OpenAI has unveiled a flagship model for complex professional work and an efficient model for everyday tasks. Businesses and developers can choose based on the difficulty and cost of tasks such as writing documents and coding.
The flagship model is GPT-6.2 Astra, and the efficient model is GPT-6.2 Sol. OpenAI says Astra outperformed its previous flagship on complex task evaluations, while API input and output prices fell by 20%.
1. Astra for professional tasks, Sol for everyday work
Astra focuses on polished deliverables such as slides, spreadsheets, and documents. OpenAI says it better reflects legal, sales, finance, and accounting conventions and follows the layout, colors, and font sizes of supplied templates.
OpenAI says it also improved the ability to implement and test complex changes in existing codebases, the source code collections that make up programs. Web and 3D design capabilities were also strengthened.
Sol handles multistep tasks while conserving output tokens. OpenAI says many employees use Sol as their default model for both technical and nontechnical work.
OpenAI says Astra scored 6.2 percentage points above its previous flagship on GDPval v2, an internal evaluation of complex, lengthy tasks, while using 22% fewer output tokens. Sol also improved by 5.3 percentage points.
2. Science task evaluation presents performance and cost together
Terminal-Bench-Science 0.1 tests science tasks in a terminal, a command-based computer interface. OpenAI reports scores of 77.1% for Astra and 68.1% for its previous flagship at maximum reasoning effort. Costs are per task in this test.
Table: Terminal-Bench-Science 0.1 scores and cost per task
Conditions: Terminal-Bench-Science 0.1, reasoning effort (the depth of a model's thinking to find an answer) set to Max, as reported by OpenAI
Units: Score (proportion), cost per task (dollars)
| Model | Score | Cost per task |
|---|---|---|
| GPT-6.2 Astra | 0.771 | 7.15 |
| GPT-6 Astra | 0.681 | 23.8 |
| GPT-6.2 Sol | 0.68 | 2.21 |
| GPT-6.1 Sol | 0.57 | 5.47 |
OpenAI says Astra scored 9 percentage points higher at less than one-third the cost per task of its previous flagship. Sol scored close to that flagship at a lower cost per task, making it an option for cost-sensitive work.
OpenAI's chart lists Claude Fable 5.2 at 73.3% and $45.21 per task with maximum reasoning effort in mini-swe-agent, the evaluation runner. Anthropic reported 74.6% at maximum reasoning effort in its own coding tool, a different setup.
3. API pricing and availability
Astra's API input and output rates are 20% lower than the previous flagship's, and cached input rates for reused prompts are 60% lower. Sol retains its previous version's rates. Standard API prices for both models are shown below.
Table: Standard API prices for both models
Conditions: Standard API pricing, per 1 million tokens
Unit: Dollars
| Item | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Input | 8 | 2 |
| Cached input | 0.40 | 0.10 |
| Output | 40 | 10 |
At standard rates, Sol is cheaper, while Astra offers both lower rates than before and reduced output token use.
Both models are available to Plus, Pro, Enterprise, and Edu users of ChatGPT Work and Codex. Eligible Enterprise and Edu users gain access when workspace administrators enable the feature.
Developers can call gpt-6.2-astra and gpt-6.2-sol through the API.
Ultrafast, a high-speed execution feature, is also available through the API for both models at launch.
Pro 500 subscribers can use Ultrafast with both models, while Pro 200 subscribers get access to Sol first.
4. Stronger cyber capabilities and safeguards
OpenAI rated both models' cybersecurity capabilities as 'Critical' and biological and chemical capabilities as 'High' under its Preparedness Framework. It tightened limits on high-risk cyber help and dual-use biology and cyber requests.
OpenAI says Astra is its first publicly released model to reach the 'High' threshold for accelerating AI research. It equates this to giving each of its researchers in 2024 an additional capable research engineer.
OpenAI says safeguards reduce misuse but also limit some legitimate security testing and research into vulnerabilities, software flaws that attackers can exploit. Two safety evaluations also found more omissions of task blockers. Additional Codex instructions sharply reduced these omissions.
5. Prompt reuse as a cost management tool
To increase cache reuse, OpenAI recommends keeping the start of a prompt, the instructions and context sent to a model, unchanged and appending new instructions. It says this lowers costs and latency and preserves reasoning between turns.
This release gives developers Astra for stronger professional performance and Sol for affordable everyday work, with both task requirements and security restrictions to consider.
