Adds 14 Abacus RouteLLM provider entries that were present in the live https://routellm.abacus.ai/v1/models endpoint but missing from the repo.
All entries use existing lab metadata via base_model and override only provider-specific cost, context/output limits, and modalities per Abacus API values.
Validation: bun validate passes.
Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>
- Set reasoning = false for GLM-4.5, GLM-4.7, GLM-5, GLM-5.1,
MiniMax-M3, gemma-4-31b-it, Qwen3.6-27B, and
gemini-3.1-flash-image-preview because RouteLLM /v1/models does not
report thinking: true for those IDs.
- Add leading RouteLLM API citation blocks to the ~19 new/updated model
files that lacked them, mapping input/output modalities, limits,
token rates, and the thinking field where present.
- Restore reasoning = false for GLM-4.6, GLM-5.2, and MiniMax-M2.7
because the RouteLLM /v1/models descriptions for those IDs do not
indicate reasoning emission.
- Replace bare source URLs with explicit claim mappings: each cited
TOML now lists the RouteLLM API field values (context_length,
max_completion_tokens, input_modalities, output_modalities, token
rates, cached_input_token_rate, thinking) that justify the override.
Add a leading source comment to provider TOMLs whose cost, limit,
modalities, or reasoning flags diverge from the upstream base-model
metadata. All cited values are confirmed against the RouteLLM model
list at https://routellm.abacus.ai/v1/models.
Revert reasoning_options to [] for openai/gpt-oss-120b. The Abacus
provider.toml audit explicitly states the RouteLLM API documents no
reasoning toggle, effort, or numeric budget request field.
- Delete lowercase qwen3-coder-480b-a35b-instruct.toml so it no longer
shadows the uppercase base_model variant (Bun resolves TOML imports
case-insensitively).
- Restore reasoning_options effort control for openai/gpt-oss-120b;
OpenAI documents low/medium/high reasoning effort for this model and
RouteLLM's OpenAI-compatible API exposes it.
- Add base_model_omit = ["limit.input"] to gpt-5.4 so the inherited
922K input cap does not exceed the 400K context window.
- Replace reasoning = false with reasoning_options = [] on base-model
overrides that inherit reasoning = true from model metadata.
- Remove duplicate lowercase GLM TOMLs that shadowed the uppercase
variants due to Bun's case-insensitive TOML import resolution.
Søren Juul