AI News October 9, 2026: Google Cloud Ships Gemini Agent for Work, Claude Haiku 5.5 Slashes Prices 90%, Pwn2Own Hackers Pwn AI Infrastructure

The agent wars went fully mainstream this week. Google stopped selling copilots and started selling coworkers, Anthropic made frontier-adjacent intelligence nearly free, and hackers proved the whole stack is softer than anyone wants to admit. Here are the five stories that matter on October 9, 2026.

1. Google Cloud Launches the Gemini Agent for Work

Google Cloud introduced the Gemini agent — a single AI agent for work that answers questions, handles tasks, creates content, and writes code. Unlike chatbots that wait for prompts, the agent plans the work, uses tools, connects to business systems, and returns finished deliverables in documents, email, and developer environments.

It runs natively inside Google Workspace — Gmail, Docs, Sheets, Calendar — and, in a striking nod to reality, also works through Microsoft 365 and Slack. Most interesting for builders: the agent picks the best model for each task, currently routing across Google’s Gemini models and Anthropic’s Claude models, with more to come.

The launch answers OpenAI’s always-on “dots” agents from September and Meta’s Muse shopping-and-booking agent. Reuters frames it plainly: tech companies are now racing to sell AI that works autonomously. The differentiator is no longer model quality — it’s who owns the workflow.

2. Claude Haiku 5.5 Arrives With a 90% Price Cut

Anthropic released Claude Haiku 5.5, and the story is the price tag: token prices cut 90% for requests under 100,000 tokens — landing at $0.50 per million input / $2.50 per million output, which deliberately matches OpenAI’s GPT-6 Luna. The system card reports Haiku 5.5 significantly outperforming Haiku 4.5 and, in a few domains, rivaling Claude Sonnet 5.5.

The target is the repetitive enterprise work that makes AI expensive at scale: summarization, classification, database queries, and sub-agent tasks. Combined with Sonnet 5.5’s updated cache-read pricing, Anthropic is clearly signaling that per-token costs are collapsing industry-wide. If your API bill hasn’t dropped this quarter, you’re overpaying.

3. GPT-6 and Intelligent UI Complete Their Global Sweep

OpenAI’s biggest consumer rollout of the year is done: after Plus, Pro, Business, and Enterprise got GPT-6 Sol on October 7, Free and Go users started receiving GPT-6 Luna on October 8 — bringing Intelligent UI to all 1.2 billion weekly ChatGPT users.

Intelligent UI means answers can be interactive diagrams, charts, calculators, and tappable tools instead of walls of text — think a retirement calculator or retro game built live in the thread. OpenAI’s framing is unapologetically grand: “Instead of people adapting to software, software will adapt to people.” Note that this is a ChatGPT product feature; GPT-6 API availability and pricing haven’t been announced yet.

4. Washington’s “Super Intelligence Force” Gets to Work

The White House’s AI reorganization moved from announcement to execution. Jay Clayton, the director of national intelligence, now doubles as the administration’s AI czar, leading the interagency Super Intelligence Force created after executives from OpenAI, Anthropic, Google, and Meta signed a voluntary accord at the White House.

Per the Wall Street Journal, the task force has a 120-day deadline to report on AI’s risks and opportunities and define the federal government’s responsibility. Clayton’s mandate from the president: “we stay the leaders in superintelligence, and the interests of the American people are put first.” The group will consult consumers, public interest groups, religious organizations, and critical infrastructure providers.

5. Pwn2Own Ireland: AI Infrastructure Is the New Attack Surface

Security researchers exploited 32 zero-days in a single day at Pwn2Own Ireland — and the freshest category was AI infrastructure. Vietnam’s VinSOC won $40,000 for a five-zero-day chain against Oracle’s Autonomous AI Database; the OpenAI Codex coding agent fell to two separate teams; and Dynamo, LiteLLM, and NVIDIA Megatron Bridge targets all went down.

The message for anyone shipping agents: the tool-calling, code-executing layer you just deployed is now a bug-bounty target with enterprise blast radius. Agent frameworks need the same threat modeling as databases — before, not after, the invoice.

Worth tracking too: Broadcom’s banks are lining up $60 billion in chip financing for Anthropic; Google launched Project Suncatcher, a step toward AI data centers in space; and Microsoft says Copilot hit 30 million users, up from 20 million in March.

The through-line this Friday: agents are the product, cents are the price, and security is the unpaid bill. The companies that treat all three as one system — not three separate roadmaps — will own the next quarter.

Sources: Reuters · VentureBeat · OpenAI · Unite.AI · Federal News Network · BleepingComputer