Back to changelog

Claude Fable 5.2 and Anthropic message threads

Route to Claude Fable 5.2, Qwen3.8-27B, and seven more models

Three of the nine run with zero data retention.

  • Claude Fable 5.2 by Anthropic
  • Qwen3.8-27B by Qwen
  • MiniMax H3 Highspeed and MiniMax H3 Lite by MiniMax for video output, both with zero data retention
  • joey-1 by Base Labs, with zero data retention
  • ScaleDown Classify 1, ScaleDown Compress, ScaleDown Extract 1, and ScaleDown Summarize 1 by ScaleDown

Read more in the Gateway model catalog docs.

/v1/messages continues Anthropic message threads from the previous response ID

After the first request, send only the new turn and the previous response ID, and Anthropic keeps the conversation in between. Threads run on the direct Anthropic route, and a continued thread stays on the kind of key that created it. Read more in the Gateway multi-turn conversations docs.

Organizations, projects, and routing policies can keep Claude clients on Claude models

claude_only_for_harnesses keeps Claude Code, Claude Desktop, and Agent SDK traffic on Claude models, and it is off by default. Set it through /v1/organization, /v1/projects/{id}/controls, and /v1/routing-policies. Read more in the Gateway Claude Code docs.

OpenAI explicit prompt caching reaches GPT-5.6 and later through /v1/openai/chat/completions

prompt_cache_breakpoint on text parts and the top-level prompt_cache_options pass through to OpenAI. Cache writes on these models bill at the catalog cache-write rate. Read more in the Gateway prompt caching docs.

Improvements

  • Clef Flash, Muse Spark Open, MiMo V2.6 Flash - Served through Makora, with Clef Flash on /v1/decisions
  • DeepSeek V4.1 Flash 1017 - Served through 12 inference providers, including Alibaba Cloud, Baseten, and Fireworks AI
  • Mercury 2, Mercury 2.5, Mercury 3, Ultramarine 1 Flash - Adds zero data retention
  • Model listings - OpenAI, LangChain, and AI SDK model lists show only the models and aliases an API key and its organization can route
  • Models API - Host-only vendors such as Baseten, Fireworks AI, and Together AI return their own logos
  • Prompt injection protection - Block mode rules apply whether or not ML screening is available
  • ScaleDown - Added as a model vendor
  • Telnyx - Added as a model vendor serving Kimi K3, GLM-5.3, DeepSeek V4.1 Flash, and four more models
  • Tool calling - Tool schemas adapt to what each vendor accepts
  • Uptime history - Model performance history reports uptime in full hourly buckets across customer traffic and probes
  • Enhanced API key lifecycle, CLI sign-in validation, and media request handling.