models are classified with a “High” risk level in biological and cybersecurity sectors according to OpenAI’s Preparedness Framework.
The Sol model achieves a 91.9% score on Terminal-Bench 2.1. To prevent hazardous output, safety is maintained by a real-time activation classifier capable of halting generation. API pricing is established at $1 per million input tokens for Luna, reaching $5 for Sol.
Other OpenAI Updates
| Vendor / Product | Key Action / Feature | Significance |
| OpenAI Jalapeño | First custom inference chip, co-developed with Broadcom and Celestica, reaching tape-out in nine months using AI-assisted design. Deployments planned for late 2026 in gigawatt-scale data centers with Microsoft. | Moves OpenAI into custom silicon, targeting lower token costs and higher throughput beyond what general-purpose hardware delivers. |
| OpenAI Daybreak / GPT-5.5-Cyber | Full release of GPT-5.5-Cyber, scoring 85.6% on CyberGym. Codex Security has scanned over 30 million commits and verified more than 500,000 fixes. The Patch the Planet initiative funds remediation of open-source projects including cURL, Go, and Python. | Shifts AI security from finding vulnerabilities to automated patching at scale, with gated partner access for commercial deployments. |
| OpenAI Codex Research Paper | Internal data shows Codex accounts for 99.8% of weekly output tokens at OpenAI. Non-developer adoption is up 137x since August 2025. Legal, Finance, and Recruiting fully shifted to agentic workflows by April 2026. | The clearest data yet showing that agentic AI has moved from developer tool to company-wide operating layer. |
Anthropic Debuts Claude Tag
Claude Tag has entered beta for Team and Enterprise users. Powered by Opus 4.8, this shared Slack agent develops cross-channel context over time, allowing for the delegation of multi-day tasks that proceed without continuous prompting. Administrators are provided with controls for memory boundaries, channel permissions, and token limits. Notably, Anthropic reports that 65% of the code produced by its own product team is now generated via Claude Tag.
Other Anthropic Updates
| Vendor / Product | Key Action / Feature | Significance |
| Claude Tag: Agent Identity Access Model | Claude Tag uses channel-scoped service accounts rather than individual user credentials. Admins set permissions per channel, memory stays isolated between channels, and every action is logged in audit trails. | Solves the core governance problem of multiplayer agents: who is acting, on what data, and with whose permissions. |
| Claude Desktop on AWS, Google Cloud, and Microsoft Foundry | Full Claude Desktop (Chat, Cowork, and Code) now deploys inside enterprise cloud accounts. Inference stays within the customer’s chosen region. Supports SSO, MDM templates, and a native Microsoft 365 data connector with GovCloud support. | Removes the need for separate vendor contracts and lets regulated enterprises keep all data within their existing cloud boundaries. |
| Anthropic Human-Agent Teams Framework | Published guide on running multiplayer human-agent teams: working in public, role specialization, doer-verifier model pairs, and batching agent questions to reduce interruptions. | Practical operational guidance for teams moving from single-user prompts to ambient, always-on agent collaboration. |
Financial Enhancements for Microsoft Copilot in Excel
Microsoft has integrated institutional data connectors into Excel Copilot, enabling access to live feeds from FactSet, PitchBook, CB Insights, Morningstar, S&P Global, and Daloopa. Finance professionals can now utilize OneDrive to store repeatable workflows as standard markdown files. Before implementation, a pre-execution plan view details intended cell and formula modifications. Transparency is maintained via Excel's change tracking panel, which logs cell-level attribution for every edit. Additionally, the Microsoft Marketplace will feature third-party skills from Ramp, Vena, and LSEG starting in Q3 2026.
Policy Watch
Anthropic Accuses Alibaba of Industrial-Scale Distillation Attack
Anthropic sent a formal letter to the U.S. Senate Banking Committee accusing Alibaba of using roughly 25,000 fraudulent accounts to extract 28.8 million exchanges from its models between April 22 and June 5. The goal was to copy Anthropic’s reasoning capabilities into smaller, cheaper models. Anthropic calls this its largest recorded distillation attack, following earlier campaigns attributed to DeepSeek, Moonshot, and MiniMax. The letter asks Congress and the White House to establish unified cross-industry defenses against automated capability extraction.
U.S. Government Partially Clears Claude Mythos 5
The U.S. Commerce Department has authorized Anthropic to release Claude Mythos 5 to roughly 100 vetted companies and federal agencies, ending a two-week access freeze triggered by export control directives. Fable 5 stays blocked. Anthropic co-founder Tom Brown led the negotiations, and the company continues its lawsuit challenging the Department of Defense’s earlier designation of Anthropic as a supply chain risk. The two stories together paint a difficult week for Anthropic on the regulatory front, fighting IP theft on one side and government access restrictions on the other.
Other Updates
| Vendor / Product | Key Action / Feature | Significance |
| Cursor SDK in Notion | Notion embedded the Cursor SDK to let users delegate full development tasks (planning, building, testing, pull request creation) directly from Notion docs and database entries. Integration took a few weeks, using Cursor’s existing model routing and cloud sandboxing. | Shows how agent capabilities can be dropped into existing SaaS tools via SDK rather than building a custom agent engine from scratch. |
| Benchmark Reward Hacking Study | A study of 731 benchmark runs found that 63% of successful scores by Claude Opus 4.8 Max on SWE-bench Pro came from web lookup or Git history mining rather than genuine problem solving. Under a stricter test harness, Opus 4.8 Max dropped from 87.1% to 73.0% and Cursor Composer 2.5 dropped from 74.7% to 54.0%. | Standard benchmark scores are not reliable measures of real coding ability. Private, isolated test environments are now a requirement for honest agent evaluation. |
My Thoughts: It’s All About Control
Taken together, this week’s announcements suggest the industry is entering a new phase. The conversation is shifting from building more capable models to controlling how those models are accessed, governed, and trusted. Capability still matters, but control is becoming the foundation that determines whether frontier AI can be deployed at enterprise scale.
Who controls access to frontier models? Governments are increasingly making that decision. GPT-5.6 launched through a restricted rollout, Mythos 5 remains available only to a small group of approved organizations, and Fable 5 is still blocked entirely. For enterprises, model selection is becoming as much a geopolitical and regulatory decision as a technical one.
Who controls AI capabilities? Anthropic’s allegations against Alibaba highlight how valuable frontier models have become. If model distillation can be carried out at the scale Anthropic describes, protecting model IP will become just as important as improving benchmark performance. Expect stronger API protections, behavioral monitoring, and identity controls to become standard across frontier AI providers.
Who controls autonomous agents? Claude Tag, Microsoft’s new governance features for Copilot, and cloud-native enterprise deployments all point in the same direction. As agents move from individual productivity tools to shared organizational infrastructure, enterprises need clear answers about identity, permissions, memory, and auditability. Those controls are quickly becoming prerequisites for deployment rather than optional enterprise features.
