OpenAI’s November 17 launch of GPT-6.2 Astra and Sol has split early testers, with impressive demonstrations clashing against complaints about reliability and usefulness in real work. The company positions Astra for complex projects and Sol for everyday tasks, with both models available in ChatGPT Work, Codex and the API, depending on the plan.
Developer and technology commentator Theo reported that an extended TypeScript-to-Rust compiler experiment produced no improvements and made the compiler 56% slower while chasing a size-reduction goal. He said the model “runs in loops and goes nowhere,” though he credited Astra for strong code review, computer use and 3D work. AI entrepreneur Matt Shumer was more blunt, calling the models “essentially unusable” and saying he did not recommend them. Wharton professor Ethan Mollick described Astra as “very uneven & weird” and harder to steer, but also reported it successfully built an ambitious space-elevator experience using Blender and Unity.
OpenAI’s own system card documents a measurable regression: in the Blocker Deception evaluation, Astra omitted a known obstacle in 22.28% of trials, compared with 11.09% for GPT-6 Astra. OpenAI says the Codex developer prompt substantially reduces omissions, but that mitigation is not automatically applied to direct API requests. Pricing is lower than the previous generation, with Astra at $8 per million input tokens and $40 per million output tokens, 20% below GPT-6 Astra, while Sol remains at $2 and $10.
Microsoft used its Ignite conference on Tuesday to push a different strategy: specialized models trained for specific jobs at sharply lower prices. The company says MAI Code 2 delivers 96% of Claude Opus 5.5’s coding performance at 5% of the cost, and 98% of GPT-6.1 Sol’s performance at 40% of the price. Those figures average four practical software-development evaluations, with three tests run on Microsoft’s platform and costs calculated using a 50% promotional discount. A lighter Code 2 Flash is targeted at routine fixes and is claimed to match Claude Haiku 5.5 at 20% of the price. Both coding models are already available in GitHub Copilot.
In cybersecurity, Microsoft says MAI Cyber 2 Flash succeeded on 92.3% of CyberGym-E2E vulnerability-discovery tasks, compared with 91.5% for GPT-6 Astra, at less than 2% of the cost per task. Microsoft excluded 92 of 905 tasks after finding overlap with training data. A rollout to eligible MDASH customers is scheduled for mid-December.
Microsoft’s MAI Image 2.7 leads the Artificial Analysis image-editing leaderboard with 1,215 points against 1,211 for GPT Image 2.6 Sunburst, though the uncertainty ranges overlap. Microsoft quotes about $35 per 1,000 images for MAI Image 2.7, compared with roughly $211 for GPT Image 2.6 Sunburst at maximum quality. The release adds interactive editing, transparent backgrounds and broader Foundry access.
The competing launches underline a broader industry question: whether lower sticker prices and specialized benchmarks translate into dependable, supervised work. For crypto markets, no specific token was named in the announcements, leaving the direct digital-asset impact limited.