This week the leading labs moved simultaneously at the infrastructure and workflow layers, while Google absorbed the visible cost of being the source from which talent is currently migrating. The result is a landscape where two labs grow harder to displace while one faces a compounding set of pressures.
Anthropic gives every Slack channel a shared AI teammate
On 23 June, Anthropic launched Claude Tag as a research preview for Enterprise and Team customers. Unlike the earlier Claude bot, which responded one-to-one in direct messages, Claude Tag gives each Slack channel a single shared @Claude identity that every member of the channel can address, observe, and hand off. Context threads across the full conversation rather than resetting with each new participant.
The agent is genuinely asynchronous: a teammate tags @Claude with a request, it breaks the work into milestones, executes them using connected tools and codebases, and posts results back in-thread. No one needs to remain in the chat while it works. Administrators can scope precisely which tools and data sources Claude may access on a per-channel basis, and an ambient mode lets it proactively surface stalled threads or material changes across its connected surfaces without being asked.
Two data points frame the ambition. Anthropic reports that 65% of its own product team’s code is now generated using an internal version of the tool. The existing Claude in Slack app is being retired on 3 August, giving Enterprise admins a 30-day migration window. For operators evaluating where AI fits in daily team workflows, this is the clearest signal yet that the IDE is not the only surface worth instrumenting. The practical question for any CxO is not whether to adopt Slack agents, but who in the organisation owns the governance model for what they can access.
OpenAI restricts its sharpest security model to vetted defenders
On 22 June, OpenAI shipped the full version of GPT-5.5-Cyber, its most capable defensive security model to date. The benchmark numbers are notable: 85.6% on CyberGym (versus 81.8% for standard GPT-5.5) and 39.5% on ExploitGym (versus 25.95%), the highest single-model CyberGym score recorded. The model can scan large codebases end to end, identify vulnerable components, generate patches, and validate fixes without human intervention in the loop.
Access is deliberately narrow. GPT-5.5-Cyber is distributed through OpenAI’s Trusted Access for Cyber programme to a vetted set of critical-infrastructure operators: Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler. OpenAI is simultaneously launching a Daybreak Cyber Partner Program that allows approved security vendors to embed the model in their commercial offerings.
The design reflects a considered asymmetry. The same profile that makes this model effective for defenders also makes it dangerous in other hands. The 13.5-point jump on ExploitGym from the base GPT-5.5 is a reminder that offensive capability rises in lockstep with defensive capability. If your organisation protects critical infrastructure and is not yet in the partner programme, that is the question to answer first. For everyone else, the relevant takeaway is that purpose-built, access-gated AI for regulated sectors is now a shipping product, not a research direction.
Micron joins the Anthropic supply ring, completing the HBM trifecta
On 22 June, Micron and Anthropic announced a four-pillar strategic agreement: co-design of memory and storage architecture, a multi-year supply deal spanning high-bandwidth memory, DRAM, and SSDs, an internal deployment of Claude across Micron’s workforce, and a strategic equity investment by Micron in Anthropic’s Series H funding round. Financial terms were not disclosed.
The structural implication matters more than the headline. Samsung and SK Hynix were named as Series H participants when the round closed on 28 May. Micron’s addition means all three global HBM manufacturers, who together supply 100% of the world’s high-bandwidth memory production, now hold equity stakes in Anthropic and have committed multi-year capacity to it at a time when supply is fully booked through 2026.
No other frontier AI lab has this configuration. For operators choosing between providers over a three-to-five year horizon, supply security is now a differentiating variable. A lab that cannot guarantee memory access cannot guarantee training runs or inference at the cost and latency you are building into your roadmap. The consolidation happening at the chip layer is, in practical terms, a moat that does not show up on a model benchmark card.
Google misses its Gemini 3.5 Pro deadline as departures accelerate
At Google I/O on 19 May, Sundar Pichai committed explicitly to a June general availability launch for Gemini 3.5 Pro, citing its 2-million-token context window, a Deep Think reasoning mode, and frontier multimodal capability. As of 25 June, the model remains in limited Vertex AI enterprise preview. Google has moved the target to July, citing quality refinements after early enterprise testing and feedback from its Antigravity evaluation platform.
The delay is happening against a difficult backdrop. In the two weeks prior to this writing, four senior AI researchers left Google for rivals. Noam Shazeer, co-lead of Gemini and co-author of the foundational 2017 Transformer paper, announced on 18 June he is joining OpenAI. John Jumper, 2024 Nobel laureate and AlphaFold lead, announced on 20 June he is joining Anthropic. Two further departures to OpenAI were also confirmed in the same period.
Google’s underlying research base remains formidable, and Gemini 3.5 Flash is already the default in Google Search’s AI Mode and the Gemini app. But a missed public commitment from the CEO, compounded by the most visible run of senior departures the company has seen in this cycle, is the kind of signal that starts to move enterprise procurement conversations. Prediction markets currently place the probability of a 30 June release at roughly 50 to 55 percent. A July launch may be fine technically; the harder question is what the market reads into the overall pattern.
Taken together, today’s four stories describe a consistent dynamic: the leading labs are hardening their positions at the infrastructure and distribution layer, through supply deals, platform integrations, and channel partnerships, while Google faces a compounding set of pressures on product delivery and talent retention. The structural advantage being built is not primarily about having the biggest model at any given moment. It is about having the most reliable stack beneath it, and the teams in place to keep improving it.