[sorenjuul]

Summary

Adds 14 missing text-generation models to the Abacus provider, sourced from the live RouteLLM models endpoint. All entries point to existing lab metadata via base_model and override only provider-specific cost, context/output limits, and modalities.

Models added

Abacus ID	Base model
claude-opus-5	anthropic/claude-opus-5
gemini-3-pro-image	google/gemini-3-pro-image
gemini-3.1-flash-image	google/gemini-3.1-flash-image
gemini-3.5-flash-lite	google/gemini-3.5-flash-lite
gemini-3.6-flash	google/gemini-3.6-flash
gemini-3.7-flash	google/gemini-3.7-flash
gpt-realtime-whisper	openai/gpt-realtime-whisper
grok-4.6	xai/grok-4.6
moonshotai/Kimi-K2.7-Code	moonshotai/kimi-k2.7-code
moonshotai/Kimi-K3	moonshotai/kimi-k3
muse-spark-1.2	meta/muse-spark-1.2
qwen3.7-max	alibaba/qwen3.7-max
qwen3.8-max	alibaba/qwen3.8-max
thinkingmachines/Inkling	thinkingmachines/inkling

Reasoning options

Abacus exposes no reasoning toggle, effort, or budget field for these IDs (confirmed in provider metadata and consistent with existing Abacus entries), so reasoning_options = [] is set where the underlying lab model supports reasoning.

Validation

• bun validate passes
• git diff --check passes

Source

• https://routellm.abacus.ai/v1/models (accessed 2026-08-14)

[github-actions]

Action items

• [high] [possible mistake] providers/abacus/models/gpt-realtime-whisper.toml:238 - Check: Provider overrides must reflect the host’s real surface for the underlying lab model, and Abacus additions in this PR are scoped as text-generation catalog entries. Why: base_model = "openai/gpt-realtime-whisper" is lab-defined as streaming speech-to-text (input = ["audio"], output = ["text"], limits 0/0), but this entry forces input/output = ["text"], context = 16_000, output = 2_000, and cost.input = 283.3333 with output = 0. That combination looks like a mis-mapped STT/realtime SKU (or audio pricing reported as text tokens), not a normal chat text model, and will mislead API consumers. Action: Re-check the RouteLLM row for this ID against Abacus request/response behavior. If it is not a usable chat/text-generation endpoint, remove the file. If it is STT/realtime audio, set modalities (and cost keys such as input_audio if applicable) to match the real wire surface; if it truly is text-only on Abacus, document that delta in the leading comment and justify the ~$283/MTok text input rate from the provider source.

[sorenjuul]

Removed "gpt-realtime-whisper" from this PR. The RouteLLM row exposes it with text modalities and an anomalous input rate, but the underlying lab model is streaming STT/audio, so it does not belong in a text-generation catalog addition.

[github-actions]

Action items

• [medium] [violation] providers/abacus/models/claude-opus-5.toml:10 - Check: After base_model, keep only provider-specific fields and real deltas (no identical restated data). Why: Several new entries restate values already correct on the lab base, which breaks override-only inheritance and will drift when lab metadata changes. Action: Drop redundant fields and keep only real Abacus deltas:
  • All 13 files: remove restated reasoning = true (inherit from base; keep reasoning_options = [] + cost).
  • Remove fully identical [limit]: claude-opus-5, gemini-3-pro-image, gemini-3.5-flash-lite, gemini-3.6-flash, gemini-3.7-flash, moonshotai/Kimi-K2.7-Code, moonshotai/Kimi-K3, muse-spark-1.2, qwen3.8-max.
  • Remove fully identical [modalities]: gemini-3-pro-image, grok-4.6, moonshotai/Kimi-K2.7-Code, moonshotai/Kimi-K3, qwen3.7-max.
  • Where only one limit differs, override just that key (deep-merge), e.g. grok-4.6 → only limit.output = 32_768; qwen3.7-max → only limit.output = 64_000.
  • Keep modality overrides that actually differ (e.g. dropped pdf/video/audio vs lab).

[sorenjuul]

Cleaned up all 13 provider files to be strictly override-only: removed restated "reasoning = true", dropped identical [limit] and [modalities] tables, and reduced partial-difference limits to only the changed key (e.g. grok-4.6 output, qwen3.7-max output).

[github-actions]

No actionable findings.

[sorenjuul]

@rekram1-node new models added so they are available for users of abacus