Days after three AI labs quietly made their models harder or more expensive to use, the same industry spent Tuesday racing in the opposite direction. OpenAI cut API prices in half on two new models. Anthropic cut its own pricing roughly two hours later. And Xiaomi released an open-source model priced at a fraction of both, closing most of the gap to frontier performance in the process.
OpenAI's New Mid-Tier: GPT-6 Sol and Luna
OpenAI released two new additions to its GPT-6 family Tuesday, Sol and Luna, joining flagship model GPT-6 Astra, which remains the company's most capable and most expensive option. As of now, there's no GPT-6 Terra to complete the set. Both new models were trained using methods similar to Astra's but optimized for speed and cost through improved caching and inference, and OpenAI says it's passing those efficiency gains directly to customers.
GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. Luna drops to $0.10 and $0.50 per million tokens, down from $0.20 and $1.20, an input cut of 50% and an output cut closer to 58%. Astra remains untouched at $10 and $50.
The capability gains aren't trivial either. OpenAI reports Sol scoring 33.2% on AutomationBench at its highest effort setting, ahead of Claude Opus 5's 26.9%, at roughly one-ninth Opus 5's cost per completed task. On DeepSWE, Sol lands just under Claude Fable 5.1 while costing about 80% less to run.
Anthropic Answered the Same Day
Anthropic cut pricing on Claude Opus 5.5 roughly two hours after OpenAI's announcement went live, close enough in timing that it reads less like coincidence and more like a company that had a price cut ready and pulled the trigger the moment a competitor moved first. The net effect: GPT-6 Sol now runs at roughly half the rate of Opus 5.5 and matches Claude Sonnet 5 on price, putting real pressure on Anthropic's mid-tier lineup specifically.
Then Xiaomi Undercut Both of Them by an Order of Magnitude
One day before OpenAI's release, Xiaomi open-sourced MiMo-V2.6, in Pro and Flash variants, releasing full model weights, a technical report, and training code under an MIT license. On Artificial Analysis's Intelligence Index, Pro scores 46, the highest of any open-weight model that firm tracks, tied with Grok 4.7, and seven points behind the strongest closed models, Claude Fable 5.1 and GPT-6 Astra.
The release also wasn't timed in isolation. It shipped the same day as Alibaba's new Zhenwu V900 chip, a pairing that reads as a deliberate signal: China assembling a complete AI stack, silicon through model, that doesn't depend on Nvidia or run into US export controls.
The Bigger Pattern Worth Watching
This is a sharp reversal from what WorkplaceAI covered just days ago: three labs quietly making access more expensive or more restricted in the same week, framed as anything but a capacity squeeze. Now the same industry is cutting prices aggressively, competing directly against each other and against an open-source Chinese alternative closing the capability gap while undercutting price by an order of magnitude.
Both patterns can be true in the same week without contradicting each other. Frontier capacity can be scarce and expensive at the top of the market while competition compresses prices hard in the middle and bottom of it. Whether cheaper tokens ultimately lower what your organization spends on AI, or just make it easier to run far more of it, driving total spending up rather than down, remains a genuinely open question analysts are actively debating.
Sources: AI Pulse · At the AI Cutting Edge · workplaceai.ai. GPT-6 Sol and Luna launch and pricing: Techzine Global, TechCrunch, gHacks Tech News, and Technology.org, September 22-23, 2026. Anthropic's same-day pricing response: Cryptopolitan, September 22-23, 2026. Xiaomi's MiMo-V2.6 release and benchmarks: Xiaomi's own announcement via X, Forkast News, and CellCog, September 21-22, 2026. Every figure and benchmark above is attributed to its original source; benchmark results are vendor-reported except where an independent index (Artificial Analysis) is specifically named, and none of this is a WorkplaceAI study.