Skip to content

feat: surface per-model cost and limit from static catalog #2

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
manojp99 merged 2 commits into from
Jun 21, 2026
Merged

feat: surface per-model cost and limit from static catalog #2

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cc07b78
docs(readme): require amplifier-agent >= 0.8.0
microsoft-amplifier Jun 21, 2026
3a527c5
feat: surface per-model cost and limit from static catalog
microsoft-amplifier Jun 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 11 additions & 2 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,7 @@ You need **three** things installed. The first two are owned by other projects;
this README walks through the official install for each so a brand-new machine
can get set up start-to-finish.

### 1. amplifier-agent — the backend server
### 1. amplifier-agent — the backend server (>= 0.8.0 required)

`amplifier-agent` is the OpenAI-compatible HTTP server this adapter talks to.
Install it via [`uv`](https://docs.astral.sh/uv/) (Astral's Python package
Expand All @@ -48,10 +48,19 @@ curl -LsSf https://astral.sh/uv/install.sh | sh
# Install amplifier-agent as a global tool
uv tool install amplifier-agent

# Verify
# Verify — output MUST report 0.8.0 or higher
amplifier-agent --version
```

> **Version requirement: `amplifier-agent >= 0.8.0` is mandatory.** Versions
> below 0.8.0 do not ship the `serve chat-completions` HTTP face, multi-provider
> routing, or the `auth` subcommand — `amplifier-opencode` will fail to spawn
> the server, populate `/v1/models`, or read persisted credentials.
>
> If `amplifier-agent --version` reports an older version, upgrade with
> `uv tool upgrade amplifier-agent` and re-run `amplifier-opencode doctor`
> to confirm.

For full install options (source builds, pinned versions, etc.) see the
[amplifier-agent README](https://github.com/microsoft/amplifier-agent#install).

Expand Down
138 changes: 133 additions & 5 deletions src/amplifier_app_opencode/cli.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,6 +79,77 @@
}


# ---------------------------------------------------------------------------
# Static per-model price catalog (USD per 1M tokens)
# ---------------------------------------------------------------------------

# opencode looks up pricing by ``(providerID, modelID)``. The models.dev
# catalog opencode bundles has entries for the canonical provider IDs
# (anthropic, openai, ...) but NOT for our custom ``amplifier`` provider
# -- so opencode renders our turns at $0.00 unless we declare cost in
# the model entry ourselves.
#
# We mirror models.dev's pricing here so the model picker and per-turn
# cost display work out of the box. Source of truth:
# https://github.com/sst/models -- the same registry opencode uses.
#
# Maintenance: when an upstream provider changes prices, update this
# table. amplifier-agent's per-turn ``cost_usd`` (PR #68 on amplifier-agent)
# still ships the authoritative dollar value on the wire for clients that
# read it; this catalog is purely for opencode's TUI cost display.
#
# Verified pricing date: 2026-06-21.
MODEL_PRICING_PER_MILLION: dict[str, dict[str, float]] = {
# ===== Anthropic =====
"claude-haiku-4-5-20251001": {"input": 1.0, "output": 5.0, "cache_read": 0.1, "cache_write": 1.25},
"claude-sonnet-4-6": {"input": 3.0, "output": 15.0, "cache_read": 0.3, "cache_write": 3.75},
"claude-opus-4-8": {"input": 5.0, "output": 25.0, "cache_read": 0.5, "cache_write": 6.25},
# ===== OpenAI =====
# GPT-4o
"gpt-4o": {"input": 2.5, "output": 10.0, "cache_read": 1.25},
"gpt-4o-mini": {"input": 0.15, "output": 0.6, "cache_read": 0.075},
# GPT-4.1
"gpt-4.1": {"input": 2.0, "output": 8.0, "cache_read": 0.5},
"gpt-4.1-mini": {"input": 0.4, "output": 1.6, "cache_read": 0.1},
"gpt-4.1-nano": {"input": 0.1, "output": 0.4, "cache_read": 0.025},
# GPT-5
"gpt-5": {"input": 1.25, "output": 10.0, "cache_read": 0.125},
"gpt-5-codex": {"input": 1.25, "output": 10.0, "cache_read": 0.125},
"gpt-5-mini": {"input": 0.25, "output": 2.0, "cache_read": 0.025},
"gpt-5-nano": {"input": 0.05, "output": 0.4, "cache_read": 0.005},
"gpt-5-pro": {"input": 15.0, "output": 120.0},
# GPT-5.1
"gpt-5.1": {"input": 1.25, "output": 10.0, "cache_read": 0.125},
"gpt-5.1-codex": {"input": 1.25, "output": 10.0, "cache_read": 0.125},
"gpt-5.1-codex-max": {"input": 1.25, "output": 10.0, "cache_read": 0.125},
"gpt-5.1-codex-mini": {"input": 0.25, "output": 2.0, "cache_read": 0.025},
# GPT-5.2
"gpt-5.2": {"input": 1.75, "output": 14.0, "cache_read": 0.175},
"gpt-5.2-codex": {"input": 1.75, "output": 14.0, "cache_read": 0.175},
# o-series (reasoning models)
"o1": {"input": 15.0, "output": 60.0, "cache_read": 7.5},
"o1-pro": {"input": 150.0,"output": 600.0},
"o3": {"input": 2.0, "output": 8.0, "cache_read": 0.5},
"o3-mini": {"input": 1.1, "output": 4.4, "cache_read": 0.55},
"o3-pro": {"input": 20.0, "output": 80.0},
"o4-mini": {"input": 1.1, "output": 4.4, "cache_read": 0.275},
# Embeddings (output is always 0; opencode schema requires both fields)
"text-embedding-3-large": {"input": 0.13, "output": 0.0},
"text-embedding-3-small": {"input": 0.02, "output": 0.0},
}


def lookup_pricing(model_id: str) -> dict[str, float] | None:
"""Return per-million-token pricing for ``model_id``, or None when absent.

Exact-match lookup against :data:`MODEL_PRICING_PER_MILLION`. Returns
``None`` for unknown models so callers can distinguish "no pricing
declared" from "free model" (emitting zeros would falsely claim the
model is free).
"""
return MODEL_PRICING_PER_MILLION.get(model_id)


def resolve_global_config_path() -> Path:
"""Locate (or pick) the global opencode config file.

Expand Down Expand Up @@ -215,18 +286,75 @@ def build_provider_block(
) -> dict[str, Any]:
"""Transform amplifier-agent's /v1/models into an opencode provider block.

opencode's per-model schema lenient-strips unknown keys but fails the
WHOLE config decode on type mismatches in known fields. Empirically
(against opencode v1.17.8): ship only ``name`` per entry. Everything
else opencode infers from defaults during provider state init. Matches
the docs' "Atomic Chat" example.
opencode's per-model schema lenient-strips unknown keys but FAILS the
WHOLE config decode on type mismatches in known fields. When decode
fails on the global config, opencode catches via ``orElseSucceed(() =>
({}))`` -- silently wiping every provider's models block. We're
therefore paranoid about types: every field we surface is strictly
validated before emission, and we skip a field rather than risk
dropping the whole entry.

Fields surfaced:

- ``name`` (always; falls back to id when no display_name)
- ``cost`` ``{input, output, cache_read?, cache_write?}`` -- USD
per 1M tokens. Pulled from the static
:data:`MODEL_PRICING_PER_MILLION` catalog because
opencode keys catalog lookups by ``(providerID,
modelID)`` and our custom ``amplifier`` providerID
doesn't match anything in opencode's bundled models.dev
registry. Without this declaration, opencode renders
every turn at $0.00. Emitted only when both ``input``
and ``output`` are present in the catalog (skipped
silently for unknown models).
- ``limit`` ``{context, output}`` -- token-count budgets opencode
uses for context-window warnings. Lifted from
amplifier-agent's /v1/models ``limit`` field
(originally ``ModelInfo.context_window`` /
``max_output_tokens``).

Note: amplifier-agent's PR #68 ALSO surfaces real per-turn
``cost_usd`` on the chat-completions response (telemetry on the
wire). The catalog here is for opencode's TUI display since
opencode's @ai-sdk/openai-compatible adapter doesn't read
``cost_usd`` today.
"""
models_block: dict[str, dict[str, Any]] = {}
for m in models:
mid = m.get("id")
if not isinstance(mid, str) or not mid:
continue

entry: dict[str, Any] = {"name": m.get("display_name") or mid}

# Cost block — from our static catalog. Skip when the model is
# not in the catalog (showing $0 zeros is worse than showing
# nothing; an empty cost block lets opencode know the data is
# missing rather than claiming the model is free).
pricing = lookup_pricing(mid)
if pricing is not None:
cost_entry: dict[str, float] = {
"input": float(pricing["input"]),
"output": float(pricing["output"]),
}
if "cache_read" in pricing:
cost_entry["cache_read"] = float(pricing["cache_read"])
if "cache_write" in pricing:
cost_entry["cache_write"] = float(pricing["cache_write"])
entry["cost"] = cost_entry

# Limit block — only emit when both context AND output are ints.
limit = m.get("limit")
if (
isinstance(limit, dict)
and isinstance(limit.get("context"), int)
and isinstance(limit.get("output"), int)
):
entry["limit"] = {
"context": int(limit["context"]),
"output": int(limit["output"]),
}

models_block[mid] = entry

return {
Expand Down