AI developer tools · What shipped

Gemini 3.6 Flash, optimized for code generation and agentic execution, supports an input token limit of 1,048,576 and output token limit of 65,536. The model includes capabilities for code execution, function calling, structured outputs, and thinking, along with support for caching, batch API, and flex inference. Latest update was July 2026.

oMLX version 0.5.4.dev1 adds a context benchmark that measures usable context window on Mac, implements prefill priority controls for choosing between maximum context or maximum throughput, and makes long-context memory admission more predictable. The release adds support for Laguna S-2.1 DFlash with a measured 2.03x speedup using adaptive speculative decoding, includes Xiaomi MiMo V2.5 text support, and delivers more representative throughput benchmarks using natural text with enforced prompt lengths. Performance improvements include 4x-4.6x faster heterogeneous embedding batches and a 1.7-3.1% decode speedup for GLM-5.2 long-context workloads on M3 Ultra hardware.

The Model Context Protocol (MCP) published its 2026-07-28 specification, introducing major changes including stateless operation over standard HTTP infrastructure, eliminating the need for session management and sticky sessions at load balancers. The protocol now includes request-level metadata for protocol version and client capabilities, removes handshake requirements, and introduces a server/discover method for capability queries. Additional improvements include support for caching with TTL metadata, distributed tracing via W3C Trace Context headers, multi-round-trip requests replacing persistent SSE streams, and a new extensions framework enabling independent governance of protocol extensions. Error handling now separates transport failures (real HTTP status codes) from application-level outcomes, and the schema system now supports full JSON Schema 2020-12 vocabulary including composition keywords and conditionals. Amazon Bedrock AgentCore Gateway now supports 2026-07-28 alongside earlier versions (2025-03-26, 2025-06-18, 2025-11-25), with version selection happening per-request rather than through handshake, allowing independent upgrade timelines for clients and servers.