Here's a roundup of AI news from September 23 (Japan time, morning). The biggest story: OpenAI and Anthropic released new models on the same day. OpenAI launched the GPT-6 family's "GPT-6 Sol" and "GPT-6 Luna" at nearly half the token price, while Anthropic countered by releasing "Claude Opus 5.5," which delivers performance on par with Fable 5.1. It turned into a collision release with pricing as the main battleground. Other stories: a follow-up on Meta officially acknowledging the similarity between its AI assistant "Muse" and OpenClaw, Microsoft's takedown of an AI-powered fraud platform, Qualcomm's chips that can run 30B MoE models locally, and from Japan, Lawson's launch of products conceived by generative AI.
GPT-6 Sol and Luna ── Astra-Class Intelligence at Nearly Half the Price, with Prompt Caching Improvements Announced the Same Day
OpenAI has announced two new models: "GPT-6 Sol" and "GPT-6 Luna." According to TechCrunch, both models belong to the same lineage as GPT-6 Astra and promise lower costs and fewer mistakes. AWS's official blog lays out their positioning clearly: Sol handles "complex tasks that recur week over week" with strong reasoning and coding capabilities, while Luna's role is to process "high volumes of repetitive tasks" at practical latency and cost. Both are offered at API prices far below the previous-generation GPT-5.6 series, and are now generally available on Amazon Bedrock.
The Decoder reads this launch as a "price-halving offensive." The new models deliver performance on par with the previous generation at roughly half the token price, while independent analyses suggest the actual intelligence gains are limited — pointing to the possibility that the pricing targets Anthropic's more expensive offerings. The report also speculates that OpenAI probably didn't expect Anthropic to drop Opus 5.5 on the very same day.
Also on the same day, OpenAI announced improvements to GPT-6 prompt caching. Higher cache hit rates, new diagnostics, explicit breakpoints, and fine-grained controls are said to cut latency and costs — a practical improvement that pays off more the heavier your usage.
Anthropic's Opus 5.5 Lands the Same Day ── Fable 5.1-Equivalent Performance at Roughly 40% Below Opus 5
On the same day as OpenAI, Anthropic released "Claude Opus 5.5," the first model in the Claude 5.5 family. According to The Decoder, Opus 5.5 delivers performance on par with Claude Fable 5.1 on most tasks, at roughly 40% lower running cost than Opus 5. Anthropic's benchmarks put it ahead of GPT-6 Astra on most tasks, and TechCrunch reports the company called it "the most capable model we have tested to date." Sonnet 5.5 and Haiku 5.5 are expected within weeks.
AWS's blog positions it as the Opus best suited for agentic coding, knowledge work, and long-running tasks. Beyond efficiency gains — getting the same work done with fewer tokens — adaptive thinking is now always on, controlled via effort rather than a manual thinking budget. At the same time, it is the first Opus to ship with the same safety classifiers as Fable 5.1 in the biology, cybersecurity, and AI development domains, and the post explicitly notes that it will refuse more requests than previous Opus models.
Community reactions came quickly: on LocalLLaMA, Opus 5.5's Artificial Analysis score of 58 has kicked off a discussion of "how much longer until open models catch up."
Meta Admits the Muse/OpenClaw Similarity Was "Not a Coincidence"
Further developments following the recent turmoil we covered, including Amazon blocking proxy purchases of Muse. According to TechCrunch, Meta — while standing by its claim that the AI assistant "Muse" was built from scratch — has acknowledged it was "heavily inspired" by OpenClaw. The influence reportedly extends to file names and content in some workspaces, meaning the similarities that until now had only been pointed out by outside observers have now been officially confirmed.
Meanwhile on Hacker News, a report that Meta is testing a "human concierge" for Muse is also drawing attention. Exactly what counts as "human completion" versus the model's job remains unclear at this point.
Microsoft Leads Industry-Wide Takedown of "EvilTokens," an AI-Chatbot-Equipped Fraud Platform
According to Ars Technica, Microsoft announced on Tuesday, September 22 that it had led a coordinated, industry-wide takedown operation. The target: "EvilTokens," a subscription-based fraud platform. The platform surfaced via a Telegram channel in February, charging $1,500 up front plus $500 per month. It reportedly compromised 12,000 Microsoft accounts within a few months.
At the center of it all was an AI-style chatbot. After centralizing the process for compromising email accounts at scale, it analyzed victims' inboxes to identify trust relationships, payment approvals, and sensitive permissions — then handled everything from selecting the targets most likely to pay to drafting fraud emails impersonating trusted contacts. The defining feature of this case is that AI handled target selection, fraud strategy recommendations, and message drafting, going well beyond the compromise itself.
Qualcomm Announces Two Smartphone Chips Capable of Running 30B MoE Models Locally
According to TechCrunch, Qualcomm has announced two smartphone chips with a strong emphasis on AI processing, and the new flagship chip among them is claimed to run 30B mixture-of-experts (MoE) models on-device. This signals a major jump in the size of models that can run on edge devices, pointing in the same direction as the local LLM practicality trend we've covered recently (such as GGUF builds of Qwen-Image-2.1 and progress on quantization). Given how MoE works, this presumably relies on a configuration with a small active-parameter footprint, though the detailed operating requirements are unclear at this point.
Lawson to Sell Three Generative-AI-Conceived Products at Around 4,700 Stores Starting September 29
Now a story from Japan. According to an explanatory article on Qiita, Lawson will sell three products whose ideas came from generative AI at around 4,700 stores in the Kanto and Koshin'etsu regions starting Tuesday, September 29, 2026 (based on a news release the company posted on its official website dated September 16). The three products are priced at 270, 181, and 475 yen. What stands out is that the ideation process reportedly did not use past sales data — which reads as an exercise in asking generative AI for "diversified ideas" without relying on data backing. As a case of practical generative AI use reaching product development in Japanese retail, it's worth following up on.
---
With OpenAI and Anthropic clashing head-on over price and performance in the same week, model selection appears to have shifted a step from "the single strongest one" toward "cost optimization per use case." Until independent evaluations fill in real-world numbers for Sol, Luna, and Opus 5.5, it's worth taking each company's claims with a grain of salt.