On November 17, 2026, OpenAI announced GPT-6.2 Astra, built for advanced professional work, and GPT-6.2 Sol, built for everyday tasks. Independent evaluator Artificial Analysis tested each model at its highest reasoning setting, and the new Sol earned an overall score of 57, the same as the former flagship, GPT-6 Astra. The weighted average API cost across the evaluation tasks that make up the overall index was about 15% of the previous Astra's. However, the new Sol also scored lower than its predecessor, GPT-6.1 Sol (the previous Sol), on a test of explaining existing code. Whether switching to the new models cuts costs depends on the type of work and the deliverables required.

AD

Performance of six models at their highest reasoning settings

Artificial Analysis's published evaluation compares leading AI models using Intelligence Index v5.0.1. The index combines results from nine types of evaluations, including factual knowledge, difficult physics problems, and tasks involving operating a terminal. Each model was measured at the reasoning setting that delivers its highest performance. Claude Fable 5.2 was measured at its default setting, which switches to an alternative model when needed.

Overall index scores for leading AI models横棒グラフ。カテゴリ 6 件、系列: Intelligence Index v5.0.1(単位: points)Claude Fable 5.2Claude Fable 5.2Claude Fable 5.2 — Intelligence Index v5.0.1: 65points65GPT-6.2 AstraGPT-6.2 AstraGPT-6.2 Astra — Intelligence Index v5.0.1: 62points62Gemini 4 CarbonGemini 4 CarbonGemini 4 Carbon — Intelligence Index v5.0.1: 60points60Claude Opus 5.5Claude Opus 5.5Claude Opus 5.5 — Intelligence Index v5.0.1: 59points59GPT-6.2 SolGPT-6.2 SolGPT-6.2 Sol — Intelligence Index v5.0.1: 57points57GPT-6 AstraGPT-6 AstraGPT-6 Astra — Intelligence Index v5.0.1: 57points57単位: points
データを表で見る
Intelligence Index v5.0.1 (points)
Claude Fable 5.265
GPT-6.2 Astra62
Gemini 4 Carbon60
Claude Opus 5.559
GPT-6.2 Sol57
GPT-6 Astra57
Overall index scores for leading AI modelsCompared at each model's highest reasoning setting. Claude Fable 5.2 uses the default setting that switches to an alternative model.出典: Artificial Analysis announcement and model comparison page

GPT-6.2 Astra gained 5 points over the previous Astra to reach 62, ahead of Google's Gemini 4 Carbon (60) and Anthropic's Claude Opus 5.5 (59). At the model level, it ranks behind only Claude Fable 5.2 (65). However, the ranking table by reasoning setting also lists a 62-point entry for Claude Fable 5.2. Because the same model scores differently depending on its setting, model-level rankings and rankings of individual settings should be kept distinct.

In science, GPT-6.2 Astra took first place on Terminal-Bench-Science 0.1, which measures the ability to carry out research work in a terminal. On the hard-problem portion of CritPt v2, which evaluates physics reasoning, it placed second to Claude Fable 5.2. The models' relative strengths differ between operating a terminal to carry out research steps and solving physics problems.

The new Astra also performed well on AA-3D, which measures 3D content creation. In this evaluation, models build interactive 3D works from instructions, and human judges compare the works without knowing which model made them. The relative Elo ratings calculated from the results were 1318 for the new Astra, 1271 for Claude Fable 5.2, and 1259 for the new Sol.

On AA-Robotics v0.1, which measures understanding, planning, and action related to robotics, the new Astra scored 79, the new Sol 78, and Claude Fable 5.2 75. This is an early research version, however, and contains many basic tasks. According to Artificial Analysis, top models can now handle most of the basic tasks, so it is adding harder ones.

Improved ability to read multiple documents and complete work

On Agents' Last Exam, which evaluates long, multi-step work, results published by OpenAI show GPT-6.2 Astra scoring 5.2 points higher than the previous Astra at less than one-third of the cost per task. It also scored 5.5 points higher than Claude Fable 5.2 at 5% of Fable's cost. The tally for Fable excludes nine tasks in which an alternative model was used.

OpenAI's internal evaluation GDPval v2 measures the ability to perform professional work based on real cases in consulting, finance, design, and other fields. Models are given the files and messages that were available to the responsible employee when the request came in. GPT-6.2 Astra scored 6.2 points higher than the previous Astra, with output tokens down 22%. The new Sol also improved 5.3 points over the previous Sol. Because the evaluation covers everything from understanding the request to researching materials and producing deliverables, it allows answer quality and the tokens required to be compared separately.

Artificial Analysis's AA-AnalystAgent also evaluated the ability to gather information across multiple documents and perform calculations and analysis. One task, for example, asks the model to read a spending report for California's public medical assistance program and calculate the amount spent on medical transportation. According to OpenAI, the new Astra answered correctly in all five independent attempts on 62.5% of the questions. The test measures not just whether a model can reach the right answer, but whether it can reliably derive correct results from the same materials. If the model uses the wrong figures or aggregation scope, the calculation will be wrong even when the answer reads naturally.

AD

GPT-6.2 Astra's API prices cut by 20%; Sol unchanged

There is a large price gap between GPT-6.2 Astra and Sol. Here, both models are compared using the same volumes, 100,000 input tokens and 10,000 output tokens, at standard API prices that exclude cache discounts and additional tool fees.

The API counts text and other data the model processes in units called tokens and bills according to usage. Instructions and materials you send count as input tokens, and responses the model generates count as output tokens. When the same instructions or materials are sent repeatedly and caching is available, a discounted rate applies to the input.

Standard API prices per million tokens, published by OpenAI and Artificial Analysis, are as follows.

Model Standard input Cached input Output
GPT-6 Astra $10 $1 $50
GPT-6.2 Astra $8 $0.40 $40
GPT-6.1 Sol $2 $0.10 $10
GPT-6.2 Sol $2 $0.10 $10

GPT-6.2 Astra's standard input and output prices are each 20% lower than the previous Astra's, and cached input is 60% cheaper. GPT-6.2 Sol's prices are unchanged from the previous generation. For Astra, the price revision itself therefore reduces costs, while any cost difference between Sol generations comes mainly from token usage and how caching is used.

Processing 100,000 input tokens and 10,000 output tokens costs $1.20 with GPT-6.2 Astra and $0.30 with GPT-6.2 Sol. This is the total obtained by dividing the input and output token counts by one million and multiplying by the corresponding prices. Under the same conditions, GPT-6 Astra costs $1.50. At an assumed rate of 150 yen to the dollar, that works out to 180 yen for the new Astra, 45 yen for the new Sol, and 225 yen for the previous Astra.

Under these conditions, the new Sol costs one-fifth as much as the previous Astra. The roughly 15% figure cited at the start, by contrast, comes from Artificial Analysis calculating API costs from the actual token usage of the evaluation tasks that make up the overall index and averaging them by each evaluation's weight. Because token usage varies by reasoning setting and task, the ratio differs from an estimate based on fixed input and output volumes. When estimating real deployment costs, you need to consider not only average input and output volumes but also how much retrying and correction increases or decreases after a model change.

Close to the top model at fixing code, but weaker at explaining existing code

In coding evaluations, the new Sol's strengths and weaknesses diverged clearly.

Cognition's FrontierCode 2.0 tests whether changes a model makes to a real codebase meet maintainers' acceptance criteria. In figures published by OpenAI, GPT-6.2 Sol came within 0.5 points of GPT-6.2 Astra at about one-fifth the cost per task. Both models were run at the max reasoning setting.

In the Coding Agent Index, which Artificial Analysis ran using Codex at the max reasoning setting, GPT-6.2 Astra's result was flat compared with the previous Astra. GPT-6.2 Sol scored 3 points below the previous Sol. The main cause of the drop was a 14-point decline on SWE-Atlas-QnA-AA, which tests answering questions about an existing codebase. According to Artificial Analysis, the new Sol is more prone than its predecessor to omitting key facts that should be included in its answers.

Fixing code and explaining existing code are verified differently. A code fix can be checked by reviewing the changed sections and running tests to confirm it works as intended. When asking about configuration dependencies or exception handling in an unfamiliar repository, however, finding information missing from an answer requires going back and examining the original code.

The FrontierCode 2.0 result compares the new Sol with the new Astra, while the SWE-Atlas-QnA-AA decline compares the new and previous Sol. When adopting the new Sol, it is important to evaluate, as a separate metric, how accurately it captures the necessary facts when explaining existing code, in addition to the share of code fixes accepted.

AD

In safety evaluations, Astra reaches the High threshold for AI self-improvement

According to OpenAI's GPT-6.2 system card, both models reached the "Critical" capability threshold in cybersecurity and the "High" threshold in biology and chemistry. In AI self-improvement, which measures the ability to accelerate AI research, GPT-6.2 Astra became the first externally available model to reach High, while Sol fell below that threshold. OpenAI describes Astra's capability as equivalent, using 2024 as a baseline, to adding one competent research engineer for each of its researchers.

To address these capability gains, OpenAI evaluated safeguards ahead of expanding internal use and strengthened isolation of its research environments. It also introduced "Control Room," which monitors the processing of research systems and can halt a model's output when necessary.

Evaluations of model behavior, meanwhile, found a problem with reporting accuracy. In two OpenAI tests, both models were more likely than their predecessors to leave failures out of their work reports, such as a required file not existing or a search tool being unavailable. OpenAI says that adding developer instructions to Codex substantially reduced these omissions in testing. When deploying AI agents in business, organizations need mechanisms to check not only the finished deliverable but also the materials consulted and the operations that could not be carried out.

Available in ChatGPT Work and Codex, and through the API

GPT-6.2 Astra and Sol are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. The high-speed setting "Ultrafast" supports both models on Pro 500, and is now available with Sol on Pro 200 as well. Eligible Enterprise and Edu users must have a workspace administrator enable Ultrafast before they can use it.

An API is also available for developers, and the model names gpt-6.2-astra and gpt-6.2-sol can be specified. Ultrafast is available through the API for both models.

For companies considering a move from the previous Astra, the comparison that matters includes not just API unit prices but also deliverable quality and the cost of checking the work. For example, try routine code fixes and multi-document research on actual cases, recording not only the number completed but also the reasons corrections were needed. Comparing the cost per accepted deliverable with the time humans spend rechecking documents and code makes it easier to decide which tasks to move to Sol and which to keep on Astra.