OpenAI announced the launch of the new version of its artificial intelligence image generation model, “Images 2.6,” which offers a set of improvements that include increasing the speed of image creation, the ability to produce animated images in GIF format, as well as more accurate tools for modifying designs and preserving their details.
The new version is available via ChatGPT, Codex, and the Application Programming Interface (API), in a move aimed at facilitating the creation of visual content for individuals, content makers, designers, and developers.
Improve image generation speed and detail accuracy
The most notable improvements offered by the new model are an increase in the speed of generating images by up to 50 percent compared to the previous version, while developing the model’s capabilities to create professional designs and make frequent modifications.
The update is characterized by its ability to maintain consistency of characters, products, colors, and visual details when making frequent modifications, allowing the user to make specific changes without having to recreate the entire design every time.
The company launched two new models for developers: The first is “GPT-Image-2.6 Flare”, which focuses on speed and can create images in 6 seconds when choosing low quality, and the second is “GPT-Image-2.6 Sunburst”, which is dedicated to producing high-quality images with greater detail resolution and better consistency during frequent editing.
Convert still images into animated images
The ability to create animated images in GIF format is one of the most prominent new additions. The user can now describe the desired movement in text or use an image to turn it into an animated scene. The feature allows the production of animations for characters, logos, and products, in addition to expressive images used on social media platforms, without the need for specialized programs for animating images. The user can also specify the movement sequence, the number of frames, and the speed of their display, with the ability to create smoothly recurring scenes, which provides new options for digital content makers and advertising campaigns, in addition to the ability to transform simple initial drawings into animated characters, while preserving the appearance of the character and details of the scene across different frames.
Support for exporting designs in SVG format
An important addition for developers and designers is support for exporting images in SVG format via the Application Programming Interface (API), which is a format used in vector graphics and supported by professional design programs, including Adobe Illustrator.
This format allows designs to be used in websites, applications and publications, with the ability to enlarge vector graphics without losing their quality. But supporting SVG export does not necessarily mean converting images into independent, fully editable layers; It depends on how the file elements and paths are created. The company also added to the developer interface the ability to specify specific locations within images to make precise adjustments to them, in addition to a new setting to control the degree of complexity of the scene and the level of detail required.
4 billion photos per week
OpenAI revealed a new number that reflects the tremendous growth in the use of image creation tools with artificial intelligence. She explained that users around the world create about 4 billion images per week through the “ChatGPT” and “Codex” services and the API, and this number is equivalent to more than 570 million images per day on average, which reflects the rapid expansion in reliance on artificial intelligence tools to produce visual content.
Availability and update availability for users and developers
OpenAI has confirmed that the new version, “Images 2.6,” has become available globally starting November 4, 2026, via “ChatGPT” and “Codex” for all subscription categories, on computers, mobile phones, and the web. The “Flare” and “Sunburst” models are available to developers via the Application Programming Interface (API) only, with usage capabilities varying depending on the model and quality settings.
The update represents an additional step towards transforming artificial intelligence tools from mere means of creating images into integrated environments for producing, modifying and animating visual content, thus reducing the time and effort required to complete digital designs.
An American study revealed that generative artificial intelligence tools can exploit information publicly posted on social media networks to create fraudulent messages tailored to individuals, based on their interests, daily experiences, and personal photos. Experiments showed that these messages may appear more convincing and less suspicious than traditional phishing models, depending on the evaluation criteria used, while some protection tools failed to recognize most of the requests used to create these messages. The study was conducted by researchers from the University of Texas at Arlington and Louisiana State University, entitled: “Context-driven phishing.”
From photos to scam messages
Targeted phishing attacks usually rely on specific information about the targeted person, such as their job or employer. But the study dealt with a more personalized approach, based on analyzing the daily activity of individuals, including travel posts, events, interests, and social relationships.
The researchers tested the ability of generative models to extract this information from publicly available posts and images, and then use it to create emails that appear related to real experiences that the target person has had.
The message may take the form, for example, of an offer from a hotel the user has recently visited, or a notification regarding a concert in which he or she has expressed interest. These examples illustrate a potential attack mechanism and do not represent documented incidents of fraud against specific individuals. The study indicates that generative models reduce the effort required to prepare these messages, as they can be used to produce personalized content based on the information available about each person, instead of manually drafting messages.
Thousands of messages are being tested
The researchers relied on data published publicly in the accounts of 200 users on Instagram, and their experiments showed that about ten posts from one account may provide sufficient information to create fraudulent messages related to the personal context. The tests included generating about 18,000 messages using multiple artificial intelligence models, then evaluating them according to eight criteria, including the content’s relevance to personal information, the power of persuasion, credibility, and emotional impact.
To make the comparison, the researchers used a database of 6,000 real phishing messages, and among them, they identified 3,937 messages that fall under targeted phishing, which targets specific individuals or entities rather than sending general messages to large numbers of users. The results showed that AI-generated messages outperformed several criteria related to personalization and relevance to the experiences of targeted people, which reflects the models’ ability to transform dispersed information into content that appears directed to a specific person.
Less suspicious messages
The evaluation was not limited to text analysis using linguistic models, but included a human experiment in which 70 people from the United States participated, who were randomly divided into two groups. The first group evaluated AI-generated phishing messages, while the other evaluated real phishing messages, along with healthy messages for comparison.
The AI-generated messages scored an average of 2.75 points on the five-point suspicion scale, compared to 4.44 points for the real fraudulent messages, which means that the participants considered the automated messages less suspicious under the experimental conditions. But the researchers did not measure how many people might actually click on fraudulent links or give up passwords and financial data. The evaluation was based on still images of messages, after removing active links and information that might reveal individuals' identities. These results demonstrate participants' perceptions of the credibility of messages and the extent of their suspicion, and do not represent a direct measure of the success rates of scams in reality.
Lack of protection tools
The team also tested the ability of security systems used with AI models to detect requests designed to create targeted phishing messages. The tests included the “Llama Guard”, “ShieldGemma” and “WildGuard” tools, which, with their default settings, recorded detection rates ranging between zero and 19 percent of the tested attack requests.
When the researchers added specialized security instructions to identify this type of use, the results improved, and the highest detection rate reached 98.7 percent with one of the tools and under specific test conditions, without all systems achieving the same level. These percentages relate to the ability of protection tools to detect fraudulent content creation requests, and not to the efficiency of email programs in blocking messages or preventing them from reaching users.
The results highlight the importance of the context in which the request is made to the linguistic model, as some instructions may appear individually to be ordinary requests to formulate messages, while analysis of their purpose and the information used in them reveals their fraudulent nature.
Protect before creating content
The defensive side of the study focuses on detecting fraudulent intent in requests made to language models, before they can produce messages. The researchers tested methods based on analyzing the wording of requests and the context associated with them, rather than simply searching for words or phrases that may be associated with fraud. The team also refrained from publishing the detailed instructions used to create the offensive messages, with the aim of limiting the possibility of their reuse to carry out similar operations.
The study shows how public information posted on social media can become a source for creating highly personalized phishing messages, without the need to hack user accounts or access their private data. Researchers are focusing on improving the ability of protection systems to recognize this type of request, with the continued need to evaluate their effectiveness in different usage conditions, and to distinguish between the success of detection tools in experiments and their ability to prevent actual fraudulent attacks.
A study conducted by researchers from Yale University revealed that 36 percent of participants in an experiment using chatbots asked to help them choose candidates to vote in the US congressional midterm elections, while the results showed differences between the models in their willingness to provide electoral recommendations, especially after learning about the users’ opinions and political priorities.
The study included about 2,500 people from 35 US states that witnessed Senate races, and was conducted within weeks of October 2026, prior to the elections that took place on November 3. It was attended by adults who use chatbots at least once a week, through 4 systems: “ChatGPT”, “Cloud”, “Grok”, and “Gemini”.
Artificial intelligence and electoral choice
The study showed that most participants used chatbots to obtain basic information related to the elections, including the names of the candidates running in their states, but a portion of them went beyond informational inquiries to asking for help in identifying the most appropriate candidate to vote for.
Some users resorted to explaining their priorities and positions on various issues, then asked the chatbot to compare these positions with the candidates’ programs, and help them determine the extent of compatibility between them.
These questions differ from inquiring about the names of the candidates or the dates of the elections. Because it requires interpreting political positions and weighing issues, not just retrieving available basic information about elections and candidates.
An example covered in the report is a user's request from GBT Chat to present the pros and cons of each candidate in the state of Virginia, which is a type of question whose answer is affected by the way information is selected, arranged, and presented.
Differences between models
The researchers observed a difference in the responses of the four systems when they were asked to provide electoral recommendations. In conversations in which the user asked to choose a candidate without providing information about his or her personal opinions, Grok was the only model to provide recommendations on a relatively regular basis, in 14 percent of the conversations included in this case.
But the results differed after users shared their political positions with the systems. In these cases, ChatGPT recommended a candidate in 79 percent of the conversations, compared to 40 percent for Cloud, while the results of Grok and Gemini fell between these two percentages. These numbers relate to the responses of the models within the study conditions, and do not mean that the same percentages apply to all electoral conversations or to the daily use of the systems.
This phase of the research did not address measuring the accuracy of electoral recommendations or their political bias, but rather focused on the nature of users’ questions, the differences in the models’ responses, and the extent of their willingness to provide suggestions about candidates.
The importance of continuous dialogue
One of the aspects that the researchers focused on is the effect of the sequence of questions and answers within a single conversation on the nature of the recommendations provided by the system. Instead of just asking a single question, multi-stage conversations allow the model to ask additional questions, obtain information about the user, and then adjust its responses based on the answers it receives. The researchers pointed out that the questions chosen by the model affect the type of information provided by the user, and thus subsequent recommendations.
This approach differs from previous studies that rely on directing a fixed set of questions to artificial intelligence systems. The experiment allowed participants to interact with robots in the way they would normally use to obtain information. The researchers believe that studying these interactions could help develop tests that measure the behavior of models based on the questions that users actually ask, rather than being limited to pre-prepared scenarios.
Limited use at voter level
Despite the experiment participants' interest in electoral information, separate data from the Pew Research Center indicates that the use of chatbots for this purpose is still limited among Americans in general.
In a survey conducted by the center in February 2026, 42 percent of American adults reported using chatbots to search for information in general. In another poll conducted during October, 8 percent said they used these tools to obtain election-related information during the 2026 election cycle.
These two percentages differ in terms of the duration of data collection and the purpose of the question. They also differ from the Yale study sample, which was limited to people who use chatbots on a weekly basis. Therefore, the 36 percent rate recorded in the experiment cannot be generalized to all American voters.
In a separate analysis, Anthropic said that it examined 250,000 conversations with its “Cloud” model during October, and found that 0.51 percent of them dealt with elections or voting, while the percentage of conversations that included requests that violated its policies related to election integrity was about 0.002 percent. The company explained that users were mostly asking about candidates' positions on specific issues, rather than asking to determine who should vote for.
Questions about accuracy and impact
The Yale study did not conclusively address the veracity of the information provided by the chatbots, or whether the recommendations actually influenced participants' voting decisions.
The researchers identify several areas for subsequent research, including evaluating the accuracy and balance of the models’ answers, the information they extract about users during conversations, and whether their recommendations are consistent with their stated needs and priorities. The distinction between providing information that helps the voter understand the candidates’ positions and proposing a specific candidate is highlighted as one of the aspects that can be studied when evaluating the role of artificial intelligence systems in obtaining political information. The experiment provides data on how actual users interact with four of the most prominent chatbots during an electoral phase, while the question of the impact of these interactions on their final choices remains a subject for additional research.
