Unpriced warning fires on subscription-billed models where $0 is correct, and its suggested model-alias fix would fabricate spend

[therickfactr]

Problem

The unpriced-models warning from #638 fires on models whose cost genuinely is zero, and its suggested fix would make the totals wrong if followed.

#638 anticipated exactly this ambiguity — "a new model id shipping ahead of our pricing table … undercounts real spend and looks identical to a genuinely free model." In practice the genuinely-free case is common: models billed under a flat-rate subscription have no per-token cost to compute.

Evidence

On my machine the warning reports these at $0:

model	tokens	calls	why it is $0
auto-genius	35.3M	898	ClinePass codename — opaque routed model, flat-rate pass
warp	17.7M	449	Warp's built-in agent, subscription
codex-auto-review	7.2M	940	runs under the ChatGPT subscription
grok-composer-2.5-fast	1.9M	10	Grok subscription
big-pickle	33.9K	4	ClinePass codename

None of these has a per-token price to discover. The marginal cost of the call is zero because the spend already happened as a subscription fee.

Why the suggested fix is harmful here

The warning advises:

Fix: codeburn model-alias "<model>" <known-model>

Applied to any of the above, that maps a flat-rate model onto another model's per-token rate and invents spend that was never billed. auto-genius alone is 35.3M tokens — aliasing it to a mid-tier model would fabricate hundreds of dollars, in the one number the tool exists to report accurately. For ClinePass specifically the underlying model is deliberately opaque and can rotate, so there is no correct alias to choose even in principle.

So today the user has two options and both are wrong: leave it and get warned indefinitely about correct data, or alias it and corrupt the totals.

Suggested direction

A third state alongside "priced" and "unpriced": flat-rate / subscription-billed, meaning $0 is the right answer and no warning is due.

There is already precedent for this shape in the codebase — local models are special-cased out of the warning via looksLikeLocalModel() / getLocalSavingsBaseline(), and surfaced separately as a "Savings (local models)" line. Subscription-billed models seem to want the same treatment: excluded from the unpriced list, and ideally attributed to their plan rather than shown as free.

Mechanically that could be a small classifier list (ClinePass codenames, warp, codex-auto-review, Grok composer variants) plus a model-alias-style escape hatch to mark one by hand, e.g. codeburn model-flat-rate "<model>".

Relationship to existing issues

• Unpriced models silently contribute $0 to totals; warning is off by default #638 (closed) introduced the warning. This is the false-positive half of the ambiguity it named, not a request to remove it — the missing-pricing case it targets is real and worth keeping.
• Cost modes: reconcile displayed cost with the source's own reported cost (auto/calculate/display) #890 (open, cost modes) overlaps but does not cover this: its display mode reconciles a source's reported cost against ours, and these sources report no cost at all. If a plan-aware attribution model is in scope for Cost modes: reconcile displayed cost with the source's own reported cost (auto/calculate/display) #890, though, this may belong there instead — happy to fold it in rather than track separately.

Environment: codeburn 0.9.20 (npm), macOS 26.6.1 (25G76) arm64, Node v26.5.0.