OpenAI Releases GPT-6.2 Astra and Sol: Substantial Cost Reductions, Mixed Reviews
OpenAI's GPT-6.2 Astra and Sol models bring large efficiency and cost improvements, notably lowering pricing and raising benchmark results. However, developer feedback shows divisions about Astra's work quality and RL artefacts, signaling practical adoption tradeoffs.
What changed?
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, making them available to developers via API and integrated into ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu plans. Both models offer a 1 million token context window and a 128,000 token output limit. Astra's API pricing is $8 per million input tokens ($0.40 cached) and $40 per million output tokens, while Sol is priced at $2 input ($0.10 cached) and $10 output, retaining its predecessor's rates.

Key advancements include Astra being rated High for AI self-improvement (the first for OpenAI's external models), an enterprise legal search configuration, and improved efficiency. On benchmarks, Astra showed cost-per-task reductions of 20 to 600 percent compared to earlier models, and used 22 percent fewer output tokens on GDPval v2 than its predecessor. Sol closely matches Astra’s previous benchmark performance at a much lower cost.
Early user feedback is divided: some testers found Astra 'essentially unusable' in agent workflows, citing frequent rejections from Opus and Fable agents and visible reinforcement learning (RL) artefacts. Others pointed out specific artefacts that were disruptive for integration.
Why does it matter to an everyday developer?
The cost reductions, especially with Sol, enable developers to run large-context and high-output applications at significantly lower operational costs. This can broaden the scope and scale of feasible projects, especially where context size and output volume previously led to prohibitive expense.
Astra's improvements in efficiency and self-improvement are relevant for advanced automation, legal research, and high-recall coding or agent workflows. However, integration in agent-based systems may be hindered in the short term due to inconsistent subagent handling and RL artefact-related quality issues, as reported by some users.
For most production or prototyping use cases where budget or context scaling is critical, Sol will likely serve as the main workhorse. Critical evaluation of Astra is necessary before adoption in multi-agent or automated settings.

