kei-registries/models.toml

# Реестр моделей. Слой 2 из трёх.
#
# Каждая модель привязана к провайдеру через `provider_ref`. Цены в
# microcents (1e-6 USD) на 1M токенов — совместимо с
# kei-ledger.cost_micro_cents и kei-model-router::pricing.rs.
#
# Источники цен: pricing-страница соответствующего провайдера.
# Каждая запись помечена `verified_at` — дата последней сверки.
#
# `deprecated_at = null` ⇒ модель живая. После deprecate kei-model-router
# не выбирает её для новых invocation-ов, но старые записи в ledger
# продолжают резолвиться.

# ─── Anthropic ─────────────────────────────────────────────────────────

[[model]]
provider_ref = "anthropic"
# TODO: pin to "claude-haiku-4-5-20251001" once router/pricing accept slug as
# the lookup key (currently hardcoded by id in pricing.rs/escalate.rs tests).
id = "claude-haiku-4-5"
slug = "haiku"
display_name = "Claude Haiku 4.5"
context_window = 200000
cost_input_per_mtok_micro = 100_000_000     # $1.00
cost_output_per_mtok_micro = 500_000_000    # $5.00
cache_write_5m_per_mtok_micro = 125_000_000
cache_write_1h_per_mtok_micro = 200_000_000
cache_read_per_mtok_micro = 10_000_000
verified_at = "2026-04-30"
deprecated_at = ""
notes = "Cheapest Claude. Single-edit, formatting, lookup."

[[model]]
provider_ref = "anthropic"
id = "claude-sonnet-4-6"
slug = "sonnet"
display_name = "Claude Sonnet 4.6"
context_window = 1000000
cost_input_per_mtok_micro = 300_000_000     # $3.00
cost_output_per_mtok_micro = 1_500_000_000  # $15.00
cache_write_5m_per_mtok_micro = 375_000_000
cache_write_1h_per_mtok_micro = 600_000_000
cache_read_per_mtok_micro = 30_000_000
verified_at = "2026-04-30"
deprecated_at = ""
notes = "RULE 0.20 default for code-implementer-* / researcher-*."

[[model]]
provider_ref = "anthropic"
id = "claude-opus-4-7"
slug = "opus"
display_name = "Claude Opus 4.7"
context_window = 1000000
cost_input_per_mtok_micro = 500_000_000     # $5.00
cost_output_per_mtok_micro = 2_500_000_000  # $25.00
cache_write_5m_per_mtok_micro = 625_000_000
cache_write_1h_per_mtok_micro = 1_000_000_000
cache_read_per_mtok_micro = 50_000_000
verified_at = "2026-04-30"
deprecated_at = ""
notes = "Architecture, novel reasoning, math derivation. Tokenizer overhead 1.35x vs Sonnet."

# ─── OpenAI ────────────────────────────────────────────────────────────

[[model]]
provider_ref = "openai"
id = "gpt-5"
slug = "gpt-5"
display_name = "GPT-5"
context_window = 400000
cost_input_per_mtok_micro = 200_000_000     # $2.00 (placeholder, verify on use)
cost_output_per_mtok_micro = 800_000_000    # $8.00
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = ""
deprecated_at = ""
notes = "[UNVERIFIED] — re-fetch pricing page before billing-grade decisions."

[[model]]
provider_ref = "codex"
id = "gpt-5-codex"
slug = "codex"
display_name = "GPT-5 Codex (via ChatGPT OAuth)"
context_window = 200000
cost_input_per_mtok_micro = 0               # subscription, not per-token
cost_output_per_mtok_micro = 0
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = "2026-05-10"
deprecated_at = ""
notes = "RULE 0.23 dual-review. ChatGPT Plus/Pro/Team subscription quota."

# ─── xAI ───────────────────────────────────────────────────────────────

[[model]]
provider_ref = "xai"
id = "grok-4"
slug = "grok-4"
display_name = "Grok 4"
context_window = 256000
cost_input_per_mtok_micro = 300_000_000     # [UNVERIFIED]
cost_output_per_mtok_micro = 1_500_000_000
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = ""
deprecated_at = ""
notes = "[UNVERIFIED] — verify before billing."

# ─── DeepSeek ──────────────────────────────────────────────────────────

[[model]]
provider_ref = "deepseek"
id = "deepseek-r1"
slug = "r1"
display_name = "DeepSeek R1"
context_window = 64000
cost_input_per_mtok_micro = 55_000_000      # $0.55 [UNVERIFIED — pricing fluctuates]
cost_output_per_mtok_micro = 219_000_000    # $2.19
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 14_000_000      # $0.14 cache hit
verified_at = ""
deprecated_at = ""
notes = "Cheapest reasoning tier. Long thinking time."

[[model]]
provider_ref = "deepseek"
id = "deepseek-v3"
slug = "v3"
display_name = "DeepSeek V3"
context_window = 64000
cost_input_per_mtok_micro = 27_000_000      # $0.27 [UNVERIFIED]
cost_output_per_mtok_micro = 110_000_000    # $1.10
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 7_000_000
verified_at = ""
deprecated_at = ""
notes = "Non-reasoning baseline. Cheap bulk inference."

# ─── Google ────────────────────────────────────────────────────────────

[[model]]
provider_ref = "google"
id = "gemini-2.5-pro"
slug = "gemini-pro"
display_name = "Gemini 2.5 Pro"
context_window = 2000000
cost_input_per_mtok_micro = 125_000_000     # $1.25 [UNVERIFIED]
cost_output_per_mtok_micro = 1_000_000_000  # $10.00
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = ""
deprecated_at = ""
notes = "2M context. Long-document analysis."

# ─── Local (zero per-token cost) ───────────────────────────────────────

[[model]]
provider_ref = "ollama-local"
id = "llama-3.3-70b"
slug = "llama-70b-local"
display_name = "Llama 3.3 70B (Ollama local)"
context_window = 128000
cost_input_per_mtok_micro = 0
cost_output_per_mtok_micro = 0
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = "2026-05-13"
deprecated_at = ""
notes = "Local. Compute cost is electricity + opportunity, not per-token."

[[model]]
provider_ref = "mlx-local"
id = "qwen-2.5-coder-32b"
slug = "qwen-coder-local"
display_name = "Qwen 2.5 Coder 32B (MLX)"
context_window = 32000
cost_input_per_mtok_micro = 0
cost_output_per_mtok_micro = 0
cache_write_5m_per_mtok_micro = 0
cache_write_1h_per_mtok_micro = 0
cache_read_per_mtok_micro = 0
verified_at = "2026-05-13"
deprecated_at = ""
notes = "Code-focused local model. Apple silicon Metal."