Back to changelog
Gateway
August 17, 2026
https://www.merge.dev/changelog/august-17-2026-gateway

A wider model catalog and stored-turn conversation continuity

Grok 4.6, Gemini 3.7 Flash, and a dozen more models join the catalog

New

Gateway now covers video and audio generation too, and most of the new models run with zero data retention.

  • Grok 4.6 (xAI), Gemini 3.7 Flash (Google), and GLM-5.3 Flash and GLM-5.3 (Z.AI) are now supported with zero data retention
  • DeepSeek V4 Pro (0813 and 0423 releases), Qwen3.8-2.4T-A95B, and Nemotron 3.5 Lightning are also new this week
  • Added Bland, NVIDIA, and Particle AI as model vendors
  • Wan3.0 Video, Grok Imagine Video 1.5, and Bland Speech v3 expand Gateway into video and audio generation
  • DeepSeek V4 Flash 0731 now routes through five inference providers, Qwen3.8 Max added zero-data-retention routing through US-based providers, and Kimi K3, MiniMax M3, and DeepSeek V4 Pro added routes

Read more in the Gateway docs.

You can now continue a conversation across requests on /v1/responses

New

In a multi-turn conversation, pass previous_response_id and the next call resumes the previous turn, so the agent keeps its earlier context instead of rebuilding it. Read more in the Gateway docs.

Improvements

  • Chat surfaces - include_routing_metadata header now returns the serving vendor and model
  • Logs - Adds Turn ID and Client trace ID filters, plus tool calls in the request drawer
  • Model discovery - Adds a claude-code base path and encoded ids for Claude Code and Desktop
  • Model explorer - Shows peak and off-peak pricing per model, with input, output, and cache rates
  • Provider warnings - Reach clients on Anthropic Messages, AI SDK, and LangChain, including the final stream frame
  • Reasoning effort - Accepts xhigh and max on Anthropic, and max on gpt-5.6 via Responses API
  • SDK compatibility - Bare paths now let SDKs use the Gateway base URL directly
  • Usage views - Display customer names instead of raw customer IDs
  • Web search - merge:web_search now routes through Parallel for lower cost and multi-query batching
  • Extended managed-key fallback to BYOK requests for Embeddings.