New model performance claims

Microsoft has rolled out a cheaper, faster version of its in-house image model as AI labs race to cut the cost of generating pictures. (techcommunity.microsoft.com) Microsoft said on April 14 that MAI-Image-2-Efficient is in public preview in Microsoft Foundry and MAI Playground. The company said the model is “up to 22% faster” than MAI-Image-2 and delivers “4x more efficiency” when measured against latency and graphics-processor use. (techcommunity.microsoft.com) MAI-Image-2 itself entered public preview on April 2, and Microsoft described it as its highest-capability text-to-image model. The company said the model already powers products including Copilot, Bing, and PowerPoint. (techcommunity.microsoft.com) Text-to-image systems turn written prompts into pictures, and the main tradeoff is usually quality versus speed. Microsoft is pitching the new “Efficient” version for high-volume jobs such as e-commerce catalogs, ad creative, and chatbot image generation where lower latency and lower graphics-processor demand matter more than squeezing out the last bit of image quality. (techcommunity.microsoft.com) The company is also using rankings to market the model. Microsoft said MAI-Image-2 debuted at No. 3 among image model families on Arena.ai when it launched in March, though the current Arena.ai text-to-image leaderboard now shows MAI-Image-2 at No. 5 behind three Google entries and one OpenAI entry. (microsoft.ai, arena.ai) That shifting leaderboard helps explain the rush to publish speed and cost numbers. Model providers are now competing on two fronts at once: public preference tests such as Arena.ai and enterprise metrics such as throughput, latency, and graphics-processor efficiency that determine how expensive a product is to run. (arena.ai, techcommunity.microsoft.com) The same week, OpenAI introduced GPT-5.4-Cyber, a restricted variant of GPT-5.4 for defensive cybersecurity work. OpenAI said on April 14 that it is expanding its Trusted Access for Cyber program to “thousands of verified individual defenders” and “hundreds of teams” while keeping access limited to vetted users. (openai.com) OpenAI said GPT-5.4-Cyber is tuned to be “cyber-permissive,” meaning it is designed to help security teams find and fix vulnerabilities while the company keeps stronger identity checks and misuse controls around access. The post ties that release to a broader push that includes cyber-specific safeguards added in 2025 and the earlier launch of Codex Security. (openai.com) The coding side of the market is moving just as quickly, but the naming is messier. Moonshot AI’s current public documentation lists Kimi K2.5 as its flagship model and says it delivers state-of-the-art open-source performance in agent tasks and code, while older K2 variants remain in preview or have been discontinued. (platform.kimi.ai) Microsoft’s own Azure Foundry catalog also lists Kimi-K2.5, not K2.6, among the models sold directly by Azure. That means some of the “K2.6” talk circulating in recent posts is ahead of what the companies have formally documented on their main product pages. (learn.microsoft.com, platform.kimi.ai) The immediate story is less about one benchmark than about how AI vendors are selling models in April 2026: image systems on cost per picture, cyber models on controlled access, and coding models on whichever leaderboard updates next. (techcommunity.microsoft.com, openai.com, arena.ai)