omni-dev 0.39.0

# AI Model Specifications
# This file defines the capabilities and limits for all supported AI models
# Used by omni-dev to configure API requests correctly
# Last updated: July 2026

# Schema version. Bumped when the YAML structure changes in a backwards-
# incompatible way. User/project files declaring a different version receive
# a warning at load time. Files without a version are accepted with a warning
# for backwards compatibility.
version: "1"

models:
  # Claude 5 / Opus 4.8 / Opus 4.7 (Current Generation)
  #
  # These models expose a 1M context window natively, at standard pricing with
  # no long-context premium — so they set `input_context: 1000000` directly and
  # carry no `context-1m-2025-08-07` beta header. That header is the older
  # mechanism and is not how this generation exposes 1M context.
  - provider: "claude"
    model: "Claude Fable 5"
    api_identifier: "claude-fable-5"
    input_token_price: 10.0
    output_token_price: 50.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 5.0
    tier: "flagship"
    supports_structured_output: true

  - provider: "claude"
    model: "Claude Opus 4.8"
    api_identifier: "claude-opus-4-8"
    input_token_price: 5.0
    output_token_price: 25.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 4.8
    tier: "flagship"
    supports_structured_output: true

  - provider: "claude"
    model: "Claude Opus 4.7"
    api_identifier: "claude-opus-4-7"
    input_token_price: 5.0
    output_token_price: 25.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 4.7
    tier: "flagship"
    supports_structured_output: true

  # Sonnet 5 has introductory pricing of $2.00/$10.00 per MTok through
  # 2026-08-31. The schema has no way to express a time-bounded price, so the
  # sticker rate is recorded here; cost estimates during the intro period are
  # therefore over-stated rather than silently wrong once it lapses.
  - provider: "claude"
    model: "Claude Sonnet 5"
    api_identifier: "claude-sonnet-5"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 5.0
    tier: "balanced"
    supports_structured_output: true

  # Claude 4.6 Series
  #
  # 1M context is native on this generation too, so `input_context` is
  # 1000000 here rather than 200000-plus-a-beta-header. The
  # `context-1m-2025-08-07` block below is retained deliberately: it is now a
  # no-op for limits (it resolves to the same 1M the base already grants), but
  # `validate_beta_header` hard-errors on any header absent from this list, so
  # dropping it would newly reject `--beta-header
  # anthropic-beta:context-1m-2025-08-07` for callers who still pass it.
  - provider: "claude"
    model: "Claude Opus 4.6"
    api_identifier: "claude-opus-4-6"
    input_token_price: 5.0
    output_token_price: 25.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 4.6
    tier: "flagship"
    supports_structured_output: true
    beta_headers:
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  - provider: "claude"
    model: "Claude Sonnet 4.6"
    api_identifier: "claude-sonnet-4-6"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 128000
    input_context: 1000000
    generation: 4.6
    tier: "balanced"
    supports_structured_output: true
    beta_headers:
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  # Claude 4.5 Series (Active, superseded by the 5 / 4.8 / 4.7 generation)
  #
  # Anthropic publishes an undated alias alongside each dated identifier
  # (e.g. `claude-haiku-4-5` for `claude-haiku-4-5-20251001`) and documents the
  # alias as the preferred form. Identifier resolution is exact-match then a
  # narrow normalization (region prefix / `anthropic.` prefix / `-v<N>:<M>`
  # suffix) — none of which maps an alias onto its dated entry. So each alias
  # needs its own entry, or `--model claude-haiku-4-5` silently misses the
  # registry and falls back to the provider default of 4096 output tokens
  # instead of this model's real 64000 (#1334).
  - provider: "claude"
    model: "Claude Sonnet 4.5"
    api_identifier: "claude-sonnet-4-5-20250929"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "balanced"
    supports_structured_output: true
    beta_headers:
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  - provider: "claude"
    model: "Claude Sonnet 4.5"
    api_identifier: "claude-sonnet-4-5"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "balanced"
    supports_structured_output: true
    beta_headers:
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  - provider: "claude"
    model: "Claude Haiku 4.5"
    api_identifier: "claude-haiku-4-5-20251001"
    input_token_price: 1.0
    output_token_price: 5.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "fast"
    supports_structured_output: true

  - provider: "claude"
    model: "Claude Haiku 4.5"
    api_identifier: "claude-haiku-4-5"
    input_token_price: 1.0
    output_token_price: 5.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "fast"
    supports_structured_output: true

  - provider: "claude"
    model: "Claude Opus 4.5"
    api_identifier: "claude-opus-4-5-20251101"
    input_token_price: 5.0
    output_token_price: 25.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "flagship"
    legacy: true
    supports_structured_output: true

  - provider: "claude"
    model: "Claude Opus 4.5"
    api_identifier: "claude-opus-4-5"
    input_token_price: 5.0
    output_token_price: 25.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4.5
    tier: "flagship"
    legacy: true
    supports_structured_output: true

  # Claude 4.x Series (Deprecated)
  #
  # `retires` is set only where a date has actually been announced. Opus 4 and
  # Sonnet 4 below are deprecated with retirement still TBD, so they carry no
  # `retires` — absence means "no published date", not "not retiring".
  - provider: "claude"
    model: "Claude Opus 4.1"
    api_identifier: "claude-opus-4-1-20250805"
    input_token_price: 15.0
    output_token_price: 75.0
    max_output_tokens: 32000
    input_context: 200000
    generation: 4.1
    tier: "flagship"
    legacy: true
    retires: "2026-08-05"

  - provider: "claude"
    model: "Claude Opus 4"
    api_identifier: "claude-opus-4-20250514"
    input_token_price: 15.0
    output_token_price: 75.0
    max_output_tokens: 32000
    input_context: 200000
    generation: 4
    tier: "flagship"
    legacy: true

  - provider: "claude"
    model: "Claude Sonnet 4"
    api_identifier: "claude-sonnet-4-20250514"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 4
    tier: "balanced"
    legacy: true
    beta_headers:
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  # Claude 3.x Series (RETIRED - these identifiers now 404 against the live API)
  #
  # Retained deliberately, not by oversight: Bedrock/AWS identifier
  # normalization resolves regional and versioned forms through these entries
  # (ADR-0011), and removing them would break that contract. A `retires` date
  # in the past is the machine-readable marker; `legacy: true` alone could not
  # distinguish "older but fine" from "already gone".
  #
  # Claude Haiku 3 is deprecated rather than retired — its date is in the
  # future, so it stays in this block but is still callable.
  - provider: "claude"
    model: "Claude Sonnet 3.7"
    api_identifier: "claude-3-7-sonnet-20250219"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 64000
    input_context: 200000
    generation: 3.7
    tier: "balanced"
    legacy: true
    retires: "2026-02-19"
    beta_headers:
      - key: "anthropic-beta"
        value: "output-128k-2025-02-19"
        max_output_tokens: 128000
      - key: "anthropic-beta"
        value: "context-1m-2025-08-07"
        input_context: 1000000

  - provider: "claude"
    model: "Claude Haiku 3.5"
    api_identifier: "claude-3-5-haiku-20241022"
    input_token_price: 0.8
    output_token_price: 4.0
    max_output_tokens: 8192
    input_context: 200000
    generation: 3.5
    tier: "fast"
    legacy: true
    retires: "2026-02-19"

  - provider: "claude"
    model: "Claude Haiku 3"
    api_identifier: "claude-3-haiku-20240307"
    input_token_price: 0.25
    output_token_price: 1.25
    max_output_tokens: 4096
    input_context: 200000
    generation: 3
    tier: "fast"
    legacy: true
    retires: "2026-04-19"

  # Claude 3 Series (RETIRED - kept for Bedrock/AWS identifier normalization)
  - provider: "claude"
    model: "Claude Opus 3"
    api_identifier: "claude-3-opus-20240229"
    input_token_price: 15.0
    output_token_price: 75.0
    max_output_tokens: 4096
    input_context: 200000
    generation: 3
    tier: "flagship"
    legacy: true
    retires: "2026-01-05"

  - provider: "claude"
    model: "Claude Sonnet 3.5"
    api_identifier: "claude-3-5-sonnet-20241022"
    input_token_price: 3.0
    output_token_price: 15.0
    max_output_tokens: 8192
    input_context: 200000
    generation: 3.5
    tier: "balanced"
    legacy: true
    retires: "2025-10-28"

  # OpenAI GPT-5 Series (Current Generation)
  - provider: "openai"
    model: "GPT-5.2"
    api_identifier: "gpt-5.2"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5.2
    tier: "flagship"

  - provider: "openai"
    model: "GPT-5.2 Pro"
    api_identifier: "gpt-5.2-pro"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5.2
    tier: "flagship"

  - provider: "openai"
    model: "GPT-5"
    api_identifier: "gpt-5"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5
    tier: "flagship"

  - provider: "openai"
    model: "GPT-5 Pro"
    api_identifier: "gpt-5-pro"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5
    tier: "flagship"

  - provider: "openai"
    model: "GPT-5 Mini"
    api_identifier: "gpt-5-mini"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5
    tier: "balanced"

  - provider: "openai"
    model: "GPT-5 Nano"
    api_identifier: "gpt-5-nano"
    max_output_tokens: 128000
    input_context: 400000
    generation: 5
    tier: "fast"

  # OpenAI o-Series Reasoning Models
  - provider: "openai"
    model: "o3"
    api_identifier: "o3"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "flagship"

  - provider: "openai"
    model: "o3-pro"
    api_identifier: "o3-pro"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "flagship"

  - provider: "openai"
    model: "o3-mini"
    api_identifier: "o3-mini"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "fast"

  - provider: "openai"
    model: "o4-mini"
    api_identifier: "o4-mini"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "fast"
    legacy: true

  - provider: "openai"
    model: "o1"
    api_identifier: "o1"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "flagship"
    legacy: true

  - provider: "openai"
    model: "o1-pro"
    api_identifier: "o1-pro"
    max_output_tokens: 100000
    input_context: 200000
    generation: 5
    tier: "flagship"
    legacy: true

  - provider: "openai"
    model: "o1-mini"
    api_identifier: "o1-mini"
    max_output_tokens: 65536
    input_context: 128000
    generation: 5
    tier: "fast"
    legacy: true

  - provider: "openai"
    model: "o1-preview"
    api_identifier: "o1-preview"
    max_output_tokens: 32768
    input_context: 128000
    generation: 5
    tier: "flagship"
    legacy: true

  # OpenAI GPT-4.1 Series (Legacy - retiring Feb 2026)
  - provider: "openai"
    model: "GPT-4.1"
    api_identifier: "gpt-4.1"
    max_output_tokens: 32768
    input_context: 1047576
    generation: 4.1
    tier: "flagship"
    legacy: true

  - provider: "openai"
    model: "GPT-4.1 Mini"
    api_identifier: "gpt-4.1-mini"
    max_output_tokens: 32768
    input_context: 1047576
    generation: 4.1
    tier: "balanced"
    legacy: true

  - provider: "openai"
    model: "GPT-4.1 Nano"
    api_identifier: "gpt-4.1-nano"
    max_output_tokens: 32768
    input_context: 1047576
    generation: 4.1
    tier: "fast"
    legacy: true

  # OpenAI GPT-4o Series (Legacy - deprecated)
  - provider: "openai"
    model: "GPT-4o"
    api_identifier: "gpt-4o"
    max_output_tokens: 16384
    input_context: 128000
    generation: 4
    tier: "flagship"
    legacy: true

  - provider: "openai"
    model: "GPT-4o Mini"
    api_identifier: "gpt-4o-mini"
    max_output_tokens: 16384
    input_context: 128000
    generation: 4
    tier: "fast"
    legacy: true

  # OpenAI GPT-4 Turbo (Legacy - deprecated)
  - provider: "openai"
    model: "GPT-4 Turbo"
    api_identifier: "gpt-4-turbo"
    max_output_tokens: 4096
    input_context: 128000
    generation: 4
    tier: "balanced"
    legacy: true

  # Google Gemini 3 Series (Preview - Latest)
  - provider: "gemini"
    model: "Gemini 3 Pro"
    api_identifier: "gemini-3-pro-preview"
    max_output_tokens: 65536
    input_context: 1048576
    generation: 3.0
    tier: "flagship"

  - provider: "gemini"
    model: "Gemini 3 Flash"
    api_identifier: "gemini-3-flash-preview"
    max_output_tokens: 65536
    input_context: 1048576
    generation: 3.0
    tier: "balanced"

  # Google Gemini 2.5 Series (Stable)
  - provider: "gemini"
    model: "Gemini 2.5 Pro"
    api_identifier: "gemini-2.5-pro"
    max_output_tokens: 65536
    input_context: 1048576
    generation: 2.5
    tier: "flagship"

  - provider: "gemini"
    model: "Gemini 2.5 Flash"
    api_identifier: "gemini-2.5-flash"
    max_output_tokens: 65536
    input_context: 1048576
    generation: 2.5
    tier: "balanced"

  - provider: "gemini"
    model: "Gemini 2.5 Flash-Lite"
    api_identifier: "gemini-2.5-flash-lite"
    max_output_tokens: 65536
    input_context: 1048576
    generation: 2.5
    tier: "fast"

  # Google Gemini 2.0 Series (Legacy - deprecating Mar 2026)
  - provider: "gemini"
    model: "Gemini 2.0 Flash"
    api_identifier: "gemini-2.0-flash"
    max_output_tokens: 8192
    input_context: 1048576
    generation: 2.0
    tier: "fast"
    legacy: true

  - provider: "gemini"
    model: "Gemini 2.0 Flash-Lite"
    api_identifier: "gemini-2.0-flash-lite"
    max_output_tokens: 8192
    input_context: 1048576
    generation: 2.0
    tier: "fast"
    legacy: true

  # Google Gemini 1.5 Series (Retired - kept for reference)
  - provider: "gemini"
    model: "Gemini 1.5 Pro"
    api_identifier: "gemini-1.5-pro"
    max_output_tokens: 8192
    input_context: 2097152
    generation: 1.5
    tier: "flagship"
    legacy: true

  - provider: "gemini"
    model: "Gemini 1.5 Flash"
    api_identifier: "gemini-1.5-flash"
    max_output_tokens: 8192
    input_context: 1048576
    generation: 1.5
    tier: "balanced"
    legacy: true

# Provider-specific configurations
providers:
  claude:
    name: "Anthropic Claude"
    api_base: "https://api.anthropic.com/v1"
    # The model every invocation gets when `--model` / `OMNI_DEV_MODEL` is not
    # passed — the highest-impact single value in this file. Sonnet 5 is the
    # current balanced tier and the direct successor to the previous default
    # (claude-sonnet-4-6), so the cost profile is unchanged at the sticker rate
    # ($3/$15 per MTok). Kept in sync with `FALLBACK_CLAUDE_MODEL` in
    # src/claude/backend.rs, which covers the case where this file fails to load.
    default_model: "claude-sonnet-5"

    # Model tier descriptions for Claude
    tiers:
      flagship:
        description: "Most capable models for complex reasoning and analysis"
        use_cases: ["complex analysis", "research", "advanced reasoning"]

      balanced:
        description: "Good balance of capability and speed"
        use_cases: ["general tasks", "coding", "writing"]

      fast:
        description: "Optimized for speed and cost efficiency"
        use_cases: ["simple tasks", "high-volume processing", "quick responses"]

    # Default fallback limits for unknown Claude models
    defaults:
      max_output_tokens: 4096
      input_context: 200000

  openai:
    name: "OpenAI"
    api_base: "https://api.openai.com/v1"
    default_model: "gpt-5-mini"

    # Model tier descriptions for OpenAI
    tiers:
      flagship:
        description: "Most capable models for complex reasoning and analysis"
        use_cases: ["complex analysis", "research", "advanced reasoning"]

      balanced:
        description: "Good balance of capability and speed"
        use_cases: ["general tasks", "coding", "writing"]

      fast:
        description: "Optimized for speed and cost efficiency"
        use_cases: ["simple tasks", "high-volume processing", "quick responses"]

    # Default fallback limits for unknown OpenAI models
    defaults:
      max_output_tokens: 16384
      input_context: 128000

  gemini:
    name: "Google Gemini"
    api_base: "https://generativelanguage.googleapis.com/v1beta"
    default_model: "gemini-2.5-flash"

    # Model tier descriptions for Gemini
    tiers:
      flagship:
        description: "Most capable models for complex reasoning and analysis"
        use_cases: ["complex analysis", "research", "advanced reasoning"]

      balanced:
        description: "Good balance of capability and speed"
        use_cases: ["general tasks", "coding", "writing"]

      fast:
        description: "Optimized for speed and cost efficiency"
        use_cases: ["simple tasks", "high-volume processing", "quick responses"]

    # Default fallback limits for unknown Gemini models
    defaults:
      max_output_tokens: 8192
      input_context: 1048576