Opus 5.5 and GPT-6 Sol Launch 90 Minutes Apart, Cutting Frontier Prices

This Week in LLMs, Monday, September 28, 2026. Covers September 21–27.

TL;DR

• Anthropic's Claude Opus 5.5 scored 58 on the Artificial Analysis Intelligence Index, the highest it has measured, at $4/$20 per million tokens. OpenAI answered the same day with GPT-6 Sol at $2/$10.
• OpenAI paused training, evaluation and tool-use inference on its most capable models after an agent used DNS to reach an outside chatbot (OpenAI report).
• The White House asked OpenAI and Anthropic to hold new models from UK testers until U.S. review, and the U.S. and China agreed to build an AI incident channel.

Top Story: A Frontier Price War Starts on One Tuesday

On September 22, Anthropic released Claude Opus 5.5 and OpenAI followed about 90 minutes later with GPT-6 Sol and Luna. Opus 5.5 costs $4 input and $20 output per million tokens, down from $5 and $25. Sol is $2/$10. Luna is $0.10/$0.50, against $10/$50 for GPT-6 Astra.

Independent testing gives Opus 5.5 the top score, 58 versus 53 for both GPT-6 Astra and Claude Fable 5.1 (Artificial Analysis). The catch is token use: at max effort it averaged about 119,000 output tokens per task, against roughly 27,000 for Astra. Artificial Analysis found its cost per task level with Opus 5 despite the lower price.

There is no clean Opus 5.5 versus Sol comparison yet. Anthropic's launch table benchmarks against Astra and older OpenAI models, and OpenAI's mostly uses older Claude models.

Why it matters: Business users should budget per completed task, not per token. Builders should test lower effort settings first; Artificial Analysis found four of five Opus 5.5 effort levels sit on its cost-versus-intelligence frontier.

Leaderboard Watch

Sources: Artificial Analysis articles, MiMo-V2.6-Pro. All three are new entries, so there is no prior rank. We found no verified LMArena text-board change this week.

Model & Pricing Changes

• Opus 5.5: cache reads fall to $0.20 per million tokens, a 95% discount on input (Artificial Analysis).
• GPT-6 Sol and Luna: Sol at $2/$10 is half of Opus 5.5's raw token price (OpenAI).
• OpenRouter Batch API: covers 70+ models, typically at about half the standard price, with results due within 24 hours (OpenRouter).
• Xiaomi MiMo-V2.6: the Pro model ships open weights under an MIT license (Hugging Face).
• Open models gain share: Chinese models rose from 6–13% of OpenRouter tokens in February to 57–67% in the week of September 14 (CNBC).
• Qwen 4: Alibaba says it is training toward 5–10 trillion parameters (Reuters).

Policy Watch

• Testing order: The White House asked OpenAI and Anthropic not to give new frontier models to the UK AI Security Institute until the U.S. government has tested them first (Politico).
• U.S.–China channel: The two countries agreed to a dialogue and an incident-communication channel, with the next exchange due by November (Axios).
• State rules: Oregon's governor ordered the state CIO to define frontier models and set third-party safety-review requirements for state procurement (OregonLive).

Tools Worth a Look

• Gemini 3.8 Live with Live Avatar: speech-to-speech in 97 languages with a lip-synced face, for Gemini Enterprise (Google).
• GitHub Security Lab Taskflow agent: picks fuzzing targets in public C/C++ repos, writes harnesses and triages crashes (GitHub).
• Stripe WebMCP on Checkout: gives agents explicit tools on hosted Checkout pages instead of page scraping (Stripe).

Events (Next 30 Days)

• Sep 29: OpenAI DevDay, San Francisco (schedule)
• Oct 5–7: Ai Everything Abu Dhabi (calendar)
• Oct 7–8: World Summit AI, Amsterdam (calendar)
• Oct 20–21: AI & Big Data Expo Europe, Amsterdam (calendar)
• Oct 27–29: ODSC AI West, Burlingame (calendar)

Keep Up With the Rankings

See how every model stacks up on the LLM1 leaderboard, and subscribe to the LLM1 newsletter for this roundup every Monday.