OpenAI has previewed Ultrafast, a new API service tier that runs GPT-5.6 Sol at up to 14 times the speed of Standard processing, with OpenAI saying the Cerebras-powered system can generate up to 750 output tokens per second. The practical shift is not simply faster chat: it makes frontier-level reasoning more viable inside incident response, financial monitoring, live support and other workflows where waiting changes the decision. The first release is an early look, not a general availability promise, but the direction is clear: latency is becoming a product surface alongside capability and price. Read the official OpenAI announcement.
What else matters
OpenAI maps the enterprise move from assistance to execution
Two new OpenAI reports examine how companies and workers are adopting AI, highlighting a gap between basic assistance and frontier organisations that redesign workflows around systems capable of taking action. That distinction matters for buyers: deployment discipline and process change, not model access alone, will determine whether usage produces measurable gains. Read OpenAI’s enterprise reports.
OpenAI adds a revenue executive as commercial pressure rises
OpenAI says Dali Rajic will join as chief revenue officer and lead its global revenue organisation. The appointment is a small announcement with a larger implication: as model costs fall and infrastructure commitments rise, distribution, pricing and enterprise conversion are becoming as strategically important as research progress. Read OpenAI’s appointment announcement.
Gemini 3.7 Flash reaches GitHub Copilot
GitHub’s changelog says Gemini 3.7 Flash is now available in GitHub Copilot. The immediate implication for developers is more model choice inside an existing coding workflow; the broader signal is that model competition is increasingly expressed through distribution channels such as IDEs and agent surfaces, not only standalone chat products. Read GitHub’s Copilot changelog entry.
New research tests capability transfer at inference time
A paper submitted to arXiv on 12 August introduces “AI4AI at Test-Time”, studying whether harnesses and prompting strategies can transfer capability from stronger systems to weaker ones at inference time. It is research, not a product claim, but it points toward a potentially cheaper route to better performance: improving the scaffolding around a model rather than scaling the model itself. Read the arXiv paper.
*Published 13 August 2026. Sources are linked directly; analysis is clearly marked as interpretation of the cited announcements and paper.*
