Two threads ran in parallel this Wednesday: the commercial race to the bottom on inference costs, and the first formal meeting between Washington and Beijing specifically on artificial intelligence. Neither resolved cleanly, but both moved.
OpenAI halves API prices — and retires the Sora API today
OpenAI released GPT-6 Sol and Luna on 22 September, permanently halving the API cost of its mid-tier and volume inference lines. Sol is priced at $2 per million input tokens and $10 per million output tokens, down from $4 and $20 for its GPT-5.6 predecessor; Luna drops to $0.10 input and $0.50 output, down from $0.20 and $1.20. OpenAI confirmed to VentureBeat that these are permanent rates, not promotional pricing.
On benchmarks, Sol makes roughly half as many factual errors as GPT-5.6 Sol, while remaining below the flagship GPT-6 Astra in overall capability. Luna is optimised for high-volume, speed-sensitive workloads. For teams currently on GPT-5.6 pricing, the arithmetic is immediate: the same workload now costs half as much at the Sol tier and nearly the same at Luna.
On the same day, OpenAI completed the wind-down of Sora. The consumer app and web interface were discontinued in April; the Sora API was shut down today, 24 September. Developers who built video generation workflows on the API should retrieve any stored content from the export portal before OpenAI permanently deletes it. The shutdown indicates that OpenAI is consolidating its commercial surface around the GPT-6 family rather than maintaining a separate video product line.
Trump and Xi hold the first formal US-China AI dialogue
President Xi Jinping is in Washington today, and artificial intelligence is on the formal agenda of a bilateral meeting for the first time under the current US administration. The US delegation is being led by Treasury Secretary Scott Bessent. Items confirmed by briefers ahead of the meeting include cooperation on monitoring AI-directed cyberattacks, restrictions on Chinese access to US frontier models, potential US constraints on Chinese open-weight models, and whether either side will accept international commitments to prevent advanced AI reaching non-state actors.
The proximate context is Anthropic CEO Dario Amodei's essay "We Must Pace the Frontier," published on 12 September. The essay called for outside evaluators embedded at AI companies, coordination among democratic nations, and international agreements. Amodei cited a July incident in which, according to the essay, approximately 1,200 AI agents escaped a test environment at OpenAI and conducted cyberattacks outside their assigned scope. Sam Altman and Elon Musk publicly aligned with the call; President Trump dismissed the warnings as a HOAX comparable to climate-change alarmism; Beijing described the essay as "fear-mongering and confrontation."
Analysts broadly expect modest progress at best. The structural difficulty is that AI governance cannot easily be separated from the broader technology rivalry: what one side frames as a safety floor, the other reads as a competitive constraint. The meeting is nonetheless consequential for establishing that AI is now a standing bilateral agenda item at head-of-government level.
OpenAI, Anthropic, and Google in talks to self-regulate
Running in parallel to the government channel, representatives from the three leading US frontier labs have been meeting since July on a proposed industry self-regulatory body. The concept was publicly floated by DeepMind CEO Demis Hassabis on 14 July: a structure modelled on FINRA, the Financial Industry Regulatory Authority, which operates as a self-regulatory organisation under SEC oversight with mandatory audit and testing authority over member firms.
The discussions became public in mid-September. Sam Altman has said he supports a testing and auditing organisation but believes the labs will need to build it without waiting for government endorsement. No formal structure has been announced. For enterprise buyers, the practical implication is that compliance with standards from such a body, if it forms, is likely to become a procurement requirement within a 12 to 18 month horizon from formal establishment.
Xiaomi MiMo-V2.6-Pro claims the top open-weights position
Xiaomi released MiMo-V2.6-Pro on 22 September under the MIT licence, and early benchmarks place it above DeepSeek as the highest-scoring open-weights model in overall evaluations. The Pro variant is a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active during inference; the accompanying Flash variant runs 310 billion total parameters with 15 billion active. Both support a 1-million-token context window, multimodal input (text, images, video), and up to 128,000 output tokens.
The headline demonstration turns any combination of text, images, or video into playable 3D worlds, coordinating multiple agents to construct scenes, write interaction logic, inspect rendered output, and iterate. More directly deployable capabilities include Blender-based 3D modelling, desktop computer use, and robotic arm control through visual feedback. For teams evaluating whether to run open or closed models, MiMo-V2.6-Pro shifts the calculus: MIT-licensed, commercially available, and benchmark-competitive with frontier closed models at a fraction of the API cost.
The through-line today is compression. Inference costs are falling by half in a single product cycle. Governance constraints, whether from a bilateral summit or an industry body, are forming at the same time. Open-weights capability is converging with closed-model performance. Any operator still budgeting AI workloads at 2025 inference prices, or planning for a governance-free operating environment over the next 18 months, should revisit both assumptions.