This Week in AI: GPT-6 Intelligent UI, Claude Haiku 5.5, and Google's Gemini Agent

This week pushed frontier models toward everyday users while vendors doubled down on agents, compliance, and defensive security. OpenAI widened GPT-6 access with a new interactive UI, Anthropic refreshed its small-model lineup and launched a cyber defense mission, and Google answered with a universal workplace agent—against a backdrop of EU watermarking rules and tighter Meta governance for capable models.

Models & Products

GPT-6 and Intelligent UI for everyone — OpenAI rolled GPT-6 out beyond paid tiers to the full ChatGPT audience, pairing the model with Intelligent UI so answers can mix text, charts, maps, and in-chat tools instead of plain paragraphs.

Introducing Claude Haiku 5.5 — Anthropic shipped its fastest Haiku yet for high-volume workloads, citing roughly 75% lower run cost than Haiku 4.5, stronger computer-use scores, and a new adjustable effort setting for trading speed against capability.

Google restricts Gemini Pro to paid AI plans — Starting October 9, free Gemini users move to Flash Lite only, while Gemini Pro stays on Google AI Pro and Ultra subscriptions—a sharper paywall as platforms hunt for paying power users.

Agents & Enterprise

Google Cloud launches Gemini agent — At Gemini at Work 2026, Google introduced a universal work agent that plans tasks, picks models, calls tools, and returns finished output inside Gmail, Docs, Drive, and other apps with enterprise governance built in.

Advancing computer use with Ironclad — OpenAI detailed reinforcement-learning work with contract-platform Ironclad, reporting higher scores and faster simulated runs on legal-ops workflows as it trains GPT-6 Astra for specialized business software.

Claude Frontier Academy and $100M talent push — Anthropic committed $100 million to train 10,000 Frontier Deployed Engineers by 2027, running residency-style programs with partners including Accenture, Deloitte, and Morgan Stanley to close the enterprise AI implementation skills gap.

Policy & Safety

OpenAI's approach to EU text provenance — To meet EU AI Act expectations, OpenAI outlined a phased rollout of its textGrain watermark for eligible ChatGPT and API output in Europe, with global API opt-in and researcher-only detector access while citing limits on short or heavily edited text.

Meta updates its Superintelligence Scaling Framework — Meta published Framework v2.1 with new loss-of-control safeguards for training and evaluation containment, open-weight release considerations, and plans for board-level oversight aligned with recent White House commitments.

Introducing the Anthropic Cyber Mission — Anthropic launched a long-term defender-focused program spanning critical infrastructure partnerships and a free OSS Scanner for open-source vulnerability finding, building on its expanded Cyber Verification Program.

Research

Sharing AI progress in mathematics — OpenAI released a batch of new results from an internal frontier model via GitHub, including Lean formalizations and transparency notes on compute use, following guidance from an IAS advisory group on how to publish AI-assisted math.

The through-line is maturation: models are cheaper, faster, and more visual for mass audiences, while the industry races to wire agents into real workflows and to prove it can meet regulation and harden infrastructure before misuse scales with capability.