Skip to content

feat(providers): add six AIGROW models - #13836

Open
guglielmo-io wants to merge 1 commit into
anomalyco:devfrom
guglielmo-io:add-aigrow-models-2026-11-16
Open

guglielmo-io wants to merge 1 commit into
anomalyco:devfrom
guglielmo-io:add-aigrow-models-2026-11-16

Conversation

@guglielmo-io

Copy link
Copy Markdown
Contributor

What

Adds six models to the AIGROW provider (merged in #13440 with nine): Claude Haiku 5.5, GPT-6 Luna, Qwen3.8 Max, Kimi K3.1 Pro, Gemini 4 Hydrogen and Gemini 3.8 Flash.

Why

AIGROW now serves fifteen models on the same endpoint. GET /v1/models lists these six, and the price page (https://aigrow.gg/ai-api) has a row for each.

Changes

  • Six model files, each with base_model and overrides only. Prices are the USD a buyer pays per 1M tokens, copied from the price page on 2026-11-16, and each file names its row.
    • claude-haiku-5.5: Messages first, then Chat Completions.
    • gpt-6-luna, qwen3.8-max, kimi-k3.1-pro, gemini-4-hydrogen, gemini-3.8-flash: Chat Completions only (the Messages endpoint returns 404 for the last four).
  • Overrides that differ from the lab entries, each from a request sent to the host on 2026-11-16:
    • claude-haiku-5.5: structured_output = false (a schema returned prose when the prompt did not ask for JSON, on both APIs); PDF left out (the file was counted but not read, twice on each API); cache TTL 5m only (a 1-hour breakpoint was written to the 5-minute cache); a higher price tier above 100,000 prompt tokens, as stated under the price table.
    • kimi-k3.1-pro: structured_output = false (prose when the prompt did not ask for JSON); cache = false (no cached tokens reported, and the price row says "No prompt cache").
    • qwen3.8-max: PDF left out (502 twice); video not tested, so left out.
    • gemini-4-hydrogen, gemini-3.8-flash: effort low, medium, high (xhigh returns 200 but runs as high on this host, so it is left out); audio and video not tested, so left out.
  • stream_usage = true on all six: with stream_options.include_usage the last chunk carried the usage. reasoning_field = "reasoning_content" on the four models that return their reasoning in that field.
  • Two cache write prices are written exactly (0.03125 and 0.46875); the page shows them rounded to four decimals (0.0313 and 0.4688). Both are 1.25 times the input price.
  • provider.toml: comment only. It names the two new families, the new heading of the price table, the Haiku tier and which models bill cache writes at the 1-hour rate.

bun validate passes. bun run v1:check reports the six models added to v1 and no change to any existing field.

Adds Claude Haiku 5.5, GPT-6 Luna, Qwen3.8 Max, Kimi K3.1 Pro, Gemini 4 Hydrogen and Gemini 3.8 Flash to the AIGROW provider, each with base_model and the overrides probed on the host on 2026-11-16. Updates the provider comment for the new model families and the price table heading.
@github-actions

Copy link
Copy Markdown
Contributor

Action items

  • [high] [violation] providers/aigrow/models/qwen3.8-max.toml:13 - Check: Relay reasoning controls must copy the lab entry's controls for the same surface (AGENTS.md → Reasoning controls: "do not leave controls out just because you could not re-test this host"; README: "Don't drop controls from uncertainty on relays"). Why: providers/alibaba/models/qwen3.8-max.toml:25 authors toggle = true (enable_thinking) and budget = { min = 0, max = 262_144 } (thinking_budget) for this exact model, and same-surface chat-completions relays (aihubmix, opencode) author them too. This file keeps only effort = ["low", "medium", "xhigh"], and its header documents probes of reasoning_effort only — no on/off or budget field was tested — so it asserts no caller off-switch/budget on a host that already forwards reasoning_effort and returns reasoning_content, i.e. the same DashScope-style surface where those fields exist. Action: Probe enable_thinking and thinking_budget (and thinking.type, which a peer relay uses) on this host; add toggle = true (and budget if a real reasoning-budget field is forwarded) with the required leading top-of-file wire comment, or document in the header that the host rejects them. While probing, also resolve explicit cache_control breakpoints, which providers/alibaba exposes on the same chat-completions surface and this file leaves out as untested.
  • [medium] [violation] providers/aigrow/models/claude-haiku-5.5.toml:14 - Check: Same relay rule — author the lab's toggle when the host exposes the same control. Why: providers/anthropic/models/claude-haiku-5-5.toml:18 authors toggle = true (thinking.type = disabled, accepted at high effort or below) alongside the identical effort list, and Messages-surface peers (opencode, aihubmix) author toggle = true for this model. The header claims "levels as on providers/anthropic" but drops the toggle without probing or documenting thinking.type on either listed API, publishing no off-switch where the lab and peers have one. Action: Probe thinking.type = "disabled" on Messages (and the off switch on Chat Completions, if any); add toggle = true with a leading top-of-file wire comment, or state in the header that the host rejects it.

This branch was previously deployed

1 inactive deployment
release — 1e350ba7 Deployed Nov 16, 2026 by guglielmo-io via review #19248
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant