Latest AI news — September 22, 2026

Anthropic: Claude Opus 5.5 is here. The first model in the Claude 5.5 family matches Claude Fable 5.1 on most work while running ~40% cheaper than Opus 5 on typical loads (input/output $4 / $20 per million tokens; cache reads $0.20). External evaluators (including Frontier Design and METR) reviewed it before launch; Anthropic says it is its strongest automated behavioral audit so far. Biology and cybersecurity capability is Mythos 5.1–like, so it ships with Fable-style safeguards; Sonnet 5.5 and Haiku 5.5 follow in the coming weeks. Source: Anthropic, Reuters.

OpenAI: GPT-6 Sol and Luna — cheaper GPT-6 for everyday work. After Astra, two more cost-efficient members of the family use the same methods; API prices are 50% lower than the GPT-5.6 promo tiers (Sol $2 / $10, Luna $0.10 / $0.50 per million tokens). Available in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu; Free/Go users get Luna on the desktop app; API ids gpt-6-sol and gpt-6-luna. Rollout in ChatGPT is gradual. Source: OpenAI, TechCrunch.

OpenAI: principles for third-party external safety evaluation. The company published four priority areas and principles for working with private and nonprofit independent assessors: deep access across training, evaluation, and deploy so safety cases and safeguards can be stress-tested from outside. The frame complements government testing, it does not replace it. Bloomberg reports external groups may also get earlier-phase model access. Source: OpenAI.

SpaceXAI: scaling customer support with Grok Bot. After the Cursor acquisition, the combined SpaceXAI+Cursor support stack was rebuilt around Grok Bot: 175% more tickets with no new hires (otherwise ~200 people would have been needed). Typical resolution cost is $0.20–$0.30 versus $1–$4 per ticket for conventional AI support; ~99% of refunds clear without a human. (More: AI support agent.) Bloomberg says the Grok Bot agent already tops 400,000 users after its first month. Source: SpaceXAI / x.ai, Bloomberg.

Meta: human “concierge” for the Muse AI agent. Citing internal posts, Reuters reports Meta is testing human contractors who quietly take over some Muse phone calls — not long after the company publicly launched Muse calling. Some staff raised privacy concerns if sensitive information can reach call-center contractors. Source: Reuters.