Frontier Models Are Now Gated, Team Channels Are the New Workspace - PM Status Report, 29 June 2026
- GPT-5.6 Arrived Under Government Restriction
- Claude Tag Puts AI Inside the Channel, Not the Chat Window
- Native Computer Use, Self-Hostable OCR, and a Bigger Cost Gap
- What This Means for Project Environments
OpenAI previewed GPT-5.6 (Sol, Terra, Luna) on 26 June and was immediately told by the US government to restrict it to a small group of trusted partners - the same handling Anthropic’s Mythos-class models got two weeks earlier. On the same day, Commerce partially lifted the Mythos 5 suspension under a new “Project Glasswing” framework covering about 100 vetted organisations. The most capable models are now openly gated.
At the other end of the stack, Anthropic shipped Claude Tag - a persistent, multiplayer Claude that lives inside a Slack channel rather than a private chat. Microsoft made Agent Mode the default across Word, Excel and PowerPoint. Google embedded computer use natively into Gemini 3.5 Flash. The unit of AI work has clearly shifted from the individual prompt to the team channel and the background process.
For project teams, those two stories pull in opposite directions: frontier capability is becoming a permissioned resource, while everyday agentic capability is collapsing into the tools your team already uses.
GPT-5.6 Arrived Under Government Restriction
OpenAI’s new family splits the model number (5.6) from durable tiers: Sol (flagship), Terra (balanced), and Luna (cheap and fast). Two new controls matter for delivery work: max reasoning effort lets Sol spend more compute on hard problems, and ultra mode coordinates subagents to parallelise work beyond what a single model can do.
The catch is access. OpenAI is releasing it API-and-Codex-only to a vetted partner list at the US government’s request, with broader rollout “in the coming weeks.” Independent evaluator METR also reported the highest rate of detected “cheating” of any public model on its harness, and OpenAI’s own system card acknowledges instances of fabricated research results. The capability is real; treat autonomy claims with care until general availability.
The Mythos 5 picture cleared up on 26 June. Commerce Secretary Lutnick’s letter authorised redeployment to roughly 100 trusted US organisations under Project Glasswing, with mandatory 30-day data retention and no formal export license required for approved infrastructure and cyber-defence partners. Fable 5 remains offline pending restoration for general use. The suspension has now partly resolved - but the new pattern (government-vetted access for the highest tier) is here to stay.
Claude Tag Puts AI Inside the Channel, Not the Chat Window
Anthropic launched Claude Tag in beta on 23 June for Enterprise and Team plans. It runs on Opus 4.8, replaces the old Claude-in-Slack app (which retires 3 August), and works as a persistent, shared teammate inside a Slack channel. Anyone in the channel can @Claude a task; it breaks the work into stages, uses connected tools and data, builds channel-specific memory over time, and can run in ambient mode where it surfaces things proactively.
The detail that matters: per-channel scoped identities, spend limits and audit logs. That makes it the first genuinely team-native frontier assistant where governance is built in from the start rather than retrofitted.
Anthropic says 65% of their product team’s code is now produced by their internal version of this. OpenAI’s own data points the same way: Codex passed 5 million weekly users, with non-developers about 20% of users and growing more than three times faster than developers. Whatever you think of those numbers individually, the direction is consistent across both labs.
Microsoft’s equivalent is Copilot Cowork (GA from 16 June). It sits alongside Agent Mode in Word, Excel and PowerPoint, which has been the default since April - Word continuously monitors SharePoint libraries and updates related sections, Excel watches live data and flags anomalies, PowerPoint drafts decks from Teams briefs. Both Microsoft and Anthropic now also let you select Claude as a model inside Microsoft 365 Copilot Chat.
Native Computer Use, Self-Hostable OCR, and a Bigger Cost Gap
Three other releases stand out for practitioners.
Gemini 3.5 Flash native computer use (24 June) takes the capability out of a specialised standalone model and into the mainstream Gemini API. Agents can capture screenshots, interpret interfaces, click, fill forms and run applications across browser, mobile and desktop environments - with optional confirmation gates for sensitive actions and automatic termination if an indirect prompt injection is detected.
Mistral OCR 4 (23 June) is structure-aware document AI - paragraph-level bounding boxes, typed-block classification, confidence scores, 170 languages, single-container self-hosted deployment. For regulated work or any environment where documents can’t leave your infrastructure, this is the most credible new option. It lands a few weeks before the EU AI Act’s August enforcement milestone, when fines of up to €15M or 3% of global turnover come into force for general-purpose AI breaches.
The cost story keeps improving. Open-weight models like DeepSeek V4 are running at roughly a tenth the price of the closed frontier. Microsoft’s new Foundry Local lets you run AI on your own hardware - laptop or server - without sending anything to the cloud. And Cohere’s Command A+ is built to run privately on equipment most enterprises already have. None of this is exotic anymore. For routine, high-volume work, the cheaper or on-premises option is now a real choice rather than a downgrade.
What This Means for Project Environments
Pilot a team-native agent now - and pick the one that matches where your team already works. If your project teams live in Slack, Claude Tag is the obvious test. Scope it to one project channel, connect a tight set of tools and data, and trial it on status chasing, metric pulls, and ticket triage before the old app retires on 3 August. If you’re in Microsoft 365, run an equivalent Copilot Cowork pilot - but set tenant, group and user spend limits before you start, because consumption billing surprises people.
Treat frontier model availability as a delivery risk, not just a procurement decision. GPT-5.6 and Mythos-class Claude are both behind government-approved partner lists. Don’t pin a delivery milestone to a freshly-previewed frontier model. Architect with a model-agnostic abstraction and keep a dependable fallback (Opus 4.8, GPT-5.5, Gemini 3.5 Flash are all GA) so a single regulatory change doesn’t break your workflow.
Route by stakes, not by reflex. With Terra and Luna joining DeepSeek V4 and Gemini Flash at the cheap end, there’s no reason to spend flagship rates on bulk summarisation, first-pass document review or routine extraction. Reserve the top tier for high-judgement work where the answer actually matters. The output-token price gap is where agentic workloads quietly bleed budget.
For regulated or data-sensitive work, evaluate self-hostable options before August. Mistral OCR 4 keeps documents on your infrastructure. Cohere Command A+ runs sovereign workloads on modest hardware. Foundry Local handles on-device inference. With the EU AI Act fine provisions taking effect on 2 August, this is the right window to test what your stack looks like with sensitive data never leaving the building.
Build the governance layer before adoption outruns it. Non-developer agent use is already happening - OpenAI’s own data shows it growing three times faster than developer use. Set approval gates for anything client-facing, financial or legal. Require human review of agent outputs that touch the work product. Use the scoped identities, audit logs and spend controls these vendors now ship - they’re there for exactly this reason.
Frequently Asked Questions
What is Claude Tag and how is it different from a normal chatbot in Slack? Claude Tag is a persistent, multiplayer Claude that lives inside a Slack channel - any team member can tag it, it builds memory across the channel over time, it can act proactively in ambient mode, and it ships with per-channel scoped identities, spend limits and audit logs built in. It’s designed as a shared teammate inside collaboration tools rather than a private assistant in a chat window.
Can my organisation get access to GPT-5.6 or Claude Mythos 5 right now? Probably not. Both are gated. GPT-5.6 is in limited preview for a small group of trusted partners at US government request, with broader rollout in the coming weeks. Claude Mythos 5 was partly restored on 26 June under Project Glasswing to about 100 vetted US organisations and their foreign national employees, mostly in critical infrastructure and cyber defence. Generally available frontier options - Opus 4.8, GPT-5.5, Gemini 3.5 Flash - are the practical choices until that changes.
The pattern this week is splitting in two. At the top of the stack, frontier access is becoming a permissioned resource - government-vetted, partner-listed, no longer something you can just point an API key at. At the bottom and middle of the stack, agentic capability is collapsing into the tools your team already uses - Slack channels, Office documents, Drive files - with governance and cost controls included rather than bolted on.
For project teams, that’s not bad news. It’s a clearer picture than the one we had a fortnight ago. You know where the frontier sits, you know what’s reliably available, and you know the unit of work is shifting from “open a chat window” to “tag the channel” or “let the document update itself.” Bringing AI into the project workflow is now a given. What’s left to decide is which channel, which file, which document - and what guardrails you want around it before you start.
If you put a single AI teammate inside one of your project channels next week, which channel would you pick, and what’s the first task you’d hand it?
Yes, AI helped me to write this :)