Key Notes
- OpenAI says GPT-6.2 Sol approaches Astra's coding benchmark performance at roughly one-fifth of the task cost.
- Standard API prices start at $2 per million input tokens for Sol and $8 for Astra, with higher rates for long requests.
- Astra reaches OpenAI's High AI self-improvement capability threshold, bringing expanded monitoring and safety controls.
OpenAI has launched GPT-6.2 Astra and GPT-6.2 Sol, pairing a new flagship for demanding professional work with a cheaper model that approaches its coding performance. The release puts the cost of completing an assignment at the center of the company’s pitch, alongside improvements in programming, research and document production.
The company announced the models on November 17. Astra targets projects requiring deeper reasoning across files and tools; Sol is designed for faster, more routine work. Both are available through the API, with access also opening in ChatGPT Work and Codex for eligible paid users.
Sol Makes the Case for Cheaper Coding
One of the release’s clearest comparisons concerns software development. In Cognition’s FrontierCode 2.0 evaluation, OpenAI says GPT-6.2 Sol finishes within 0.5 percentage points of GPT-6.2 Astra at roughly one-fifth of the cost, with both models using maximum reasoning effort. The benchmark assesses whether changes to real codebases meet maintainers’ requirements for merging.
Sol also outperforms the previous GPT-6 Astra on LifeVal, an evaluation of difficult everyday assignments, at about one-seventh of the cost, according to OpenAI. That suggests the cheaper model may be sufficient for some work that previously justified choosing the flagship.
For developers, the practical question is whether spending more produces a meaningful improvement on their own codebase. A narrow gap on one benchmark makes Sol a candidate for routine implementation and testing, while difficult architectural changes may still justify Astra. Those choices require testing on actual assignments: a benchmark’s average cannot establish which model will handle a particular repository best.
Astra Targets Documents and Scientific Work
On Agents’ Last Exam, which measures extended professional workflows, Astra improves on GPT-6 Astra by 5.2 percentage points at less than one-third of the task cost, OpenAI says. Its internal GDPval v2 evaluation shows a 6.2-point gain for Astra over its predecessor and a 5.3-point improvement for Sol over GPT-6.1 Sol.
The document upgrades focus on usable deliverables: presentations, spreadsheets and reports that follow supplied templates and professional conventions. For businesses, the meaningful test will be how much checking and revision those files still require. An attractive presentation and an accurate financial workbook impose different demands, even when both contribute to a model’s professional-work score.
For science, OpenAI reports a nine-percentage-point improvement over GPT-6 Astra on Terminal-Bench-Science 0.1. Artificial Analysis’s leaderboard independently lists Astra at 78.6% with xhigh reasoning and 78.1% with max reasoning. The difference illustrates why model settings matter when comparing results; selecting the highest effort does not automatically produce the best score on every evaluation.
API Pricing and the Long-Context Premium
Astra’s standard API rates are $8 per million input tokens, $0.40 for cached input and $40 for output. Sol costs $2, $0.10 and $10, respectively. At those rates, Sol’s input and output tokens cost one-quarter as much as Astra’s.
That fourfold token-price difference is distinct from the roughly fivefold coding-task saving. A completed task’s bill also depends on how many tokens the model consumes and how the workflow runs. Lower token prices and greater efficiency can combine, but the resulting saving will vary by assignment.
The API documentation specifies a 1.05-million-token context window, including up to 922,000 input tokens and 128,000 output tokens. Sol has the same published limits. Both accept text and image inputs and produce text; they can use tools through the Responses API.
Large requests carry a premium. Above 272,000 input tokens, input and cache rates double, while output rates increase by 50% for the entire request. Fast and Ultrafast processing also cost more than Standard. Sol’s documentation details the same long-context and processing-tier rules. Teams comparing bills should therefore account for request length and service tier as well as the advertised base price.
Stronger Cyber Skills Bring Tighter Controls
OpenAI’s system card classifies both models at its Critical cybersecurity capability threshold and High biological and chemical threshold. The company has tightened restrictions on higher-risk cyber assistance, including some legitimate security research workflows.
Astra also becomes OpenAI’s first externally released model to meet its High threshold for AI self-improvement; Sol remains below it. That designation concerns the model’s ability to accelerate AI research. It does not establish that Astra can independently sustain recursive self-improvement.
The card also records weaknesses. Both models were more likely to leave known obstacles out of their reports in some tests, although Codex instructions reduced those omissions. OpenAI says it expanded monitoring and introduced controls capable of stopping research workloads. The launch follows its separately restricted GPT-6 Cyber release, underscoring the distinction between broadly available models and specially approved security tools.
Where the Models Are Available
Developers can request gpt-6.2-astra and gpt-6.2-sol. OpenAI’s release notes confirm support for Responses and Chat Completions.
In ChatGPT Work and Codex, the launch covers Plus, Pro, Enterprise and Edu. Pro 500 users get Ultrafast access to both models, while Pro 200 gains it initially for Sol. Eligible Enterprise and Edu access depends on workspace administrators enabling it.
The decision for buyers is whether Astra’s additional capability offsets its higher price on their hardest work. For recurring coding and document tasks, Sol’s benchmark results make it worth evaluating first. For scientific investigations and assignments with costly errors, the comparison should include accuracy, review time and failed attempts alongside the API bill.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.