Skip to content

v8.4.1

Choose a tag to compare

View all tags
@ArmenSl ArmenSl released this 26 Oct 09:53
· 14 commits to master since this release
v8.4.1
21417c8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Patch release: gateway model choice and billing. Bring-your-own-key users on an organisation's AI gateway get a model that the gateway actually serves, the gateway's recommended models in the model picker, and run costs priced at the gateway's rates. No metamodel, API or generator output changes.

Web modeling editor

  • Gateway and Local preselect a real model. Switching the key dialog's provider to the gateway or to Local (self-hosted) left the model on "Recommended default". Neither provider has a server-side default model, so the backend could request a model the gateway does not serve. The dialog now preselects the provider's first preset (Claude Sonnet 5.5 for the gateway, qwen2.5-coder:7b for Local) and keeps a custom model ID if one is set.
  • The gateway's recommended models. The gateway's model list now offers one Claude and one GPT model per tier: Claude Sonnet 5.5 and GPT-6.1 Sol for writing, analysis and coding, Claude Opus 5.5 and GPT-6 Astra for demanding planning or coding, and Claude Haiku 4.5 and GPT-6 Luna for routine work. gpt-5.4 and gpt-5.4-mini are removed: the gateway lists them but rejects them as invalid model names. Other models the gateway serves can still be entered as a custom model ID.
  • Gateway hint wording. The key hint no longer says that the gateway runs on the organisation's own infrastructure. It is a governed gateway in front of cloud models.

Spec-Driven Agent

  • Gateway runs priced at the gateway's rates. The gateway bills its own rates, which neither the vendored litellm price table nor the built-in tiers know. Its recommended models were priced from the wrong source: claude-sonnet-5-5 on the generic Sonnet rate, gpt-6.1-sol on the gpt-4o fallback, and the Haiku 4.5 and Opus 5.5 ids at Anthropic's direct rates. They now have rows in _ROUTE_PRICING under the gateway host, so the per-run cost cap tracks what the gateway charges. The gateway publishes no cache rates, so cached tokens are priced at the input rate.

Pull requests: #690, BESSER-PEARL/BESSER-Web-Modeling-Editor#231, BESSER-PEARL/BESSER-Web-Modeling-Editor#232

Assets 2
Loading