The AI Brief

France, holding the rotating Security Council presidency in September, convened the session on September 23 during the high-level week of the UN General Assembly. French Minister for Europe and Foreign Affairs Jean-Noël Barrot chaired. The confirmed briefers were OpenAI CEO Sam Altman, who attended in person, and Anthropic CEO Dario Amodei, who joined virtually. Also presenting: Yoshua Bengio, co-chair of the UN's Independent International Scientific Panel on AI (IISP-AI), and Hugging Face CEO Clément Delangue. Chinese lab DeepSeek sent company representatives; DeepSeek founder Liang Wenfeng did not attend personally. Chinese coding-model startup Moonshot was also invited to present.

The Security Council Report, which tracks Council activity as a primary monitor, described the session as "the Council's first meeting focused specifically on the safety risks posed by increasingly capable AI systems, including the potential loss of human control." The Council had previously held six formal and informal meetings on AI, but those examined broader peace-and-security dimensions rather than capability risk specifically. No binding resolution was expected from today's session; France's concept note positioned the Council as having a key role in addressing AI risks, a framing that Russia and China are expected to contest on sovereignty grounds.

The session lands as Amodei's pacing essay, published September 12, continues to drive policy debate, and as OpenAI CEO Altman is separately scheduled to brief the UN on the framework OpenAI published September 22 calling for US-led global technical standards on frontier AI and recursive self-improvement (RSI) oversight. The timing places both lab chiefs before the full permanent membership within 24 hours of OpenAI's framework publication.

OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, 2026, nineteen days after GPT-6 Astra opened the family. Sol is priced at $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4/$20 promotional rate. Luna lands at $0.10/$0.50, down from $0.20/$1.20. An OpenAI spokesperson confirmed to VentureBeat that both rates are permanent, not promotional. Cached-input pricing drops 90% on both models, with Sol cached input at $0.20 per million and Luna at $0.01. The cuts were attributed to improved caching and inference efficiency built into the GPT-6 architecture.

Independent evaluator Artificial Analysis found GPT-6 Sol improved roughly 2 points over GPT-5.6 Sol on its Coding Agent Index, with measurable gains on Terminal-Bench 4.0 (43% versus 37%) and SWE-Atlas-QnA (58% versus 54%) at roughly half the cost per task. GPT-6 Luna, however, scored 2 points below GPT-5.6 Luna on the same index, with declines on SWE-Atlas-QnA and DeepSWE v1.1, making Luna primarily a price cut rather than a capability upgrade. OpenAI's internal factuality test found Sol makes roughly half as many mistakes as its predecessor, though the test used de-identified real-world conversations flagged for errors and was not controlled for answer length or task distribution. Sol scored 68.8% on DeepSWE v1.1, within about 1 point of Claude Fable 5.1, at roughly one-fifth the price.

Both models are currently available in ChatGPT Work and Codex for paid subscribers; they are not yet in the standard ChatGPT chat interface. Free and Go users can access Luna on the desktop app. The launch arrived 89 minutes after Anthropic released Claude Opus 5.5 at $4/$20, meaning both price-cutting moves landed on the same trading day.

Anthropic released Claude Opus 5.5 on September 22, 2026 at 16:31 UTC. The model is priced at $4 per million input tokens and $20 per million output tokens, down 20% from Opus 5 at $5/$25, with cache reads cut 60% from $0.50 to $0.20. On typical workloads it costs roughly 40% less to run than Opus 5 because it also generates fewer tokens per task. It carries a 1 million-token context window, 128,000-token maximum output, and a June 2026 knowledge cutoff. On Anthropic's internal benchmark table, Opus 5.5 leads on SWE-bench Pro (89.9%) and Terminal-Bench 4.0 (66.4%), and leads 7 of 9 benchmark rows overall, with GPT-6 Astra ahead on AutomationBench and Terminal-Bench-Science.

Preserved thinking ships on both Claude Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026. Older accounts are unaffected. The control stops API users from injecting or editing content into Claude's prior-context thinking blocks in an attempt to extract reasoning. Anthropic's announcement links this directly to the distillation threat detailed in its September 2026 threat intelligence report, which documented Alibaba's 151-million-exchange extraction operation and named seven other suspected distillation actors. Anthropic also raised Claude Code five-hour session limits 20% on September 22, and said Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks.

The sequence on September 22: Anthropic posted Claude Opus 5.5 at 16:31 UTC at $4/$20 per million input/output tokens, down 20% from Opus 5 on list price and roughly 40% cheaper on typical workloads due to lower token generation per task. OpenAI followed at 18:12 UTC with GPT-6 Sol ($2/$10) and Luna ($0.10/$0.50), both 50% below GPT-5.6 promotional rates and confirmed as permanent. The preceding day, SpaceXAI's Grok 4.7 had arrived at $2/$6, itself 80% below the prior frontier, starting the cascade. The practical result is that Sol undercuts Anthropic's new flagship on both input and output on list price, while Opus 5.5 leads Sol on seven of nine benchmarks including SWE-bench Pro.

The framing that one company "won" the day misses the structure: Opus 5.5 leads on pure benchmark performance across most coding and knowledge-work tasks; Sol leads on unit cost for equivalent output on some agentic benchmarks; Luna establishes a new cost floor for high-volume, lower-complexity tasks. OpenAI CFO Sarah Friar noted at the Goldman Sachs Communacopia conference that an earlier 80% Luna price cut drove roughly a tenfold usage increase, providing the economic rationale for today's further reduction. For builders evaluating model routing, the three-tier GPT-6 family now covers extraction (Luna), production coding (Sol), and peak reasoning (Astra) within a single API integration.

Mirendil was founded by Behnam Neyshabur, who previously co-led Anthropic's scientific AI reasoning team, and Harsh Mehta, a former Anthropic senior research scientist. Both left Anthropic in December 2025. The company raised $200 million in a seed round in June 2026 at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia participating. Bloomberg reported on September 22 that Mirendil is now in talks to raise up to $1 billion in a new round at a $5 billion post-money valuation, with Kleiner Perkins in discussions to lead and Andreessen Horowitz again considering participation. The terms are not yet final.

The company now has roughly 20 researchers drawn from Anthropic, xAI, Google DeepMind, and OpenAI. It secured a $100 million Google Cloud deal in August 2026 for TPU and GPU access. Its stated target is a frontier engineering and research model by early 2027. No public model has shipped. If the round closes at $5 billion, Mirendil will have raised approximately $1.2 billion in total within nine months of founding, before generating any disclosed external revenue. The company's valuation trajectory sits above Thinking Machines Lab's $40 billion raise at over $100 million in annualized revenue on a price-to-revenue basis, underscoring the premium the market is placing on RSI research specifically.