This Week in AI: Claude Opus 5.5, GPT-6 Sol and Luna, and Meta's Muse Expansion

This week turned into a pricing chess match at the frontier: Anthropic and OpenAI shipped new flagship-tier models within about ninety minutes of each other, both leaning on cheaper inference and stronger agent benchmarks. Behind the launches, labs pushed governance and evaluation into the open, while Meta tried to make its Muse agent family feel everyday—and researchers tested what happens when models and agents step into biology, mental health, and peer-to-peer markets.

Models

Introducing Claude Opus 5.5 — Anthropic opened its Claude 5.5 line with a model that it says matches Fable 5.1 on most work while cutting typical run costs about forty percent versus Opus 5, with faster output and stricter safeguards for biology- and cyber-sensitive use cases.

Introducing GPT-6 Sol and Luna — OpenAI refreshed its mid-tier GPT-6 pair shortly afterward, halving API prices versus promotional GPT-5.6 rates and pitching Sol for heavy coding and Luna for high-volume clerical workflows while keeping Astra as the top tier.

Better prompt caching for GPT-6 — Alongside the model drop, OpenAI rolled out GPT-6 prompt-caching upgrades, a usage dashboard, and diagnostics so long-running agents can reuse shared context for up to thirty minutes and cut cached-input spend sharply.

Agents & Tools

Meta Connect 2026 recap for developers — At Connect, Meta expanded Muse with Spark 1.3 for agentic coding, Muse Code for terminal and CI workflows, broader Model API availability, and new AI-glasses and Horizon building blocks aimed at shipping consumer agents faster.

Bringing your Muse to life — Meta also unveiled Muse Realtime Avatar, streaming low-latency voice-and-video personas built on a redesigned real-time inference stack co-optimized with NVIDIA for concurrent sessions on GB200 hardware.

Project Swap: What happens when agents trade for us? — Anthropic ran an internal book-swap marketplace where Claude agents negotiated on employees' behalf, surfacing frictions around trust, refunds, and the need for agent identity registries before autonomous commerce scales.

Research

Claude discovers a novel enzyme system with CRISPR-like repeats — Anthropic's new life-sciences lab reported that Claude flagged an uncharacterized bacterial repeat array reminiscent of early CRISPR discoveries, a early win for its agent-assisted biology workflow separate from last week's large-scale virtual biotech demo.

Introducing MentalHealthBench — OpenAI released an open benchmark co-built with more than eighty licensed clinicians across twenty-two countries to score how models handle realistic mental-health conversations beyond emergency refusals, spanning teens, caregivers, and varying acuity levels.

Policy & Governance

Sam Altman's remarks at the United Nations Security Council — OpenAI CEO Sam Altman told the Security Council that rapid model progress compresses timelines for both opportunity and risk, urging international guardrails against loss of human control and concentration of AI power in too few hands.

Priorities and principles for effective third party assessments — OpenAI published a framework for deeper independent scrutiny of safety cases, safeguards, and deployment claims, part of its broader push to pair faster releases with embedded external evaluation.

The through-line is familiar but sharper: frontier labs are racing to make capable models affordable enough for always-on agents, while simultaneously inviting outsiders—and diplomats—to inspect whether speed is still paired with credible control. Consumer platforms are betting that agents will live in glasses and terminals, not just chat boxes, even as researchers document new scientific and social domains where those agents already leave fingerprints.