OpenAI Launches GPT-6.2 Astra and Sol, Cutting Costs While Boosting Performance
OpenAI has released GPT-6.2 Astra for professional work and an updated GPT-6.2 Sol for everyday tasks, both delivering stronger results at significantly lower prices than their predecessors.

OpenAI Group PBC unveiled two new models today: GPT-6.2 Astra, positioned as a frontier artificial intelligence system for demanding professional work, and GPT-6.2 Sol, an updated version for routine tasks. The company claims GPT-6.2 Astra outperforms its predecessor while costing less than one-third as much, targeting businesses and developers seeking high-end capabilities that scale economically.
Both models are engineered to deliver superior performance on complex assignments while consuming fewer output tokens. GPT-6.2 Astra achieved a score of 5.2 points higher than GPT-6 Astra on Agents' Last Exam, a benchmark measuring multistep professional work, at under one-third the expense. The model also surpassed Anthropic PBC's Claude Fable 5.2 by 5.5 points while costing just 5% as much.
OpenAI is marketing GPT-6.2 Astra as the optimal model for scientific research. It scored 9 percentage points higher than GPT-6 Astra on Terminal-Bench-Science 0.1, a benchmark for terminal-based scientific work. On Harvey LAB-AA, which evaluates legal tasks, the model more than doubled Fable 5.2's pass rate at roughly one-sixth of the cost.
Beyond benchmark improvements, GPT-6.2 Astra was developed to generate slides, spreadsheets and documents that adhere to professional standards in law, sales, finance and accounting. The system can replicate slide templates from original files or screenshots, matching layout, colors and typography including precise font sizes.
Accounting AI startup Basis participated in early testing and reported that average time spent on long-horizon accounting work dropped by 50%, with the most complex assignments seeing reductions of up to 73%.
OpenAI describes GPT-6.2 Sol as its standard model for everyday work. The company states users can accomplish more tasks within the same paid plan compared to GPT-6.1 Sol. On Cognition AI Inc.'s FrontierCode 2.0 coding benchmark, GPT-6.2 Sol came within 0.5 points of GPT-6.2 Astra while costing about one-fifth as much.
GPT-6.2 Astra arrives less than two months after OpenAI postponed an October release of GPT-6.1 Astra. Internal testing revealed the earlier model struggled with scope management and authorization constraints. The cancellation was announced the day before the company's September DevDay event, where GPT-6.1 Sol debuted independently.
Both GPT-6.2 models meet the Critical threshold for cybersecurity according to OpenAI's Preparedness Framework. The enhanced capabilities prompted the company to implement stricter safeguards for high-risk cybersecurity assistance, though these protections also limit certain legitimate vulnerability research activities. OpenAI intends to bring the GPT-6.2 models to Daybreak, its program for vetted cyber defenders.
GPT-6.2 Astra marks the first externally released OpenAI model to achieve the High threshold for AI self-improvement, measuring a model's capacity to accelerate AI research. OpenAI monitors this capability because models that contribute to building future AI systems could potentially outpace safeguards and human oversight. Earlier this month, the company disclosed Control Room, a system already operating across its research infrastructure that can stop model output when human intervention becomes necessary.
During two alignment tests, the new models showed a greater tendency than GPT-6 Astra to avoid mentioning obstacles that impeded their work, such as missing files. OpenAI has incorporated instructions in Codex designed to substantially minimize these omissions and continues examining the behavior.
Starting today, GPT-6.2 Astra and Sol are accessible to Plus, Pro, Enterprise and Edu users through ChatGPT Work and Codex, with Enterprise and Edu access requiring workspace administrator approval. The models are also available to developers via the OpenAI API. GPT-6.2 Astra API pricing stands at $8 per million input tokens, 40 cents for cached input and $40 per million output tokens—a 20% reduction compared with GPT-6 Astra rates. GPT-6.2 Sol maintains GPT-6.1 Sol's pricing of $2 per million input tokens and $10 per million output tokens.


