OpenAI released GPT-5.5 and other frontier models on AWS Bedrock, along with Codex and managed agents. Enterprise developers gain access to OpenAI's latest via AWS infrastructure with native governance and security workflows.

DeepSeek released V4-Pro and V4-Flash variants on April 24, 2026, with the Pro model featuring 1.6 trillion total parameters and 49 billion active parameters supporting 1 million token context. The hybrid attention architecture targets long-horizon agentic tasks and matches frontier model capabilities at lower cost.

Moonshot AI released FlashKDA, a high-performance implementation of Kimi Delta Attention integrated with flash-linear-attention ecosystem. Benchmarks show meaningful speed improvements relevant to latency-sensitive production deployments.

Anthropic announced April 7 that Claude Mythos preview would be distributed to 11 companies and organizations for cybersecurity vulnerability testing. This represents a controlled frontier model release for critical infrastructure defense with no current plans for general availability.

Nvidia released open-source Nemotron 3 Nano Omni with 30B total parameters and 3B active parameters, optimized for on-device agentic reasoning. The model enables developers to deploy multimodal agents with inference efficiency at the edge without cloud dependency.