AI Trends

OpenAI Releases GPT‑6.2 Astra and GPT‑6.2 Sol, Boosting Efficiency for Complex Tasks and Everyday Work

ZhidxUpdated Nov 19 08:26

AI summary

OpenAI has released GPT‑6.2 Astra and GPT‑6.2 Sol, designed respectively for challenging projects and efficient general-purpose work, and says both can complete difficult tasks with fewer output tokens. GPT‑6.2 Astra offers enhanced capabilities for presentations, spreadsheets, documents, complex coding, and 3D design. Its API costs $8 per million input tokens and $40 per million output tokens, while GPT‑6.2 Sol costs $2 per million input tokens and $10 per million output tokens. Starting today, both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.

Why read this

SelectedAI score 89

Multiple benchmarks, token usage figures, and API prices are presented side by side, enabling comparisons of the performance-cost trade-offs between the flagship and everyday models.

Topics

Tags

Body · AI translation

Zhidx reported on November 18 that OpenAI today unveiled its most powerful models yet, GPT‑6.2 Astra and GPT‑6.2 Sol. In multiple benchmarks released by OpenAI, the two models performed on par with or even surpassed other mainstream models, while costing only a fraction as much per task.

GPT‑6.2 Astra is a frontier model designed for challenging projects, while GPT‑6.2 Sol is an efficient, general-purpose model for everyday work. OpenAI said the two models deliver better results on difficult tasks with fewer output tokens, allowing users to accomplish more within their usage limits.

In the overall Artifical Analysis rankings, GPT-6.2 Astra placed second, behind only Claude Fable 5.2, while GPT‑6.2 Sol ranked fifth, behind Gemini 4 Carbon and Claude Opus 5.5. AA’s chart shows that OpenAI dominates the Pareto frontier for output tokens per task, with Claude Fable 5.2 the only model on it from a non-OpenAI lab.

OpenAI co-founder and CEO Sam Altman said on social media that although the new models are already highly intelligent, OpenAI still has more work to do to make them “easy to use and collaborate with,” and will prioritize this in subsequent releases.

OpenAI said slides, spreadsheets, and documents generated by GPT‑6.2 Astra no longer have an obvious AI feel. It can also write more concise code and build front-end interfaces with better visuals and user experiences.

In terms of API pricing, GPT‑6.2 Astra’s input and output token prices are 20% lower than those of GPT‑6 Astra, while cached input receives a 60% discount. Per million tokens, input costs $8 (about RMB 53.33), cached input costs $0.40 (about RMB 2.67), and output costs $40 (about RMB 266.65). GPT‑6.2 Sol retains GPT‑6.1 Sol’s API token pricing: $2 (about RMB 13.33) per million input tokens, $0.10 (about RMB 0.67) for cached input, and $10 (about RMB 67) for output.

Starting today, GPT‑6.2 Astra and Sol are available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. For the first time, OpenAI is making the higher-tier Ultrafast mode available to Pro 500 premium individual subscribers, initially offering it with the GPT‑6.2 Sol model.

However, users in OpenAI’s comments section were unimpressed. The top commenters complained that OpenAI’s charts were extremely difficult to read and suggested that OpenAI ask Claude Opus 5.5 to help create its bar charts.

One developer called it crazy, saying that Astra 6.2 delivers performance close to Fable 5.2 at less than half the price of Sonnet and Opus 5.5. Even more remarkably, GPT Sol 6.2 delivers Opus 5.5-level performance at just one-tenth the price—only one dollar. He remarked that Sol’s price is rapidly approaching DeepSeek’s, differing by only a few cents.

One user compared GPT-6.2 Sol and Claude Opus 5.5 using the same prompt and found that GPT-6.2 Sol cost just 1/15 as much, ran five times faster, and produced results that were every bit as good.

I. GPT‑6.2 Sol and Astra post nearly identical scores across multiple tests, while the former costs only a fraction as much

Compared with GPT‑6.1 Sol, GPT‑6.2 Sol is smarter and more efficient, enabling users to complete more tasks under the same paid plan. Many OpenAI employees have already adopted it as their default model for both technical and non-technical tasks.

In the FrontierCode 2.0 programming benchmark, GPT‑6.2 Sol and GPT‑6.2 Astra, both running at their highest configurations, scored within 0.5 points of each other, while Sol cost only about one-fifth as much.

In terms of speed and intelligence, OpenAI said GPT‑6.2 Sol is also well suited to helping people use AI for important tasks outside work. In LifeVal, a benchmark released by OpenAI to measure performance on complex tasks, GPT‑6.2 Sol scored higher than GPT‑6 Astra while costing only about one-seventh as much.

OpenAI’s main focus in improving GPT‑6.2 Astra was to make its output more professional. The company said the model writes more directly and concretely, with substantially fewer generic conclusions and less redundant content.

In Agents’ Last Exam, which tests multi-step work, GPT‑6.2 Astra set a new record, scoring 5.2 points higher than GPT‑6 Astra while costing less than one-third as much. It also led Claude Fable 5.2 by 5.5 points while costing only 5% as much.

OpenAI believes GPT‑6.2 Astra is particularly well suited for scientists. The model set a new high on Terminal-Bench-Science 0.1, which evaluates models’ ability to conduct scientific research in a terminal environment, scoring 9 percentage points higher than GPT-6 Astra while costing less than one-third as much. The model also broke the record on the Vals Science Index, which asks models to interpret real biological data and conduct experiments to infer how simulated systems work.

Harvey LAB‑AA v1.1 tests whether models can find relevant facts in source files in legal scenarios, apply them to detailed case summaries, and produce deliverables that meet specified requirements. In this test, GPT‑6.2 Astra’s pass rate was more than twice that of Claude Fable 5.2, while costing only about one-sixth as much.

In addition, OpenAI has created a legal version of GPT-6.2 Astra specifically for lawyers, which scored even higher on this benchmark.

OpenAI’s internal GDPval v2 benchmark measures complex, long-horizon workflows drawn from real projects in industries including consulting, finance, and design. In its tests, GPT‑6.2 Astra ranked first, improving by 6.2 percentage points over GPT‑6 Astra while using 22% fewer output tokens. GPT‑6.2 Sol improved by 5.3 points over GPT‑6.1 Sol.

Writing an investment memo requires first identifying relevant data in source files and then understanding the relationships among the data. The AA‑AnalystAgent benchmark tests this work using real spreadsheets and reports. In one task, the model had to use California Medicaid expenditure reports to calculate how much the program spent on medical transportation. Across all five independent attempts, GPT‑6.2 Astra answered 62.5% of the questions correctly.

II. OpenAI: The best model yet for presentations is here—and it can even reproduce font sizes from screenshots

OpenAI demonstrated applications of the two models in several key scenarios.

OpenAI said GPT-6.2 Astra is its best-performing model ever for presentation design. Using an original file or a single screenshot, the model can directly reproduce the layout, color scheme, typography, and even exact font sizes.

Early testers reported that presentations generated by GPT-6.2 Astra were ready to use, required fewer revisions, stated the key conclusion directly in the title, and featured concise body text.

▲ Reproducing a presentation from a screenshot

For computer use and interactive web design, OpenAI said GPT‑6.2 Astra can work down to individual parts and connection points in CAD and simulation tools. Engineers can use this finer level of detail to check how components fit together, observe how loads transfer as a design moves, and test modifications before building a prototype.

In one test, OpenAI asked GPT‑6.2 Astra to visualize the chassis dynamics of a car driving on a mountain road. Astra built a 3D model, wrote a Python solver to simulate the complex dynamics between the tires and suspension, and then rendered an animation of the results in Blender.

GPT‑6.2 Astra can also model mechanical systems component by component, giving engineers a way to show how systems are assembled and test how design changes affect their performance.

OpenAI researchers also used GPT‑6.2 Astra to transform the scalloped balconies of Chicago’s Marina City towers into a table lamp. The animation below documents the process from reference photos and CAD revisions to the hand-assembled physical lamp. GPT‑6.2 Astra also generated a website for creating custom paper lamps.

For website creation, OpenAI said the model has stronger visual reasoning capabilities. It can select a distinctive look, layout, and visual assets for each website, then build page elements around the site’s actual objectives.

In 3D design, both GPT‑6.2 Astra and Sol can turn written requirements into interactive browser experiences. In Artificial Analysis’s AA‑3D test, human judges evaluated two blinded experiences generated from the same prompt and voted for the one they preferred. GPT‑6.2 Astra ranked first with an Elo score of 1318, followed by Claude Fable 5.2 at 1271 and GPT‑6.2 Sol at 1259.

In programming, GPT‑6.2 Astra excels at implementing and testing complex changes in existing codebases. Early testers reported cleaner coding style, more focused tests, and fewer rounds of revision.

In cybersecurity, OpenAI said GPT‑6.2 Astra and GPT‑6.2 Sol can more effectively identify and exploit vulnerabilities, write more secure code, and produce higher-quality patches.

Conclusion: The efficiency curve becomes a key battleground in large-model competition

The release of GPT-6.2 Astra and Sol continues the tiered strategy adopted by leading large-model companies: Astra targets scientific research, legal work, complex coding, and long-horizon agentic workflows, while Sol handles high-frequency everyday tasks. Beyond their performance advantages, OpenAI is placing greater emphasis on the models’ cost advantages. Competition among today’s models is no longer just about benchmark scores, but also about the efficiency curve.

For enterprise users, efficiency gains are narrowing the cost gap between flagship and everyday models. Combined with adjustable reasoning intensity, this makes it easier to achieve ROI at scale in areas such as high-frequency coding, document processing, customer service, and internal knowledge assistants. The cost advantage of leading models is most apparent in their ability to use fewer tokens when executing complex tasks.

Source: OpenAI

Source

Zhidx