各家疯狂出新模型，OpenAI今天正式发GPT-5.6，谷歌的Gemini 3.5 Pro快了，中国也有新动作，竞争太激烈了。
OpenAI于今日向公众全面发布GPT-5.6，此前仅对合作伙伴开放，并应美国政府要求延迟上市。GPT-6预计约一个月后发布，基于全新预训练基础而非Spud架构。谷歌Gemini 3.5 Pro仍在内测中，目标提升推理深度和智能体性能，计划7月发布。MiniMax正准备2.7T参数模型M3 Pro，DeepSeek筹备V4 GA，中国当局讨论限制先进模型海外访问。
Frontier releases are piling up fast. OpenAI is r…
Frontier releases are piling up fast.
OpenAI is responding to Fable 5 with GPT-5.6, scheduled for full public release today after an initial limited rollout to vetted partners. The US government had requested the delay over national security concerns, a pattern seen earlier with Fable 5. According to the Leo, GPT-5.6 will close out the 5.x series, with GPT-6 expected in roughly a month on an entirely new and significantly larger pretrain base rather than extending the Spud foundation used for GPT-5.5 and 5.6. They state OpenAI sees the new base as necessary to compete directly with Fable 5 and the anticipated Fable 5.1.
Google’s Gemini 3.5 Pro remains in testing with a public release now targeted for July (maybe next week?). Early indications and leaks suggest efforts to strengthen reasoning depth, agentic performance and context handling to close the gap on the current leaders. Without material gains on long-horizon coding and multi-step tasks, the model risks leaving Google further behind on frontier capability benchmarks where Anthropic and OpenAI currently set the pace and Grok now a clear third place.
Chinese labs are adding scale and policy variables to the picture. MiniMax is preparing a 2.7T params model internally called M3 Pro for potential Q3 release, with open-source plans. DeepSeek is preparing V4 GA, positioned as competitive with or ahead of GLM-5.2 on relevant tasks. Recent reporting shows Chinese authorities have discussed restricting overseas access to the most advanced models, both closed and open-weight versions. The discussions, involving major firms and drawing on a May roundtable, floated a tiered approach under which frontier models could face domestic-only limits or additional security reviews.
The combination of rapid US releases, government oversight on the highest-capability models, and shifting Chinese policy on openness creates a more fragmented and accelerated environment. Timelines are compressing as labs react to each other’s moves on base models and agentic performance, while national security considerations are now directly influencing rollout speed and access for both American and Chinese frontier systems.
Interesting times ahead. Some good, some bad.