Today's Headlines

  • Anthropic launches Claude Tag in beta, embedding Claude as a persistent team member inside Slack channels
  • OpenAI reports that an immunologist used GPT-5 to solve a research mystery that had stumped him for three years
  • A Stanford HAI study documents that a single vendor's hiring AI screened out 26% of Black and 15% of Asian applicants

AI worked its way into office conversations and delivered results in scientific research on the same day. Anthropic moved Claude to live inside a chat tool, while OpenAI showcased a case where it helped a researcher solve a long-standing puzzle.

On that same day, however, the shadow side of AI surfaced in the sensitive arena of hiring. A Stanford study uses large-scale data to show that AI-driven screening can systematically exclude certain racial groups. How far, and into which decisions, AI should reach grew heavier as a question today.

Today's Top Three

Anthropic Launches Claude Tag, Putting Claude Inside Slack

Anthropic launched Claude Tag in beta, a feature that lets teams invite Claude into Slack channels as a team member.

It is available to Claude Enterprise and Team plan customers. Unlike earlier @-mention integrations, Claude Tag stays resident in a channel and continuously accumulates the context of the ongoing conversation.

The heart of the feature is its ambient mode. Without being explicitly prompted, it can autonomously flag relevant information and follow up on tasks that have stalled. Administrators can scope each instance's channel access to prevent data from leaking across departments.

Anthropic says that 65% of its own product teams' code is already generated through an internal version of the feature. The product reframes Claude from a passive assistant into a resident participant in an organization's conversations.

GPT-5 Helps an Immunologist Solve a Three-Year Mystery

OpenAI published a case in which immunologist Derya Unutmaz used GPT-5 to resolve an immunology puzzle that had eluded him for three years.

GPT-5 acted not as a mere literature-search assistant but as a partner in shaping hypotheses and working through reasoning, according to the account. It is a case study of an LLM contributing to an actual discovery in specialist research.

The example carries weight because it is a practical report from science and medicine, fields with demanding standards of verification. It adds to the view that AI can move beyond summarizing papers to serve as a thinking partner for researchers.

This is a single case report, and reproducibility and scope warrant careful evaluation. Even so, as a concrete instance of AI use translating into results in a specialist domain, it moves the discussion a step forward.

Stanford HAI Documents Racial Bias in Hiring AI

Stanford's Institute for Human-Centered AI (HAI) published findings that a single vendor's AI hiring tool screened against 26% of Black applicants and 15% of Asian applicants.

The study tracked 3.4 million people across 4 million job applications. Had the AI recommended candidates fairly, an estimated 40,000 additional applications would have advanced to the next round.

More troubling is the problem of "algorithmic monoculture." The more employers use the same vendor's tool, the more a candidate rejected by one is rejected by the next. Ten percent of people who applied to four companies were turned down by all of them — a concentration of rejection that would not occur if each firm decided independently.

The reality that AI can create structural disadvantage at the gateway of hiring is strengthening calls for audits and regulation of recruiting AI. It is a case showing how the choice of where to place AI bears directly on individual opportunity.

Other Developments

Models & APIs

  • Mistral released OCR 4, a document extraction model supporting 170 languages that outputs block-type classification with bounding boxes and confidence scores. It claims top benchmark results, including 93.07 on OmniDocBench, and prices the API at $4 per 1,000 pages ($2 for batch). Mistral OCR 4 (Mistral)
  • All Claude services experienced elevated error rates for about 85 minutes on June 23, from 14:08 to 15:33 UTC. The incident affected claude.ai, the API, and Claude Code, and error rates returned to normal after a fix was applied. Elevated error rate across multiple models (Claude Status)
  • ITmedia examined whether OpenAI's on-site deployment unit and Anthropic's enterprise support could break Japanese companies' status-quo inertia. A survey found only about 11% of firms have the organizational readiness to leverage AI agents. Can hands-on support from OpenAI and Anthropic break enterprise AI inertia? (ITmedia AI+)

Research

Policy

Business

Other

Source: Selected by the editorial desk from the AI news inbox (collected June 24, 2026; 30 items, 5 primary and 25 secondary).