harn-vm 0.8.22

# providers.toml — Harn's built-in LLM provider/model catalog defaults.
#
# This file is the single source of truth for Harn's bundled defaults:
# providers, model aliases, inference + tier routing rules, canonical
# model metadata + pricing, qc defaults, and per-pattern hyperparameter
# overrides. It deserializes into `ProvidersConfig` via the same Serde
# pipeline that loads HARN_PROVIDERS_CONFIG / ~/.config/harn/providers.toml
# / harn.toml [providers] / package-manifest [llm] sections at runtime.
#
# Resolution order at startup (later overlays win on per-key basis):
#   1. This file (embedded into the VM via include_str!)
#   2. ~/.config/harn/providers.toml (user-global override)
#   3. HARN_PROVIDERS_CONFIG env var (explicit per-process override)
#   4. Per-run programmatic overlays installed by hosts via
#      llm_config::set_user_overrides()
#
# Edit this file directly to change defaults. Do not re-add the equivalent
# data as Rust literals in llm_config.rs — that creates the parallel
# system this file exists to eliminate.

default_provider = "anthropic"

# ── Providers ────────────────────────────────────────────────────────────────
# Each [providers.X] block defines an LLM endpoint Harn can dial. The
# `auth_env` field can be a single string or an array (tried in order).
# `cost_per_1k_in/out` are coarse provider-level fallbacks used when a
# specific [models.X] entry has no `pricing` table.

[providers.anthropic]
base_url = "https://api.anthropic.com/v1"
auth_style = "header"
auth_header = "x-api-key"
auth_env = "ANTHROPIC_API_KEY"
chat_endpoint = "/messages"
features = ["prompt_caching", "thinking"]
cost_per_1k_in = 0.003
cost_per_1k_out = 0.015
latency_p50_ms = 2500
extra_headers = { "anthropic-version" = "2023-06-01" }

[providers.anthropic.healthcheck]
method = "POST"
path = "/messages/count_tokens"
body = '{"model":"claude-sonnet-4-6","messages":[{"role":"user","content":"x"}]}'

[providers.openai]
base_url = "https://api.openai.com/v1"
auth_style = "bearer"
auth_env = "OPENAI_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0025
cost_per_1k_out = 0.010
latency_p50_ms = 1800

[providers.openai.healthcheck]
method = "GET"
path = "/models"

[providers.openrouter]
base_url = "https://openrouter.ai/api/v1"
auth_style = "bearer"
auth_env = "OPENROUTER_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.003
cost_per_1k_out = 0.015
latency_p50_ms = 2200

[providers.openrouter.healthcheck]
method = "GET"
path = "/auth/key"

[providers.huggingface]
base_url = "https://router.huggingface.co/v1"
auth_style = "bearer"
auth_env = ["HF_TOKEN", "HUGGINGFACE_API_KEY"]
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0002
cost_per_1k_out = 0.0006
latency_p50_ms = 2400

[providers.huggingface.healthcheck]
method = "GET"
url = "https://huggingface.co/api/whoami-v2"

# Ollama defaults to /api/chat (native NDJSON) so the test stubs keep
# working; hosts can flip to /v1/chat/completions via a providers.toml
# overlay to bypass Ollama's per-model tool-call post-processors
# (qwen3coder.go, qwen35.go) that raise HTTP 500s on text-mode responses
# for the Qwen3.5 family.
[providers.ollama]
base_url = "http://localhost:11434"
base_url_env = "OLLAMA_HOST"
auth_style = "none"
chat_endpoint = "/api/chat"
completion_endpoint = "/api/generate"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 1200

[providers.ollama.healthcheck]
method = "GET"
path = "/api/tags"

[providers.gemini]
base_url = "https://generativelanguage.googleapis.com"
base_url_env = "GEMINI_BASE_URL"
auth_style = "header"
auth_header = "x-goog-api-key"
auth_env = ["GEMINI_API_KEY", "GOOGLE_API_KEY"]
chat_endpoint = "/v1beta/models"
cost_per_1k_in = 0.00125
cost_per_1k_out = 0.005
latency_p50_ms = 1800

[providers.gemini.healthcheck]
method = "GET"
path = "/v1beta/models"

[providers.together]
base_url = "https://api.together.xyz/v1"
base_url_env = "TOGETHER_AI_BASE_URL"
auth_style = "bearer"
auth_env = "TOGETHER_AI_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0002
cost_per_1k_out = 0.0006
latency_p50_ms = 1600

[providers.together.healthcheck]
method = "GET"
path = "/models"

# Groq — OpenAI-compatible LPU-hosted fast inference. Headline ~840 tok/s
# on Llama 3.1 8B, ~594 tok/s on Llama 4 Scout. Useful executor target
# when sub-100ms TTFT matters more than raw quality.
[providers.groq]
base_url = "https://api.groq.com/openai/v1"
base_url_env = "GROQ_BASE_URL"
auth_style = "bearer"
auth_env = "GROQ_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0001
cost_per_1k_out = 0.0003
latency_p50_ms = 450

[providers.groq.healthcheck]
method = "GET"
path = "/models"

# Cerebras — OpenAI-compatible wafer-scale inference. Headline ~3,000 tok/s
# on gpt-oss-120b and ~2,100 tok/s on Llama 3.3 70B, fast enough to keep
# binder-style single-tool-call hops under ~50ms p50. Provider-level
# pricing here is a coarse default; per-model rows under [models.X] hold
# the authoritative numbers from the Cerebras pricing page.
[providers.cerebras]
base_url = "https://api.cerebras.ai/v1"
base_url_env = "CEREBRAS_BASE_URL"
auth_style = "bearer"
auth_env = "CEREBRAS_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.00025
cost_per_1k_out = 0.00069
latency_p50_ms = 80
features = ["native_tools"]

[providers.cerebras.healthcheck]
method = "GET"
path = "/models"

[providers.deepseek]
base_url = "https://api.deepseek.com/v1"
base_url_env = "DEEPSEEK_BASE_URL"
auth_style = "bearer"
auth_env = "DEEPSEEK_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.00014
cost_per_1k_out = 0.00028
latency_p50_ms = 1800

[providers.deepseek.healthcheck]
method = "GET"
path = "/models"

[providers.fireworks]
base_url = "https://api.fireworks.ai/inference/v1"
base_url_env = "FIREWORKS_BASE_URL"
auth_style = "bearer"
auth_env = "FIREWORKS_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0002
cost_per_1k_out = 0.0006
latency_p50_ms = 1400

[providers.fireworks.healthcheck]
method = "GET"
path = "/models"

[providers.dashscope]
base_url = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
base_url_env = "DASHSCOPE_BASE_URL"
auth_style = "bearer"
auth_env = "DASHSCOPE_API_KEY"
chat_endpoint = "/chat/completions"
completion_endpoint = "/completions"
cost_per_1k_in = 0.0003
cost_per_1k_out = 0.0012
latency_p50_ms = 1600

[providers.dashscope.healthcheck]
method = "GET"
path = "/models"

# AWS Bedrock — resolves credentials through env, profile, container, or
# EC2 instance roles, then signs Converse API calls with SigV4.
[providers.bedrock]
base_url = ""
base_url_env = "BEDROCK_BASE_URL"
auth_style = "aws_sigv4"
chat_endpoint = "/model/{model}/converse"
features = ["native_tools"]
latency_p50_ms = 2600

# Azure OpenAI — deployment name is routed in the URL; callers can
# either pass the deployment as the Harn model field or set
# AZURE_OPENAI_DEPLOYMENT.
[providers.azure_openai]
base_url = "https://{resource}.openai.azure.com"
base_url_env = "AZURE_OPENAI_ENDPOINT"
auth_style = "azure_openai"
auth_env = ["AZURE_OPENAI_API_KEY", "AZURE_OPENAI_AD_TOKEN", "AZURE_OPENAI_BEARER_TOKEN"]
chat_endpoint = "/openai/deployments/{deployment}/chat/completions?api-version={api_version}"
features = ["native_tools"]
cost_per_1k_in = 0.0025
cost_per_1k_out = 0.010
latency_p50_ms = 1900

[providers.vertex]
base_url = "https://aiplatform.googleapis.com/v1"
base_url_env = "VERTEX_AI_BASE_URL"
auth_style = "bearer"
auth_env = ["VERTEX_AI_ACCESS_TOKEN", "GOOGLE_OAUTH_ACCESS_TOKEN", "GOOGLE_APPLICATION_CREDENTIALS"]
chat_endpoint = "/projects/{project}/locations/{location}/publishers/google/models/{model}:generateContent"
features = ["native_tools"]
cost_per_1k_in = 0.00125
cost_per_1k_out = 0.005
latency_p50_ms = 2100

[providers.local]
base_url = "http://localhost:8000"
base_url_env = "LOCAL_LLM_BASE_URL"
auth_style = "none"
chat_endpoint = "/v1/chat/completions"
completion_endpoint = "/v1/completions"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 900

[providers.local.healthcheck]
method = "GET"
path = "/v1/models"

# llama.cpp — separate from `local` so capability rules can isolate Qwen
# chat-template thinking quirks from other local OpenAI-compatible hosts.
[providers.llamacpp]
base_url = "http://127.0.0.1:8001"
base_url_env = "LLAMACPP_BASE_URL"
auth_style = "none"
chat_endpoint = "/v1/chat/completions"
completion_endpoint = "/v1/completions"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 900

[providers.llamacpp.healthcheck]
method = "GET"
path = "/v1/models"

# Apple Silicon MLX. Harn owns readiness probing; hosts that want
# script-based auto-start should launch the process first, then call
# Harn again to verify readiness.
[providers.mlx]
base_url = "http://127.0.0.1:8002"
base_url_env = "MLX_BASE_URL"
auth_style = "none"
chat_endpoint = "/v1/chat/completions"
completion_endpoint = "/v1/completions"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 900

[providers.mlx.healthcheck]
method = "GET"
path = "/v1/models"

[providers.vllm]
base_url = "http://localhost:8000"
base_url_env = "VLLM_BASE_URL"
auth_style = "none"
chat_endpoint = "/v1/chat/completions"
completion_endpoint = "/v1/completions"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 800

[providers.vllm.healthcheck]
method = "GET"
path = "/v1/models"

[providers.tgi]
base_url = "http://localhost:8080"
base_url_env = "TGI_BASE_URL"
auth_style = "none"
chat_endpoint = "/v1/chat/completions"
completion_endpoint = "/v1/completions"
cost_per_1k_in = 0.0
cost_per_1k_out = 0.0
latency_p50_ms = 950

[providers.tgi.healthcheck]
method = "GET"
path = "/health"

# ── Inference rules ──────────────────────────────────────────────────────────
# Map a model ID shape to a default provider when the caller doesn't
# specify one. First match wins. User overlays prepend, so they can
# preempt these defaults without removing them.

[[inference_rules]]
pattern = "claude-*"
provider = "anthropic"

[[inference_rules]]
pattern = "gpt-*"
provider = "openai"

[[inference_rules]]
pattern = "o1*"
provider = "openai"

[[inference_rules]]
pattern = "o3*"
provider = "openai"

[[inference_rules]]
pattern = "o4*"
provider = "openai"

[[inference_rules]]
pattern = "anthropic.claude-*"
provider = "bedrock"

[[inference_rules]]
pattern = "meta.llama*"
provider = "bedrock"

[[inference_rules]]
pattern = "amazon.*"
provider = "bedrock"

[[inference_rules]]
pattern = "mistral.*"
provider = "bedrock"

[[inference_rules]]
pattern = "cohere.*"
provider = "bedrock"

[[inference_rules]]
pattern = "gemini-*"
provider = "gemini"

# Cerebras model IDs come back as bare names ("gpt-oss-120b",
# "llama-3.3-70b") from /v1/models, so callers slash-prefix
# them as "cerebras/<model>" to disambiguate from OpenRouter's
# one-slash convention. Match the prefix before the generic
# single-slash rule routes it elsewhere.
[[inference_rules]]
pattern = "cerebras/*"
provider = "cerebras"

# ── Tier rules ───────────────────────────────────────────────────────────────
# Classify a model into small / mid / frontier for default routing and
# evaluation budgets. First match wins.

[[tier_rules]]
contains = "9b"
tier = "small"

[[tier_rules]]
contains = "a3b"
tier = "small"

[[tier_rules]]
contains = "gemma-4-e2b"
tier = "small"

[[tier_rules]]
contains = "gemma-4-e4b"
tier = "small"

[[tier_rules]]
contains = "gemma-4-26b"
tier = "mid"

[[tier_rules]]
contains = "gemma-4-31b"
tier = "frontier"

[[tier_rules]]
contains = "gemma4:26b"
tier = "mid"

[[tier_rules]]
contains = "gemma4:31b"
tier = "frontier"

[[tier_rules]]
pattern = "claude-*"
tier = "frontier"

[[tier_rules]]
exact = "gpt-4o"
tier = "frontier"

[tier_defaults]
default = "mid"

# ── Aliases ──────────────────────────────────────────────────────────────────
# Short symbolic names → (model id, provider, optional tool_format). The
# tier-resolution path (`resolve_tier_model("frontier", None)`) reads
# `frontier`, `mid`, `small`; provider-scoped tiers like `tier/mid` let
# callers force a specific resolution per provider.

# Short flagship aliases — these track whatever the current
# generation is. Bump these when a successor lands.
[aliases.sonnet]
id = "claude-sonnet-4-6"
provider = "anthropic"

[aliases.opus]
id = "claude-opus-4-7"
provider = "anthropic"

[aliases.haiku]
id = "claude-haiku-4-5-20251001"
provider = "anthropic"

[aliases.frontier]
id = "claude-sonnet-4-6"
provider = "anthropic"

[aliases."tier/frontier"]
id = "claude-sonnet-4-6"
provider = "anthropic"

[aliases.mid]
id = "gpt-4o-mini"
provider = "openai"

[aliases."tier/mid"]
id = "gpt-4o-mini"
provider = "openai"

[aliases.small]
id = "Qwen/Qwen3.5-9B"
provider = "openrouter"

[aliases."tier/small"]
id = "Qwen/Qwen3.5-9B"
provider = "openrouter"

# Local Gemma 4 variants (vLLM / OpenAI-compat backend at `providers.local`).
[aliases.local-gemma4]
id = "gemma-4-26b-a4b-it"
provider = "local"

[aliases.local-gemma4-26b]
id = "gemma-4-26b-a4b-it"
provider = "local"

[aliases.local-gemma4-31b]
id = "gemma-4-31b-it"
provider = "local"

[aliases.local-gemma4-e4b]
id = "gemma-4-e4b-it"
provider = "local"

[aliases.local-gemma4-e2b]
id = "gemma-4-e2b-it"
provider = "local"

[aliases.ollama-gemma4]
id = "gemma4:26b"
provider = "ollama"
tool_format = "text"

[aliases.ollama-gemma4-26b]
id = "gemma4:26b"
provider = "ollama"
tool_format = "text"

# Qwen3.6 — Ollama (text tool calling is the safe default; the `-native`
# variant opts into the experimental native path).
[aliases."qwen3.6-coding"]
id = "qwen3.6:35b-a3b-coding-nvfp4"
provider = "ollama"
tool_format = "text"

[aliases."qwen3.6-35b-coding"]
id = "qwen3.6:35b-a3b-coding-nvfp4"
provider = "ollama"
tool_format = "text"

[aliases."qwen3.6-coding-nvfp4"]
id = "qwen3.6:35b-a3b-coding-nvfp4"
provider = "ollama"
tool_format = "text"

[aliases."qwen3.6-coding-native"]
id = "qwen3.6:35b-a3b-coding-nvfp4"
provider = "ollama"
tool_format = "native"

# llama.cpp — Unsloth Dynamic 2.0 GGUF served by llama-server.
[aliases."llamacpp-qwen3.6"]
id = "qwen3.6-35b-a3b"
provider = "llamacpp"
tool_format = "text"

[aliases."llamacpp-qwen3.6-q4"]
id = "qwen3.6-35b-a3b-ud-q4-k-xl"
provider = "llamacpp"
tool_format = "text"

[aliases."local-qwen3.6"]
id = "qwen3.6-35b-a3b-ud-q4-k-xl"
provider = "llamacpp"
tool_format = "text"

[aliases."local-qwen3.6-gguf"]
id = "qwen3.6-35b-a3b-ud-q4-k-xl"
provider = "llamacpp"
tool_format = "text"

# MLX (Apple Silicon).
[aliases.mlx-qwen36-27b]
id = "unsloth/Qwen3.6-27B-UD-MLX-4bit"
provider = "mlx"

[aliases."mlx-qwen3.6-27b"]
id = "unsloth/Qwen3.6-27B-UD-MLX-4bit"
provider = "mlx"
tool_format = "native"

[aliases."mlx-qwen3.6-27b-q4"]
id = "unsloth/Qwen3.6-27B-UD-MLX-4bit"
provider = "mlx"
tool_format = "native"

[aliases."local-qwen3.6-27b"]
id = "unsloth/Qwen3.6-27B-UD-MLX-4bit"
provider = "mlx"
tool_format = "native"

# Devstral (Mistral's agentic-coding tune).
[aliases.devstral-small-2]
id = "devstral-small-2:24b"
provider = "ollama"
tool_format = "text"

[aliases.ollama-devstral-small-2]
id = "devstral-small-2:24b"
provider = "ollama"
tool_format = "text"

[aliases.ollama-devstral-small-2-native]
id = "devstral-small-2:24b"
provider = "ollama"
tool_format = "native"

# ── Alias tool-calling probe state ───────────────────────────────────────────
# Per-alias overrides recording the last-observed native vs. text vs.
# streaming tool-call probe outcome and the desired fallback. Hosts may
# update these via providers.toml overlays as they re-probe a model.

[alias_tool_calling."qwen3.6-coding"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "text"

[alias_tool_calling."qwen3.6-coding-native"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "native"

[alias_tool_calling.ollama-gemma4]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "disabled"
failure_reason = "requires_tool_probe"

[alias_tool_calling."llamacpp-qwen3.6-q4"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "text"
failure_reason = "requires_tool_probe_and_cache_probe"

[alias_tool_calling."mlx-qwen3.6-27b"]
native = "unknown"
text = "unknown"
streaming_native = "unknown"
fallback_mode = "native"
failure_reason = "requires_served_identity_and_tool_probe"

# ── QC defaults ──────────────────────────────────────────────────────────────
# Default low-cost model per provider for cheap quality-check / repair
# passes. Scripts read these via `qc_default_model(provider)`.

[qc_defaults]
anthropic = "claude-haiku-4-5-20251001"
openai = "gpt-4o-mini"
openrouter = "google/gemini-2.5-flash"
ollama = "llama3.2"
local = "gpt-4o"

# ── Models ───────────────────────────────────────────────────────────────────
# Canonical model metadata: display name, provider, context window,
# capabilities, pricing (USD per 1M tokens), and deprecation status.
# Pricing reflects public provider pages snapshotted at the comment
# beside each section; edit the literal here and the change shows up in
# `git blame`.

# Anthropic ─ pricing pages: https://www.anthropic.com/pricing &
# https://platform.claude.com/docs/en/about-claude/model-deprecations.
# Sonnet 4.5 retired 2026-05-15; Sonnet 4 and Opus 4 retire 2026-06-15.

[models."claude-3-5-haiku-20241022"]
name = "Claude Haiku 3.5"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 0.80, output_per_mtok = 4.00, cache_read_per_mtok = 0.08, cache_write_per_mtok = 1.00 }

[models."claude-haiku-4-5-20251001"]
name = "Claude Haiku 4.5"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 1.00, output_per_mtok = 5.00, cache_read_per_mtok = 0.10, cache_write_per_mtok = 1.25 }

[models."claude-3-5-sonnet-20240620"]
name = "Claude Sonnet 3.5 (2024-06-20)"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }

[models."claude-3-5-sonnet-20241022"]
name = "Claude Sonnet 3.5 (2024-10-22)"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }

[models."claude-sonnet-4-20250514"]
name = "Claude Sonnet 4"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }
deprecated = true
deprecation_note = "Sunset 2026-06-15 per Anthropic deprecations page. Replaced by claude-sonnet-4-6."

[models."claude-sonnet-4-5"]
name = "Claude Sonnet 4.5"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }
deprecated = true
deprecation_note = "Sunset 2026-05-15 per Anthropic deprecations page. Replaced by claude-sonnet-4-6."

[models."claude-sonnet-4-6"]
name = "Claude Sonnet 4.6"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }

[models."claude-sonnet-4-7"]
name = "Claude Sonnet 4.7"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 15.00, cache_read_per_mtok = 0.30, cache_write_per_mtok = 3.75 }

[models."claude-3-opus-20240229"]
name = "Claude Opus 3"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 75.00, cache_read_per_mtok = 1.50, cache_write_per_mtok = 18.75 }

[models."claude-opus-4-20250514"]
name = "Claude Opus 4"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 75.00, cache_read_per_mtok = 1.50, cache_write_per_mtok = 18.75 }
deprecated = true
deprecation_note = "Sunset 2026-06-15 per Anthropic deprecations page. Replaced by claude-opus-4-7."

[models."claude-opus-4-1-20250805"]
name = "Claude Opus 4.1"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 75.00, cache_read_per_mtok = 1.50, cache_write_per_mtok = 18.75 }
deprecated = true
deprecation_note = "Superseded by claude-opus-4-7. No formal sunset yet; switch when convenient."

[models."claude-opus-4-6"]
name = "Claude Opus 4.6"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 75.00, cache_read_per_mtok = 1.50, cache_write_per_mtok = 18.75 }

[models."claude-opus-4-7"]
name = "Claude Opus 4.7"
provider = "anthropic"
context_window = 200000
capabilities = ["tools", "streaming", "prompt_caching", "thinking"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 75.00, cache_read_per_mtok = 1.50, cache_write_per_mtok = 18.75 }

# OpenAI ─ pricing pages: https://platform.openai.com/docs/pricing.
# GPT-4o retired from ChatGPT 2026-02-13; chatgpt-4o-latest removed
# from API 2026-02-17 (Enterprise/Edu grace until 2026-04-03).

[models."gpt-4o"]
name = "GPT-4o"
provider = "openai"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 2.50, output_per_mtok = 10.00, cache_read_per_mtok = 1.25 }
deprecated = true
deprecation_note = "API sunset 2026-02-17 per OpenAI deprecations page. Switch to gpt-5-mini for cheap routing or gpt-5 for frontier."

[models."gpt-4o-mini"]
name = "GPT-4o Mini"
provider = "openai"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.15, output_per_mtok = 0.60 }
# Not yet deprecated as of 2026-05 — OpenAI's deprecation page lists
# gpt-4o (Feb 17 2026 API sunset) but gpt-4o-mini has no announced
# sunset. Still the canonical `mid` tier default until gpt-5-mini ships
# with confirmed pricing.

[models."gpt-4-turbo"]
name = "GPT-4 Turbo"
provider = "openai"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 10.00, output_per_mtok = 30.00 }
deprecated = true
deprecation_note = "Superseded by gpt-5 family. Listed for cost-attribution backfill only."

[models.o1]
name = "OpenAI o1"
provider = "openai"
context_window = 200000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 60.00, cache_read_per_mtok = 7.50 }

[models."o1-mini"]
name = "OpenAI o1-mini"
provider = "openai"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 3.00, output_per_mtok = 12.00, cache_read_per_mtok = 1.50 }

[models.o3]
name = "OpenAI o3"
provider = "openai"
context_window = 200000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 15.00, output_per_mtok = 60.00, cache_read_per_mtok = 7.50 }

[models."o3-mini"]
name = "OpenAI o3-mini"
provider = "openai"
context_window = 200000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 1.10, output_per_mtok = 4.40, cache_read_per_mtok = 0.55 }

# Google Gemini ─ pricing: https://ai.google.dev/pricing.
# Gemini 1.0 / 1.5 already retired. Gemini 2.0 Flash + Flash-Lite shut
# down 2026-06-01 per the deprecations page.

[models."gemini-2.5-flash"]
name = "Gemini 2.5 Flash"
provider = "gemini"
context_window = 1048576
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.10, output_per_mtok = 0.40, cache_read_per_mtok = 0.025 }

# OpenRouter-routed variant of the same model — kept as a distinct
# catalog entry so the `qc_defaults.openrouter` lookup resolves to a
# registered ID. Pricing matches the native Gemini API; OpenRouter adds
# its own margin at request time.
[models."google/gemini-2.5-flash"]
name = "Gemini 2.5 Flash (via OpenRouter)"
provider = "openrouter"
context_window = 1048576
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.10, output_per_mtok = 0.40, cache_read_per_mtok = 0.025 }

[models."gemini-2.5-pro"]
name = "Gemini 2.5 Pro"
provider = "gemini"
context_window = 2097152
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 1.25, output_per_mtok = 5.00, cache_read_per_mtok = 0.3125 }

# Mistral hosted via OpenRouter.

[models."mistral-large-latest"]
name = "Mistral Large"
provider = "openrouter"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 2.00, output_per_mtok = 6.00 }

[models."mistral-small-latest"]
name = "Mistral Small"
provider = "openrouter"
context_window = 128000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.20, output_per_mtok = 0.60 }

# Open-weight executor candidates (<$2/Mtok with function calling). Use
# these via OpenRouter or Fireworks for fast secondary-model dispatch.
# Pricing snapshot 2026-05 from OpenRouter / Artificial Analysis.

[models."qwen/qwen3-coder"]
name = "Qwen3 Coder 480B A35B"
provider = "openrouter"
context_window = 262144
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.22, output_per_mtok = 1.80 }

[models."deepseek/deepseek-v3.2"]
name = "DeepSeek V3.2"
provider = "openrouter"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.28, output_per_mtok = 0.42 }

[models."moonshotai/kimi-k2.6"]
name = "Kimi K2.6"
provider = "openrouter"
context_window = 200000
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.55, output_per_mtok = 2.20 }

[models."openai/gpt-oss-120b"]
name = "GPT-OSS 120B"
provider = "openrouter"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.15, output_per_mtok = 0.60 }

# Cerebras-hosted open-weight models. Pricing snapshot from
# https://www.cerebras.ai/pricing (2026-05). The headline binder-substrate
# candidate is gpt-oss-120b at ~3,000 tok/s; Llama 3.3 70B trails at
# ~2,100 tok/s but is included as a fallback when the binder hop wants
# function-calling-trained Llama instead of GPT-OSS.
#
# Catalog keys are bare wire IDs (Cerebras's /v1/chat/completions wants
# the raw model name). Users routing via `model: "cerebras/<name>"` get
# the slash-prefixed selector stripped by `normalize_model_id` while
# `infer_provider` routes them to this provider.

[models."gpt-oss-120b"]
name = "GPT-OSS 120B (Cerebras)"
provider = "cerebras"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.25, output_per_mtok = 0.69 }

[models."llama-3.3-70b"]
name = "Llama 3.3 70B (Cerebras)"
provider = "cerebras"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 0.85, output_per_mtok = 1.20 }

[models."qwen-3-coder-480b"]
name = "Qwen3 Coder 480B (Cerebras)"
provider = "cerebras"
context_window = 131072
capabilities = ["tools", "streaming"]
pricing = { input_per_mtok = 2.00, output_per_mtok = 2.00 }

# Open-router Qwen3.5 9B (kept for the `small` tier alias).

[models."Qwen/Qwen3.5-9B"]
name = "Qwen3.5 9B"
provider = "openrouter"
context_window = 131072
capabilities = ["tools", "streaming"]

# Ollama / local models — no `pricing` (free); context_window reflects
# model card ceiling. `runtime_context_window` caps what Harn will
# actually feed the runtime (host memory budget).

[models."llama3.2"]
name = "Llama 3.2"
provider = "ollama"
context_window = 32000
stream_timeout = 300.0
capabilities = ["tools", "streaming"]

[models."gemma4:26b"]
name = "Gemma 4 26B MoE"
provider = "ollama"
context_window = 262144
runtime_context_window = 32768
stream_timeout = 300.0
capabilities = ["tools", "vision", "streaming", "thinking"]

[models."qwen3.6:35b-a3b-coding-nvfp4"]
name = "Qwen3.6 35B A3B Coding (NVFP4)"
provider = "ollama"
context_window = 262144
runtime_context_window = 32768
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]

[models."devstral-small-2:24b"]
name = "Devstral Small 2 24B"
provider = "ollama"
context_window = 262144
runtime_context_window = 32768
stream_timeout = 600.0
capabilities = ["tools", "streaming"]

# llama.cpp — Unsloth Dynamic 2.0 GGUF served by llama-server.

[models."qwen3.6-35b-a3b-ud-q4-k-xl"]
name = "Qwen3.6 35B (Unsloth Q4_K_XL, llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]

[models."qwen3.6-35b-a3b-ud-q5-k-xl"]
name = "Qwen3.6 35B (Unsloth Q5_K_XL, llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]

[models."qwen3.6-35b-a3b"]
name = "Qwen3.6 35B (llama.cpp)"
provider = "llamacpp"
context_window = 262144
runtime_context_window = 65536
stream_timeout = 900.0
capabilities = ["tools", "streaming", "thinking"]

# Apple Silicon MLX.

[models."unsloth/Qwen3.6-27B-UD-MLX-4bit"]
name = "Qwen3.6 27B (MLX 4-bit)"
provider = "mlx"
context_window = 262144
stream_timeout = 900.0
capabilities = ["tools", "vision", "streaming", "thinking"]

# Local OpenAI-compatible servers (vLLM / bring-your-own).

[models."gemma-4-e2b-it"]
name = "Gemma 4 E2B (local)"
provider = "local"
context_window = 131072
stream_timeout = 300.0
capabilities = ["streaming", "thinking"]

[models."gemma-4-e4b-it"]
name = "Gemma 4 E4B (local)"
provider = "local"
context_window = 131072
stream_timeout = 300.0
capabilities = ["streaming", "thinking"]

[models."gemma-4-26b-a4b-it"]
name = "Gemma 4 26B MoE (local)"
provider = "local"
context_window = 131072
stream_timeout = 600.0
capabilities = ["streaming", "thinking"]

[models."gemma-4-31b-it"]
name = "Gemma 4 31B (local)"
provider = "local"
context_window = 131072
stream_timeout = 600.0
capabilities = ["streaming", "thinking"]