Gemini 3.7 Flash and GPT-5.6 Sol Ultrafast Race for Speed

Two Labs, One Race for Speed

Google and OpenAI have both rolled out faster, cheaper variants of their flagship AI models this week, underscoring a growing industry emphasis on latency and cost as much as raw capability. Google's Gemini 3.7 Flash is publicly available now and pitched as an efficient backbone for building low-cost AI agents, while OpenAI's answer, GPT-5.6 Sol running in a new "Ultrafast" mode, remains locked behind an invite-only waitlist 1.

What Ultrafast Actually Delivers

OpenAI describes Ultrafast as a preview mode that lets GPT-5.6 Sol, its most powerful model, respond up to 14 times faster than standard operation 23. The push is explicitly aimed at enterprise customers who need near-instant responses for production workloads, rather than casual consumer chat use 2. Coverage of the launch frames it as OpenAI's attempt to court businesses that have previously balked at the latency of large, heavily reasoning-capable models, positioning speed as a competitive differentiator against rivals like Google 23.

Google's Contrasting Approach

Whereas OpenAI is restricting early access to Ultrafast, Google has pushed Gemini 3.7 Flash out broadly, emphasizing affordability and throughput for developers building autonomous agents rather than headline-grabbing raw intelligence gains 1. The divergent rollout strategies — one open, one gated — highlight differing bets on how to introduce high-speed models: Google wants adoption at scale immediately, while OpenAI appears to be testing demand and stability among select enterprise partners before a wider release 1.

A Broader Wave of GPT-5.6 Variants

The speed-focused release is only one piece of a larger GPT-5.6 rollout. OpenAI has also introduced GPT-5.6-Cyber, a specialized model aimed at cybersecurity work, alongside an expansion of its Daybreak platform to broaden organizational access to its tools 4. Notably, GPT-5.6-Cyber ships with reduced safeguards relative to OpenAI's general-purpose models, and reporting indicates it can complete roughly 95% of advanced cyber-related requests, including identifying zero-day vulnerabilities 6. That combination of capability and loosened restrictions has drawn scrutiny given the dual-use potential of a model explicitly tuned for exploit development 6.

Context From Earlier Releases

These moves follow OpenAI's earlier launch of GPT-5, which the company touted as a major step in multimodal reasoning, integrating text, image, and video understanding with a claimed 40% improvement in complex problem-solving over GPT-4 5. Taken together, the recent announcements suggest OpenAI and Google are now iterating rapidly along multiple axes simultaneously — general reasoning, domain-specific specialization, and raw inference speed — rather than treating any single model as a singular flagship release. For enterprises and developers, the practical question emerging from this cluster of announcements is less about which model is smartest and more about which combination of speed, cost, and access restrictions best fits real-world deployment.