Back to changelog
Gateway
October 30, 2026
https://www.merge.dev/changelog/october-30-2026-gateway

GPT-6.1 Terra, Gemini 4 Argon, and Mistral Large 4

Gateway adds GPT-6.1 Terra, Gemini 4 Argon, Mistral Large 4, and 33 more models

Fourteen of the 19 new chat, embedding, and transcription models run with zero data retention.

  • GPT-6.1 Terra by OpenAI, and Gemini 4 Argon and Gemini Embedding 2 by Google, all with zero data retention
  • Mistral Large 4, Mistral Embed, Voxtral Mini Transcribe 3, and Voxtral Small Transcribe 3 by Mistral, all with zero data retention
  • DeepSeek V4.1 Flash 1017 by DeepSeek, Kimi K3.1 Pro by Moonshot AI, MiniMax M3.1 Flash by MiniMax, and Qwen3.8 Flash by Qwen, all with zero data retention
  • Muse Spark Open by Meta AI, Nemotron-3-Nano-30B-A3B by NVIDIA, and Particle 1 by Particle AI, all with zero data retention
  • Mercury 2, Mercury 2.5, and Mercury 3 by Inception, solar-pro4 by Upstage AI, and Ultramarine 1 Flash by Paper Instruments

The catalog also adds 11 image, video, speech, and embedding models, including GPT Image 1, MiniMax Image-01, and Voxtral TTS 3, plus 6 decision and extraction models such as Clef and Parallel Decisions (fast).

Read more in the Gateway model catalog docs.

Build Your Own Router adds a model picker and ability to evaluate routing policies

New Build Your Own Router adds additional flexibility and customization, and new evaluation suite allows users to easily test routing policies for cost, quality, and latency trade offs

  • The Build Your Own Router model picker limits a policy to the scored models you choose
  • Evaluation suite allows you to compare multiple routing policies and models against baselines, with full view into each result and a breakdown on key attributes

Read more in the Gateway Intelligent routing docs.

‍

Improvements

  • Evaluations - Run evaluations directly within Merge Gateway to compare different models, routing policies, and evaluate cost, quality, and latency tradeoffs between each
  • Amazon Bedrock - Serves Claude Haiku 5.5 with zero data retention, plus Claude Sonnet 5.5, GPT-6 Astra, GPT-6 Luna, and GPT-6.1 Sol
  • Cloudflare Workers AI, Inception, Luminal, Novita AI, Paper Instruments, Parallel, SubQ, Upstage AI - Added as model vendors
  • DeepInfra - Serves Claude Opus 5.5, DeepSeek V4 Pro, Kimi K2.6, and several other models
  • DeepSeek V4.1 Flash - Available through DeepInfra, Luminal, Novita AI, and Pareto
  • Google Vertex AI - Serves Gemini 3 Flash Preview, Gemini 3.1 Pro Preview, Gemini 3.8 Flash TTS, and three more Gemini models with zero data retention
  • Novita AI - Serves Kimi K3, Qwen3.8 Max, MiniMax M3, and several other open-weight models
  • Model pricing - Per-request fees appear beside token prices on the models table and model pages
  • Zero data retention - Extends to Claude Fable 5, Codestral 2508, GLM-5.3, Mistral Medium 3.5, and ten more models