OpenAI's first open-weights model in years has been delayed until later this summer, as CEO Sam Altman announced on X that the company needs more time following an unexpected breakthrough by their research team that will make the model "very very worth the wait," despite originally targeting an early summer release date.
In March 2025, Sam Altman announced via X that OpenAI would release "a powerful new open-weight language model with reasoning in the coming months," marking the company's first open-weights release since GPT-2 in 2019.12 Altman emphasized the company's excitement about making this "a very, very good model" and invited developers to provide input on how to make it "maximally useful."2 This announcement represented a significant shift in OpenAI's strategy, with Altman previously acknowledging that the company had been "on the wrong side of history" regarding open models.34
The announcement came amid growing competitive pressure from companies like Meta, Mistral, and DeepSeek, whose open-weight models have gained significant traction.3 Altman assured users that despite being open-weight, the model would undergo rigorous evaluation according to OpenAI's preparedness framework, with "extra work" being done knowing the model would be modified post-release.2 To engage the developer community, OpenAI planned events in San Francisco, Europe, and the Asia-Pacific region where participants could interact with prototypes and provide feedback.5
OpenAI's o3 model has delivered several unexpected breakthroughs that likely contributed to the company's decision to delay their open-weights release. In late 2024, o3 achieved a remarkable 75.7% score on the ARC-AGI-Pub benchmark, demonstrating a qualitative shift in AI capabilities rather than just incremental improvement1. The system has shown unprecedented reasoning abilities, with researcher Sean Heelan revealing that o3 autonomously discovered CVE-2025-37899, a zero-day vulnerability in the Linux kernel's SMB implementation—marking the first time an AI independently identified such a complex security flaw2.
Beyond security applications, OpenAI has been developing specialized versions of o3, including the o3 mini expected in early 2025, which focuses on enhancing capabilities in science, mathematics, and coding3. The company has also partnered with U.S. National Laboratories for nuclear security research4 and created "Deep Research," a system combining autonomous agents with powerful reasoning abilities5. These developments suggest OpenAI's delayed open model will incorporate significantly advanced reasoning capabilities worth the extended wait.
Mistral AI has entered the reasoning model race with Magistral, released on June 10, 2025, in two versions: Magistral Small, a 24B parameter open-source model available under Apache 2.0 license, and Magistral Medium, a more powerful enterprise version accessible through Le Chat and Mistral's API.12 The French startup designed Magistral for "transparent, multilingual reasoning" with capabilities in structured calculations, programming logic, and rule-based systems.1 On the AIME2024 benchmark, Magistral Medium scored 73.6% (90% with majority voting), while the open-source Small version achieved 70.7% (83.3% with majority voting).13
While Magistral represents Europe's first reasoning model with strong multilingual capabilities across English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese, it currently lags behind competitors like OpenAI's o3 and Google's Gemini Pro 2.5 in benchmark performance.13 Mistral has integrated innovative features into Le Chat, including a "Think Mode" that, combined with Flash Answers, reportedly delivers responses up to 10 times faster than competitors.3 The company's dual-track release strategy—offering both enterprise-grade and open research options—directly challenges OpenAI's upcoming open model and positions Mistral as a significant player in the reasoning AI landscape.4
