OpenAI Launches GPT-6.2 Astra and Sol With Lower API Costs
OpenAI's two new GPT-6.2 models bring sharper coding, computer use in engineering tools, and professional outputs to developers and businesses at lower API costs.
Key points
- OpenAI released GPT-6.2 Astra and Sol, two models that complete tasks with fewer output tokens, reducing cost per job for developers and businesses.
- GPT-6.2 Astra API rates are 20% lower than GPT-6 Astra's, with a 60% discount on cached input tokens, while Sol carries forward GPT-6.1 Sol's pricing.
- On the FrontierCode 2.0 coding benchmark, Sol scored within 0.5 points of Astra at about one-fifth the cost, making it a practical choice for most coding work.
- GPT-6.2 Astra is the first model OpenAI has released externally that meets the High threshold for AI Self-improvement under the company's Preparedness Framework.
OpenAI has released two AI models at once: GPT-6.2 Astra, aimed at the most demanding professional work, and GPT-6.2 Sol, a faster option built for everyday tasks. Both generate fewer output tokens than their predecessors to finish the same jobs, which lowers the cost per task for developers and businesses that build on the OpenAI API, according to OpenAI.
The release covers ChatGPT Work and Codex as well as the OpenAI API. OpenAI said both models are live immediately for subscribers on the Plus, Pro, Enterprise, and Edu tiers.
GPT-6.2 Astra Sets a New Professional Quality Bar
The model also finished 5.5 points clear of Claude Fable 5.2 on that same benchmark while running at about 5% of that rival's per-task cost, OpenAI reported.
Scientific work saw particular improvement. GPT-6.2 Astra set a record on Terminal-Bench-Science 0.1, a test of how well a model handles science tasks from a command-line environment. Its score came in 9 percentage points above GPT-6 Astra's, again at below one-third the cost per task, OpenAI said. It also set a new high on the Vals Science Index, which measures a model's ability to interpret real biological data and run experiments inside a simulated system.
Professional documents and presentations also improved. OpenAI trained Astra to follow the conventions that practitioners in law, sales, finance, and accounting actually apply. Slides match provided templates down to font size and color, written output is more direct, and titles lead with a clear point rather than a vague summary. Early testers said the finished documents required fewer rounds of editing before they were ready to use.
Sol Handles Everyday Work at a Fraction of the Price
OpenAI said many employees at the company rely on Sol as their standard model for both technical and non-technical work.
On Cognition's FrontierCode 2.0 coding benchmark, Sol's score trailed Astra's by just 0.5 points while costing about one-fifth as much per task, OpenAI said. That narrow gap means most coding jobs do not require the pricier frontier model. On OpenAI's LifeVal evaluation, which measures results on challenging everyday tasks, Sol's score surpassed GPT-6 Astra's at roughly one-seventh the per-task cost.
Sol's API pricing matches GPT-6.1 Sol's: $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens, OpenAI said. For Pro 200 subscribers, OpenAI is extending Ultrafast speed to that tier for the first time, starting with Sol.
Coding and Computer Use Get Sharper
Software developers who work inside large existing codebases should notice the clearest gains from Astra. Modifying established code - finding bugs, adding new functions, checking that changes do not break other parts of a system - has been an area where earlier AI tools often fell short. OpenAI said early testers found cleaner code and more focused tests, with fewer revision rounds, after moving to GPT-6.2 Astra.
Computer use in engineering software advanced as well. GPT-6.2 Astra can now interact with CAD and simulation tools at the level of individual components and connections, rather than treating a design as a single unit, OpenAI said. That precision lets engineers check how parts fit, observe where structural loads shift as a design changes, and test revisions before constructing a physical prototype. OpenAI demonstrated this by asking Astra to simulate suspension dynamics for a car on a mountain road; the model built a 3D scene, wrote a Python solver, and animated the output in Blender.
Three-dimensional design benchmarks reflected similar gains. On Artificial Analysis's AA-3D test, human judges compared browser experiences built by competing models from the same prompt. Astra led the standings at an Elo score of 1,318, with Claude Fable 5.2 second at 1,271 and Sol third at 1,259, OpenAI said.
Token Efficiency Drives Lower API Costs
The cost reduction for Astra has two parts. The model uses fewer output tokens to complete a given task, so less billable text is generated per job. Beyond that, OpenAI cut input and output token rates by 20% from GPT-6 Astra's prices. Standard API rates for GPT-6.2 Astra are $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens, OpenAI said.
The largest savings apply to cached input tokens, where a 60% discount is in place. Apps that repeatedly pass the same background information to the model - such as system prompts or lengthy reference documents - will capture the most benefit. On OpenAI's internal GDPval v2 benchmark, which tests complex workflows drawn from real consulting, finance, and design projects, Astra's score rose 6.2 percentage points above GPT-6 Astra's, all while generating 22% fewer output tokens, OpenAI said.
Financial analysis tasks showed strong results as well. On Artificial Analysis's AA-AnalystAgent benchmark, which runs models through real spreadsheets and financial reports, GPT-6.2 Astra answered 62.5% of questions correctly across all five independent attempts, OpenAI said. The test included assignments such as calculating government health program spending from actual expenditure documents.
Safety Thresholds and What to Watch
Both models arrived with stricter safety controls than their predecessors. OpenAI said GPT-6.2 Astra and Sol each meet the Critical threshold for cybersecurity under its Preparedness Framework, which led the company to impose tighter limits on assistance with higher-risk security work. Those limits affect some legitimate vulnerability research and testing workflows alongside broader protections against misuse. Both models also meet the High threshold for biological and chemical capabilities under that framework, OpenAI said.
GPT-6.2 Astra is the first model OpenAI has made available outside the company that meets the High threshold for AI Self-improvement, a measure of a model's ability to speed up AI research. OpenAI described that designation as equivalent to giving each of its researchers an additional capable research engineer, compared with a 2024 staffing baseline. To manage those risks, the company built Control Room, a system that tracks active research workloads and can pause further model output when human review is needed.
As model capability and pricing both shift with each release cycle, businesses and developers have reason to revisit use cases they once set aside as too expensive. The performance gap between frontier and mid-tier models keeps narrowing, while costs on both sides continue to fall.
Sources
More from the network
Junious Media Network ↗Yellowjackets Season 4 Is the Show's Final Chapter
Showtime's survival drama Yellowjackets concludes after four seasons, with creators Ashley Lyle and Bart Nickerson saying the story is complete and no spinoff is planned.
junidaily.media ↗S&P 500 Drops Into Correction as AI Stocks Fuel Worst Week in 18 Months
AI-stock losses pushed the S&P 500 down 3.7% Friday, dragging the index into correction territory for the first time since March 2025.
junidaily.media ↗Macy's 100th parade will open with a 10-minute Broadway medley
NBC and Macy's will open the 100th Thanksgiving Day Parade on Nov. 26 at 8:30 a.m. with a 10-minute all-star Broadway medley introduced by Julianne Hough.
junidaily.media ↗Army Delays Iowa Lead Azide Blast After Toxic Plume Concerns
The U.S. Army postponed a controlled detonation of degraded lead azide in Iowa after residents and lawmakers raised safety concerns about a toxic lead plume.
junidaily.media ↗