On 17 November 2026, OpenAI launched its latest generation of foundation models, GPT-6.2 Astra and GPT-6.2 Sol. Pitched as a bifurcated approach to artificial intelligence, Astra is designed for demanding, agentic projects, while Sol is positioned as a lighter model for everyday workflows. However, the release has immediately highlighted a widening and contentious gap between controlled laboratory benchmarks and real-world practical utility.
On paper, OpenAI’s new suite demonstrates formidable capabilities. According to the company's release data, the top-tier Astra model recorded a striking 78.3 per cent on the 'Agents' Last Exam' benchmark. Concurrently, OpenAI aggressively expanded its enterprise footprint by opening a specialised configuration, 'Astra for Law', to all ChatGPT Enterprise customers via its Trusted Access Program. The legal model ranked first on the Harvey LAB-AA benchmark, achieving full credit on 20.8 per cent of complex assignments and edging out competing models like Muse Wonder Full.
Yet, despite these aggressive quantitative metrics, the qualitative reception from the global software engineering and research communities has been notably hostile. Independent developers testing GPT-6.2’s practical capabilities have reported severe regressions. Prominent industry critics, including developers Dan Shipper and Theo Browne, have publicly castigated the release. Browne went as far as labelling it "the worst model OpenAI has ever released." Early complaints heavily feature failures in common coding tasks and the generation of excessively dense, hard-to-read text that severely limits the tool's everyday utility.
These capability stumbles arrive at a particularly vulnerable moment for OpenAI. During the same week, Microsoft moved to reduce its reliance on its generative AI partner, releasing its own proprietary MAI Code 2 and MAI Code 2 Flash models at its Ignite 2026 conference. Microsoft's new models are specifically aimed at common enterprise coding tasks—the exact arena where GPT-6.2 Astra is currently drawing its fiercest criticism.
For technology leaders and research teams in Aotearoa and across the Asia-Pacific, the rocky debut of GPT-6.2 offers a cautionary tale. As local enterprises increasingly seek to integrate generative systems into high-stakes environments—such as legal analysis and software engineering workflows—the reliance on headline benchmark scores is proving to be a critical vulnerability. The stark contrast between Astra's 78.3 per cent agentic benchmark score and its reported failure to handle basic conversational and coding prompts underscores the absolute necessity for independent, contextual testing before wide-scale deployment.
The mixed reception to GPT-6.2 suggests that the broader research community is losing patience with foundation models seemingly optimised for testing leaderboards rather than practical, everyday reliability. As the global race for enterprise AI dominance continues, capability on paper is no longer enough to guarantee capability in practice.
Sources
- The Rundown AI; OpenAI’s GPT-6.2 draws rough early reviews
- BigGo Finance; OpenAI Opens Legal AI Model to All Enterprise Users
- Local AI Foundry; AI has started raising AI
- InfoWorld; Microsoft launches MAI Code 2 models
AI disclosure: Automated systems assisted with research and drafting. Gemini independently peer-reviewed this article before publication. Sources are linked for verification.
