Frontier
Model launches, benchmarks, multimodal systems, and eval wars.
OpenAI introduced GPT-Realtime-2 in the API, their most advanced voice model with GPT-5-level reasoning for real-time collaboration. It pairs with GPT-Realtime-Translate and GPT-Realtime-Whisper for seamless audio processing. Now available for building next-gen voice interfaces.
In response to rising AI-fueled threats, OpenAI is providing limited access to GPT-5.5-Cyber, a specialized variant optimized for vulnerability detection and patching. This follows Anthropic's Mythos debut and aims to bolster defenders.
Using Anthropic's Claude Mythos Preview, the Firefox team identified and fixed 271 vulnerabilities in April alone—more than the previous 15 months combined. Additional fixes rolled out in recent updates.
OpenAI has rolled out GPT-5.5 Instant as the default model for ChatGPT, promising fewer hallucinations in critical domains like law and medicine, plus boosted math scores (81.2% on AIME 2025). It enhances memory across chats, files, and integrations.
xAI's advanced image generation model, powering 300M+ images on Grok, now available via API with improved realism, text rendering, and control.
Gemini 3.1 Flash-Lite is now generally available, optimized for cost-efficiency, low latency, and tasks like translation and data processing with 1M token context. Priced at $0.25/M input tokens, it's ideal for scalable agentic workflows.
OpenAI launched a specialized cybersecurity version of GPT-5.5, dubbed 'Spud', to select defenders, intensifying the AI arms race in vuln discovery post-Mythos.
Z.ai (Zhipu) publishes report on GLM-5V-Turbo, a multimodal coding foundation model strong in vision-based tasks, tool use, and agent frameworks. Highlights RL and toolchain advances.
Google Cloud announces general availability of Gemini 3.1 Flash-Lite, the fastest and most cost-efficient model in the Gemini 3 series, optimized for low-latency agentic workflows like tool calling. Adopted by JetBrains, Ramp, and others for coding, customer service, and finance. Marks push into high-volume enterprise AI deployment.
Anthropic introduces 'Dreams,' a new capability allowing Claude to generate internal simulations for self-improvement and better performance over time. This could pave the way for more autonomous AI agents.
xAI adds Image Generation Quality Mode to API, powering 300M+ Grok images with improved realism, text rendering, and control. Aimed at business pros for higher fidelity outputs.
Following initial limited rollout, OpenAI is making GPT-5.5 'Spud' available to more vetted cybersecurity teams to counter emerging AI threats.
Google AI Studio updates Gemini interactions API to support rich agentic workflows with flexible step-based actions.
Zyphra launches ZAYA1-8B, a new open-source reasoning model optimized for complex tasks.
Meta postponed its next-gen frontier model release due to underwhelming performance. Sources say the model failed internal benchmarks. This highlights challenges in scaling amid competition.