models.txt

# Roster documentation only — the EXECUTED roster is run_matrix_v41.py
# (EXISTING_*/NEW_* lists + FRONTIER_2026). See RUNBOOK.md "Add a model".
anthropic/claude-fable-5
anthropic/claude-opus-4-6
anthropic/claude-opus-4-7
anthropic/claude-opus-4-8
anthropic/claude-sonnet-4-6
anthropic/claude-haiku-4-5
google/gemini-2.5-flash
google/gemini-2.5-pro
google/gemini-3-flash-preview
google/gemini-3.1-pro-preview
google/gemini-3.5-flash
google/gemini-3.1-flash-lite-preview
openai/gpt-4.1
openai/gpt-5-codex
openai/gpt-5.2
openai/gpt-5.4
openai/gpt-5.5
openai/gpt-5.5-pro
openai/o3
openai/o4-mini
x-ai/grok-3
x-ai/grok-3-mini
x-ai/grok-4
x-ai/grok-4-fast
x-ai/grok-4.1-fast
x-ai/grok-4.20
x-ai/grok-code-fast-1
mistralai/codestral-2508
mistralai/devstral-2512
mistralai/devstral-medium
mistralai/devstral-small
deepseek/deepseek-r1
deepseek/deepseek-r1-0528
deepseek/deepseek-v3.2
deepseek/deepseek-v4-pro
qwen/qwen3-coder
qwen/qwen3-coder-flash
qwen/qwen3-coder-plus
qwen/qwen3.6-35b-a3b
moonshotai/kimi-k2.5
moonshotai/kimi-k2.6
meta-llama/llama-4-maverick
mlx/ternary-bonsai-8b

# ---------------------------------------------------------------------------
# PENDING — ids not live yet. When each id lands, uncomment its line here and
# apply the one-line edits listed below it (prices TBD until announced).
# ---------------------------------------------------------------------------
# openai/gpt-5.6
#   run_matrix_v41.py FRONTIER_2026:   "gpt-5.6": ["native", "kernel-cpu"],
#   consolidate_scores.py RATE_CARD:   add ONLY when the price is announced:
#                                      "gpt-5.6": (<$/M in>, <$/M out>),
#     DO NOT paste a placeholder — a literal `(TBD, TBD)` is a NameError at
#     import of consolidate_scores.py, which nothing exercised until
#     POSTFLIGHT (i.e. AFTER the paid run); preflight now import-checks the
#     consolidator and refuses to launch on it. Until the price lands, simply
#     OMIT the RATE_CARD entry: compute_cost_ex falls back to the harness's
#     self-reported cost, else 'unpriced' — never a silent $0.
#   NOTE: like gpt-5.5, do NOT add to CPU_DRIVER_MODELS — ostk has no native
#   OpenAI Responses driver, so the kernel-cpu execution is generic_kernel
#   (B*) via OpenRouter/api.openai.com and must be labeled as such.
# google/gemini-3.5-pro
#   run_matrix_v41.py FRONTIER_2026:   "gemini-3.5-pro": ["native", "kernel-cpu"],
#   consolidate_scores.py RATE_CARD:   add ONLY when the price is announced
#                                      (same rule as gpt-5.6 above — never
#                                      paste TBD; omit until priced):
#                                      "gemini-3.5-pro": (<$/M in>, <$/M out>),
#   consolidate_scores.py CPU_DRIVER_MODELS: "gemini-3-5-pro",   # normalized: dots -> dashes
#   CAUTION: gemini-3.5-flash was DEFERRED 2026-06-14 for a kernel-arm stall
#   in the gemini-3.x thinking path (run_matrix_v41.py:90-97); 3.5-pro sits in
#   the same is_gemini_3 family gate — verify with ONE cheap bench before a
#   paid matrix run. bench.sh has no --bench flag; the mandated probe is the
#   direct runner (one cell, one sample):
#     python3 run_matrix_v41.py --frontier --model gemini-3.5-pro \
#       --arm kernel-cpu --bench off-by-one-pagination --samples 1