BTC $64,559.09 -0.03%
ETH $1,813.26 -0.14%
BNB $594.10 +0.61%
XRP $0.9755 +2.18%
SOL $74.69 +4.30%
TRX $0.3079 +0.92%
DOGE $0.0595 +1.16%
ADA $0.1534 +0.81%
BCH $210.25 +1.27%
LINK $9.22 +1.31%
HYPE $69.14 +0.15%
AAVE $111.77 +2.74%
SUI $0.7311 +2.25%
XLM $0.1405 +2.35%
ZEC $957.22 +6.08%
AAPL $350.19 -1.03%
AMZN $259.45 +0.19%
GOOGL $283.25 -0.52%
MSFT $508.86 -2.19%
META $621.16 -0.81%
NVDA $161.66 +3.10%
TSLA $383.92 +0.82%
SNDK $832.04 +2.01%
INTC $56.28 +3.88%
SPCX $143.03 +3.68%
MU $739.83 +1.65%
AMD $518.87 +1.32%
BTC $64,559.09 -0.03%
ETH $1,813.26 -0.14%
BNB $594.10 +0.61%
XRP $0.9755 +2.18%
SOL $74.69 +4.30%
TRX $0.3079 +0.92%
DOGE $0.0595 +1.16%
ADA $0.1534 +0.81%
BCH $210.25 +1.27%
LINK $9.22 +1.31%
HYPE $69.14 +0.15%
AAVE $111.77 +2.74%
SUI $0.7311 +2.25%
XLM $0.1405 +2.35%
ZEC $957.22 +6.08%
AAPL $350.19 -1.03%
AMZN $259.45 +0.19%
GOOGL $283.25 -0.52%
MSFT $508.86 -2.19%
META $621.16 -0.81%
NVDA $161.66 +3.10%
TSLA $383.92 +0.82%
SNDK $832.04 +2.01%
INTC $56.28 +3.88%
SPCX $143.03 +3.68%
MU $739.83 +1.65%
AMD $518.87 +1.32%

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

Core Viewpoint
Summary: OpenAI released GPT-6.2 Astra and Sol: Astra rose to a score of 62 in Artificial Analysis, ranking second globally, with a 20% reduction in API input and output prices, capable of replicating PPT, performing financial modeling, and Blender 3D modeling; however, the AA-Omniscience hallucination rate is 59%, the programming intelligence index is on par with the previous generation, and security has reached the "critical threshold" for cyber attacks.
Tencent Technology
2026-11-18 09:02:17
OpenAI released GPT-6.2 Astra and Sol: Astra rose to a score of 62 in Artificial Analysis, ranking second globally, with a 20% reduction in API input and output prices, capable of replicating PPT, performing financial modeling, and Blender 3D modeling; however, the AA-Omniscience hallucination rate is 59%, the programming intelligence index is on par with the previous generation, and security has reached the "critical threshold" for cyber attacks.

Article | Xiaojing

Editor | Xu Qingyang

On November 17, local time in the United States, OpenAI officially launched GPT-6.2 Astra and GPT-6.2 Sol.

This release came later than expected. As early as the end of October, some users had discovered traces of a suspected GPT-6.2 in ChatGPT usage records, and since then, model identifiers and testing screenshots have continuously leaked. Previously, the media had predicted that the new model would be released as early as November 6, but the final release date was 11 days later than this expectation.

OpenAI emphasized the ability of AI to take over professional work this time.

From replicating a complete presentation from a PPT screenshot, to extracting financial data, building valuation models, operating Blender to generate 3D models of cars, simulating the physical motion of vehicle suspensions, and even using CAD to design mechanical parts and verify assembly relationships, GPT-6.2 has begun to delve into tasks that previously required professionals to use complex software.

OpenAI particularly highlighted that the new model can reduce ineffective outputs and repeated modifications, making the results delivered by AI closer to professional work standards.

In the past two months, OpenAI has successively launched models such as GPT-6 Astra, Sol, and Luna, as well as GPT-6.1 Sol and Terra. The newly released GPT-6.2 Astra is an upgrade of the flagship model following GPT-6 Astra in September, focusing on enhancing the handling of complex professional tasks; GPT-6.2 Sol mainly strengthens the efficiency of daily work, programming, and multi-step agent tasks.

According to the latest list from Artificial Analysis, Astra rose to second place globally with a score of 62, an increase of 5 points from the previous generation; Sol reached a score of 57. In terms of pricing, both the input and output unit prices for Astra's API were reduced by 20%, and the price for cached input dropped by 60%; Sol maintained its original price, but the overall cost of completing certain tasks further decreased.

It is noteworthy that GPT-6.2 Astra is also the first publicly assessed model from OpenAI to reach the "High" capability threshold for AI autonomous improvement.

However, the impressive demonstrations and scores have also attracted mixed reviews. On the day of the release, several developers who had tested the new model weeks in advance publicly expressed disappointment, believing that Astra might still stop prematurely during long-term software development tasks, and that overly concise outputs sometimes miss necessary information.

At the same time, as the model's capabilities come closer to real professional work, the boundaries of its capabilities and safety controls have also become important highlights of this release.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

OpenAI CEO Sam Altman. Image generated by AI

01 Performance and Cost

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

The API price for GPT-6.2 Astra is $8 per million input tokens and $40 per million output tokens, with cached input priced at $0.40. Compared to the previous generation, both input and output prices have decreased by 20%, and cached input has dropped from $1 to $0.40, a reduction of 60%. GPT-6.2 Sol maintains a price of $2 per million input tokens and $10 per million output tokens.

For multi-step agents, the API unit price does not fully reflect actual expenses. When the model executes tasks, it may require multiple calls to tools, reading files, handling errors, and retrying. Even if the unit price remains unchanged, if fewer tokens are needed to complete the task, the final bill will decrease accordingly.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

In the FrontierCode 2.0 programming test by Cognition, GPT-6.2 Sol scored only 0.5 points lower than Astra, with costs about one-fifth of the latter. In the LifeVal assessment aimed at daily complex tasks, Sol's score surpassed the previous flagship GPT-6 Astra, with the cost of individual tasks being about one-seventh of the latter.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

Astra also showed improvement in high-difficulty tasks. In the Agents' Last Exam assessment, it improved by 5.2 points compared to GPT-6 Astra, with the cost of completing tasks being less than one-third of the previous generation. Compared to Claude Fable 5.2, Astra scored 5.5 points higher, with the cost of individual tasks being about 5% of the latter.

02 Delivery of Professional Work

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

Professional work in finance, law, and consulting requires AI to handle a large amount of raw data, verify information, and produce deliverables according to industry standards. OpenAI stated that GPT-6.2 Astra has improved in these tasks, capable of performing financial modeling, merger analysis, legal research, and data science analysis.

OpenAI's internal GDPval v2 test covers fields such as consulting, finance, and design, with Astra's score improving by 6.2 percentage points compared to the previous generation, and output tokens reduced by 22%.

In legal research, in the Harvey LAB-AA v1.1 test simulating real law firm tasks, Astra's pass rate exceeded that of Claude Fable 5.2 by more than double, with the cost of individual tasks being about one-sixth of the latter.

OpenAI's internal GDPval v2 also covers consulting, finance, and design work. The test results showed that Astra's score improved by 6.2 percentage points compared to the previous generation, with output tokens reduced by 22%.

03 Entering the 3D World

Another capability demonstrated by GPT-6.2 Astra is direct involvement in 3D modeling and engineering design. The model no longer just generates a segment of modeling code; it can also operate professional software, handle parts, adjust structures, and create visual content from computational results.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

GPT-6.2 Astra simulating mountain road driving and chassis dynamics

In the mountain road driving simulation case, Astra first established a 3D model of a sports car, then wrote a Python physics solver to calculate tire ground force and dynamic changes in the suspension system, and finally imported the calculation results into Blender to generate vehicle driving animations. Through simulation, engineers can observe load transfer situations when the vehicle traverses complex road conditions before manufacturing a physical prototype.

The input data for the mechanical modeling case was much simpler. OpenAI only provided public photos and assembly manuals of a 1:8 scale remote-controlled off-road vehicle and a straight-six engine. Based on this, Astra established 3D models of various mechanical parts and generated step-by-step assembly animations and illustrated parts catalogs.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

A test also linked CAD design with actual production. The OpenAI team provided photos of the fan-shaped balcony of the Chicago landmark Marina City for the model to reference in its design. After several adjustments, Astra ultimately created a paper art table lamp that can be printed and assembled, and wrote a supporting webpage called Lamp Builder for users to adjust parameters online and generate custom drawings.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

GPT-6.2 Astra completing mechanical modeling of the remote-controlled off-road vehicle and engine

In web development, Astra created an interactive configurator for a camper van. Users can switch the body color and add a rooftop tent, with the page quote updating in real-time. It also generated playable pinball tables, a model of the Roman Colosseum, and a futuristic city racing game called "Meridian Sound Racing" based on text descriptions.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

From the Chicago Marina City building to the paper art table lamp

The AA-3D human evaluation blind test by Artificial Analysis also compared the model's 3D generation capabilities. GPT-6.2 Astra scored 1318 Elo points, ranking first; Claude Fable 5.2 scored 1271 points, and GPT-6.2 Sol scored 1259 points.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

The interactive configurator for the camper van generated by GPT-6.2 Astra

In the AA-Robotics v0.1 preview testing physical world understanding and action planning capabilities, Astra and Sol scored 79 and 78 points respectively, higher than Claude Fable 5.2's 75 points. However, the evaluation team stated that the leading models are now capable of completing the vast majority of basic tasks in the initial tests and are increasing the difficulty of the questions.

04 Scientific Reasoning and Shortcomings

Scientific research was also a key focus of OpenAI's demonstration this time.

In the Terminal-Bench-Science 0.1 test, GPT-6.2 Astra improved by 9 percentage points compared to the previous generation, ranking first. This test mainly assesses the model's ability to complete scientific workflows in terminal environments. In the Vals Science Index, which requires the model to interpret real biological data, design experiments, and infer system mechanisms, Astra also set a new high.

The results of several other tests were slightly inferior to Claude Fable 5.2. In the high-difficulty subset of the physical reasoning benchmark CritPt v2, the factual knowledge assessment AA-Omniscience v1.1, and the reasoning benchmark HLE-Diamond, Astra ranked second.

One issue that still exists with the model is that it may provide incorrect answers when faced with uncertain facts.

GPT-6.2 Released: Can replicate PPT, build valuation models, do 3D modeling, but the hallucination rate is still as high as 59%

The hallucination rate test of Artificial Analysis's AA-Omniscience shows that at the highest reasoning intensity, the hallucination rates for GPT-6.2 Astra and Sol are 59% and 55%, respectively. This test primarily examines whether the model tends to guess answers when faced with factual questions that it may not have the answers to. It does not represent the model's error rate in general Q&A, but for scenarios like financial auditing and legal references, manual verification is still necessary.

The programming tests did not show a comprehensive improvement either. In the Artificial Analysis programming agent index, GPT-6.2 Astra's score is on par with its predecessor, while GPT-6.2 Sol has dropped 3 points compared to GPT-6.1 Sol.

The detailed results show that Sol's score loss mainly comes from SWE-Atlas-QnA-AA, with a decline of 14 points. This test assesses the model's understanding and Q&A ability regarding complex codebases. The test results indicate that Sol is more likely to overlook key facts when answering questions related to codebases compared to its predecessor.

This is a point that developers who need to handle large legacy codebases should pay attention to. The model can write code, but that does not mean it has accurately grasped the project's history, dependencies, and specific constraints.

05 Security Boundaries

As the model can operate terminals, review code, and call professional software, the cybersecurity risks have also increased.

OpenAI disclosed that both GPT-6.2 Astra and Sol have reached the "critical threshold" for cybersecurity set by the company's Preparedness Framework. Both models can more efficiently discover and exploit software vulnerabilities, as well as write safer underlying code and generate patches that retain the original business logic as much as possible.

This means that the same capability can be used for vulnerability scanning and may also be used for attacks. As a result, OpenAI has strengthened the restrictions on the public version for high-risk cybersecurity tasks.

For security researchers who need to conduct vulnerability testing, these restrictions will also affect some normal work. OpenAI is currently tiered opening related capabilities through the Daybreak project: defenders can use the lower rejection rate GPT-6 Sol through Daybreak Blue; red team testers with advanced authorization can call the alpha version of GPT-6 Astra and GPT-6 Cyber in Daybreak Red. OpenAI has stated that the GPT-6.2 series will also be introduced into this system in the future.

Additionally, OpenAI has incorporated defense-oriented training signals during the training phase. Previously, AI had been used to discover high-risk vulnerabilities such as PHP2Shell. The GPT-6.2 series further enhances the ability to identify security flaws from the development source and learns to generate patches that balance security and functional integrity.

ChainCatcher reminds readers to view blockchain rationally, enhance risk awareness, and be cautious of various virtual token issuances and speculations. All content on this site is solely market information or related party opinions, and does not constitute any form of investment advice. If you find sensitive information in the content, please click , and we will handle it promptly.

app_icon
ChainCatcher Building the Web3 world with innovations.