AI News September 23, 2026: Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol/Luna Collide in a Same-Day Frontier Price War, Xiaomi Open-Sources MiMo-V2.6, OpenAI Opens Training-Phase Safety Evals, and Cisco Talos Flags the First Autonomous AI Malware C2

The frontier got dramatically cheaper in a single afternoon. On September 22, Anthropic and OpenAI shipped major releases just 90 minutes apart — and the combined effect is a price reset across the entire top tier of AI models, one week before OpenAI’s DevDay and two days before Sam Altman briefs the UN Security Council.

1. Claude Opus 5.5: Fable-Class Performance, 40% Cheaper

Anthropic released Claude Opus 5.5 on Tuesday, calling it its most powerful model to date — performing on par with its top-tier Claude Fable 5.1 on most tasks while costing 40% less to run on typical workloads than its predecessor, Claude Opus 5 (Reuters).

The numbers: $4 per million input tokens and $20 per million output — 20% below Opus 5 on sticker price — with cache reads down 60% to $0.20/MTok. Output generation is more than 30% faster, and a fast mode in Claude Code and the Claude Platform multiplies speed by 2.5× at a premium rate. On Terminal-Bench 4.0, Anthropic reports 66.4% versus Opus 5’s 50.3%.

Notably, Anthropic framed the release against its own “pacing the frontier” framework: Opus 5.5 is the first release since Dario Amodei’s September 12 call to slow AI development, was tested before release by external evaluators including Frontier Design and METR, and posted the strongest score to date on Anthropic’s automated behavioral alignment audit — 85% less likely than Opus 5 or Mythos 5.1 to attempt to bypass its prescribed boundaries. Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks.

2. GPT-6 Sol and Luna Land 90 Minutes Later at Half of GPT-5.6 Pricing

OpenAI didn’t wait for DevDay. Roughly 90 minutes after the Opus 5.5 drop, it launched GPT-6 Sol and Luna, positioned as “cut from the same cloth” as flagship GPT-6 Astra but priced at half the GPT-5.6 tier: $2/$10 per million tokens for Sol versus the previous $4/$20 (TechCrunch).

Sol targets complex coding and reaches Astra-level reliability on internal factuality evals, making about half as many mistakes as its GPT-5.6 predecessor; Luna is the lightweight tier for high-volume clerical work. OpenAI claims both beat Claude Fable and Opus on many tasks. Shipping workhorse models a week ahead of DevDay (September 29, San Francisco) reads as clearing the decks for something bigger.

Why the timing matters: with Grok 4.7 launching at $2/$6 just a day earlier, three labs have now compressed frontier-adjacent pricing within 48 hours. The cost of frontier-grade coding agents is falling faster than any margin model predicted — and consumers of tokens are the immediate winners.

3. Xiaomi Open-Sources MiMo-V2.6 Pro — Opus 5-Class, MIT License

Xiaomi quietly became a frontier-lab story this week. It released the MiMo-V2.6 series on Hugging Face: an omnimodal Pro model plus a 309B-parameter, 15B-active-parameter Flash MoE with 256K context — both under MIT license (Xiaomi).

Xiaomi claims Pro performs on par with Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks, scoring 46 on Artificial Analysis’s index. Coming the same month Z.ai’s GLM-5.3 shipped open weights and CISA alleged industrial-scale distillation by Chinese labs, MiMo-V2.6 deepens a clear pattern: the open-weight frontier is now arriving within weeks — not years — of closed labs, at a fraction of the price.

4. OpenAI Opens Training-Phase Safety Evals to Outside Groups

Ahead of Thursday’s UN Security Council briefing, OpenAI said it will let third-party groups conduct technical safety assessments during training, evaluation, and deployment of new models — earlier than the post-development testing that has been standard practice (Bloomberg).

The company published priorities for such reviews: strong independence mechanisms, scientific rigor, robust security practices, and clear responsibilities — and Sam Altman committed to embedding independent evaluators. The move follows reporting questioning whether embedded evaluators can truly stay independent, and lands alongside The Information’s revelation that an OpenAI–Anthropic cross-lab stress-testing pact nearly signed before the Hugging Face incident killed its momentum.

5. Cisco Talos Discloses CLOSEDQUORUM — First Fully Autonomous AI Malware C2

The security side of the AI story hardened this week. Cisco Talos released CAIRN, a metadata-first toolkit for hunting AI-integrated malware by extracting prompt templates, API keys, and jailbreak strings without executing binaries — and disclosed CLOSEDQUORUM, which it calls the first reported fully autonomous AI-driven malware command-and-control infrastructure (Talos).

Combined with Patrick Wardle’s disclosure of a Muse for Mac 0-day allowing any local app to steal Meta AI auth tokens, and Treasury Secretary Bessent blaming “OpenAI management, not a bunch of agents” for the July Hugging Face incident, the offensive-AI threat model moved from theory to tracked infrastructure in a single week.

Quick hits: Ex-Anthropic startup Mirendil is in talks to raise up to $1B at a $5B valuation (Kleiner leading, a16z in discussions) to build AI that does the work of an AI researcher; AstroForge is putting its transformer-based “Solo” AI in autonomous command of the Autonomy-1 asteroid-mining probe, flying shadow-mode from late 2026; OpenAI fired contractors for using AI to train its AI, banning GPTZero and Grammarly in reviewer workflows; Grok Bot hit 418K weekly users, up 24% week-over-week; and Google’s Gemini-powered “Googlebook” laptop is available for preorder.

Two frontier models, a same-day price war, an open-weight shock from Xiaomi, and the first autonomous AI malware — September 2026 is compressing a year of AI history into a week. Follow our daily AI news coverage for tomorrow’s update.