The global competition in large AI models is intensifying this July, with major players from around the world unveiling their latest offerings almost simultaneously.
OpenAI's flagship model, GPT-5.6 Sol, is set for official release this Thursday. Elon Musk's SpaceX AI has announced that Grok 4.5 will also become publicly available tomorrow. Google's Gemini 3.5 Pro is reportedly launching on July 17th, while the official release of the domestic DeepSeek V4 is also scheduled for mid-July.
This concentrated rollout of multiple top-tier models within a few weeks directly escalates competition within the AI industry and will profoundly impact the commercialization strategies and pricing of the involved companies.
For the market, this wave of releases is not merely a contest of technical prowess but a comprehensive battle on multiple fronts including API pricing, inference efficiency, and ecosystem integration. The core question for investors is identifying which player can find the optimal balance between performance and cost, and be the first to translate technological advantages into a sustainable commercial moat.
GPT-5.6 Sol: A Flagship with Staged Rollout
OpenAI CEO Sam Altman announced on social media that GPT-5.6 Sol will be officially launched this Thursday. Positioned as the flagship product within the GPT-5.6 series, it features a new ultra multi-agent mode and max reasoning intensity, setting new records in core benchmarks for coding, biology, and cybersecurity.
Regarding pricing, Sol adopts a rate of $5 per million input tokens and $30 per million output tokens, placing it at the premium end of the three products in the GPT-5.6 series. OpenAI also announced that Sol will be launched on Cerebras hardware in July, achieving inference speeds of up to 750 tokens per second.
OpenAI is employing a phased release strategy, initially granting API and Codex access only to select trusted partners, with plans to roll out the full GPT-5.6 series to a broader user base in the coming weeks. This approach helps manage service load while also providing a window for competitors to respond.
Grok 4.5: Musk Accelerates AI Integration
Elon Musk announced on social media that, based on strong positive feedback from beta testers, SpaceX AI will make Grok 4.5 publicly available tomorrow. Musk stated the model is at the Opus level but is faster, more token-efficient, and lower cost.
Previously disclosed information indicates Grok 4.5 is built on the V9 foundational model with 1.5 trillion parameters and was supplemented with data from the AI programming tool Cursor during additional training. Early evaluations suggest its performance is close to or may even surpass Anthropic's flagship Claude Opus model, with reinforcement learning continuing to optimize it.
Notably, SpaceX AI and the programming tool company Cursor plan to jointly release their first co-developed AI model, posing a direct challenge to Anthropic and OpenAI. According to an internal memo, the model's release was previously delayed as both companies sought to further enhance its operational efficiency, but it is expected to compete with Anthropic's Opus 4.8 and OpenAI's GPT-5.5 on certain metrics.
These developments coincide with the progress of SpaceX's all-stock $60 billion acquisition of Cursor, signaling an accelerated integration of Musk's AI portfolio. Musk also stated that SpaceX will release completely new, from-scratch trained models every month for the remainder of the year.
Gemini 3.5 Pro: Rebuilt from the Ground Up, Betting on Quality
For Google, leaked information suggests Gemini 3.5 Pro will officially launch on July 17th. Reports indicate that Google DeepMind abandoned the original 2.5 Pro base and instead conducted a completely new pre-training for Gemini 3.5 Pro, pushing its release date from the originally planned June 2026 to July 17th of this year.
In terms of capabilities, Gemini 3.5 Pro is said to have made a leap in front-end and visual code generation, outperforming Anthropic's Fable 5 in several tests, though it still lags in hardcore reasoning and complex engineering tasks. This decision is seen as Google prioritizing quality over speed to counter pressure from both OpenAI's GPT-5.6 and Anthropic's Fable 5.
Additionally, Google is reportedly developing an image model, Nano Banana Pro, based on this new foundation, targeting OpenAI's GPT-Image 2. This indicates Google's intent to advance simultaneously on both the text/code and image generation fronts.
DeepSeek V4: Peak/Off-Peak Pricing and Faster Inference
On the domestic front, the DeepSeek team announced on June 29th that the official V4 release is planned for mid-July, introducing a peak and off-peak pricing strategy concurrently. According to the published price list, API rates during peak hours will be double the standard price, which remains the same as the current V4 API pricing. Peak hours are defined as 9 AM to 12 PM and 2 PM to 6 PM daily. The company stated this move aims to allocate resources more rationally and improve service stability.
On the technical side, DeepSeek, in collaboration with Peking University, released the DSpark inference acceleration framework on June 27th and simultaneously open-sourced the full-stack speculative decoding toolkit DeepSpec, with a paper authored by the company's founder, Liang Wenfeng. Real-world tests show that after deploying DSpark, the single-user generation speed for V4-Flash increased by 60% to 85%, and for V4-Pro by 57% to 78%, with the results fully validated in online services. This marks the first open-source technical achievement released by DeepSeek following its 50 billion yuan financing round.
For API users, the peak/off-peak pricing will directly increase usage costs during working hours. For developers, the significant improvement in inference speed may partially offset cost pressures in high-concurrency scenarios and further lower the barrier to implementing inference optimizations.
Comments