OpenAI is reportedly planning to release the next generation of its artificial intelligence (AI) model before the end of the year. As per a report, the next frontier model of the company will be significantly more powerful and capable than the GPT-4 AI model. The large language model is said to be internally called Orion. While the company is reportedly planning for a December release, it is said that the AI model will not be released in the public domain first. Instead, it could first be accessed by the enterprises OpenAI closely works with.

OpenAI’s Next Generation of AI Model

The Verge reported that the AI firm is looking at a timeline of December 2024 for the launch of the next generation of its frontier large language model. Citing people familiar with the matter, the publication claimed that the model is internally being called Orion. Notably, earlier it was said to be Strawberry, but that turned out to be the GPT-4o AI model.

While OpenAI has released the GPT-4 Turbo and GPT-4o AI models after the release of GPT-4, neither of these were new AI models. They were simply upgraded and tweaked versions of the LLM, built on the foundation of the GPT-4 architecture. The next iteration of its AI model, which could be called GPT-5, is expected to include a new architecture and capabilities.

As per the report, the codenamed Orion AI model will not be released directly to the public, unlike the previous AI model releases by the company. Instead, the AI model will reportedly be shared with companies that OpenAI works closely with. The reason for this is said to be that the AI firm wants to let the enterprises build their own products and features before the model is available in the public domain.

Microsoft could be one of them, given that it is a major investor in the company. Interestingly, the report also claims that the Orion model could be hosted on the Azure servers by November. However, the official name of the AI model is not confirmed as of yet.

Additionally, the report claims that the Orion AI model could be up to 100 times more powerful than GPT-4. It is also expected to introduce improved AI agentic capabilities. Eventually, OpenAI wants to merge its AI models together to create artificial general intelligence (AGI), the report added.

Google announced a significant upgrade for Gemini, its in-house artificial intelligence (AI) model, on Wednesday. The company announced that the image generation capability of the chatbot will now be handled by the Imagen 3 AI model for all users. Imagen 3 is the Mountain View-based tech giant’s latest and most capable image generation model. Apart from the Gemini app, the feature is also being extended to the API version of Gemini to let developers build apps and experiences based on this capability.

Gemini Users Get Access to Imagen 3 AI Model

In a post on X (formerly known as Twitter), the official handle of the Google Gemini App revealed that all users, including those on the free tier, will be able to generate images using Imagen 3. The post highlighted that the AI model offers a high degree of photorealism, better prompt adherence, and adds fewer unwanted elements to images.

Gadgets 360 staff members were able to verify that the Gemini app is indeed using Imagen 3 to generate images. To test its capabilities and compare it with Meta AI, we gave both chatbots the same prompt. The prompt was, “Draw an image of a golden retriever dog sitting on a train berth, looking out through the window at the Alps. The train has a wooden interior and the seats are green in colour. All other passengers on the train are also animals. One human conductor is checking for tickets.”

Meta AI vs Gemini

 

The generated images can be seen above. While both AI models failed to incorporate one or more elements instructed in the prompt, Gemini was able to incorporate more elements. Additionally, while Meta AI generates images in 1280 x 1280 resolution, Imagen 3 images are generated in 2048 x 2048 resolution.

Imagen 3 can generate images in a wide range of styles such as photorealistic, textured oil paintings, and claymation scenes. Users can also request images to appear as if it has been taken from a specific camera such as a Nikon DSLR, GoPro style, wide-angle lens, and more.

Google has said that the AI model comes with inbuilt safeguards to reduce the risk of deepfakes. Every generated image also comes watermarked with SynthID, a technology that adds an invisible AI label within the pixels of the image. It cannot be cropped out or removed and is present even in screenshots.

Google announced a significant upgrade for Gemini, its in-house artificial intelligence (AI) model, on Wednesday. The company announced that the image generation capability of the chatbot will now be handled by the Imagen 3 AI model for all users. Imagen 3 is the Mountain View-based tech giant’s latest and most capable image generation model. Apart from the Gemini app, the feature is also being extended to the API version of Gemini to let developers build apps and experiences based on this capability.

Gemini Users Get Access to Imagen 3 AI Model

In a post on X (formerly known as Twitter), the official handle of the Google Gemini App revealed that all users, including those on the free tier, will be able to generate images using Imagen 3. The post highlighted that the AI model offers a high degree of photorealism, better prompt adherence, and adds fewer unwanted elements to images.

Gadgets 360 staff members were able to verify that the Gemini app is indeed using Imagen 3 to generate images. To test its capabilities and compare it with Meta AI, we gave both chatbots the same prompt. The prompt was, “Draw an image of a golden retriever dog sitting on a train berth, looking out through the window at the Alps. The train has a wooden interior and the seats are green in colour. All other passengers on the train are also animals. One human conductor is checking for tickets.”

Meta AI vs Gemini

 

The generated images can be seen above. While both AI models failed to incorporate one or more elements instructed in the prompt, Gemini was able to incorporate more elements. Additionally, while Meta AI generates images in 1280 x 1280 resolution, Imagen 3 images are generated in 2048 x 2048 resolution.

Imagen 3 can generate images in a wide range of styles such as photorealistic, textured oil paintings, and claymation scenes. Users can also request images to appear as if it has been taken from a specific camera such as a Nikon DSLR, GoPro style, wide-angle lens, and more.

Google has said that the AI model comes with inbuilt safeguards to reduce the risk of deepfakes. Every generated image also comes watermarked with SynthID, a technology that adds an invisible AI label within the pixels of the image. It cannot be cropped out or removed and is present even in screenshots.

Gemini 1.5 Flash-8B, the latest entrant in the Gemini family of artificial intelligence (AI) models, is now generally available for production use. On Thursday, Google announced the general availability of the model, highlighting that it was a smaller and faster version of the Gemini 1.5 Flash which was introduced at Google I/O. Due to being fast, it has a low latency inference and more efficient output generation. More importantly, the tech giant stated that the Flash-8B AI model is the “lowest cost per intelligence of any Gemini model”.

Gemini 1.5 Flash-8B Now Generally Available

In a developer blog post, the Mountain View-based tech giant detailed the new AI model. The Gemini 1.5 Flash-8B was distilled from the Gemini 1.5 Flash AI model, which was focused on faster processing and more efficient output generation. The company now claims that Google DeepMind developed this even smaller and faster version of the AI model in the last few months.

Despite being a smaller model, the tech giant claims that it “nearly matches” the performance of the 1.5 Flash model across multiple benchmarks. Some of these include chat, transcription, and long context language translation.

One major benefit of the AI model is its price effectiveness. Google said that the Gemini 1.5 Flash-8B will offer the lowest token pricing in the Gemini family. Developers will have to pay $0.15 (roughly Rs. 12.5) per one million output tokens, $0.0375 (roughly Rs. 3) per one million input tokens, and $0.01 (roughly Rs. 0.8) per one million tokens on cached prompts.

Additionally, Google is doubling the rate limits of the 1.5 Flash-8B AI model. Now, developers can send up to 4,000 requests per minute (RPM) while using this model. Explaining the decision, the tech giant stated that the model is suited for simple, high-volume tasks. Developers who wish to try out the model can do so via Google AI Studio and the Gemini API free of charge.