usage.cost field is wrong by ~5 orders of magnitude for claude-opus-4-7

Summary

On claude-opus-4-7 responses, openclaw's computed usage.cost.cacheWrite (and related per-category cost fields) appear to be using an incorrect or uninitialized price multiplier. Observed values are around $1.9×10⁻⁵ per million tokens, which is ~5 orders of magnitude below any real Anthropic price tier.

Downstream effects: openclaw usage-cost and per-session cost estimates are unusable; session logs' cost.total field is effectively zero for any Opus 4.7 activity.

Evidence

Single captured assistant turn on Opus 4.7:

{
  "usage": {
    "input": 6,
    "output": 15,
    "cacheRead": 45612,
    "cacheWrite": 285595,
    "totalTokens": 331228,
    "cost": {
      "input": 9E-11,
      "output": 1.125E-9,
      "cacheRead": 6.8418E-8,
      "cacheWrite": 0.00000535490625,
      "total": 0.000005424539249999999
    }
  }
}

Math check:

• cacheWrite: 285,595 tokens, reported cost $0.00000535490625 → implies $1.875×10⁻⁵/M.
• Expected at cacheRetention: "long" (1h TTL, 2× input price): for a $5/M input model that's $10/M. For a $15/M input model that's $30/M.
• Either way, the reported value is off by ~5 orders of magnitude.

Same session shows correct cost calculations for earlier turns on claude-opus-4-6:

• cacheWrite: 25716 tokens, cost $0.160725 → $6.25/M (1.25× input at $5/M → correct for 5m TTL).

So the pricing table has Opus 4.6 entries but not Opus 4.7 entries, and Opus 4.7 is falling through to a default / zero-initialized entry that gives near-zero costs.

Root cause (hypothesis)

openclaw maintains a per-model pricing table used to convert usage token counts → cost estimates for session logs and CLI reports. The claude-opus-4-7 entry is either missing, or present with zero/near-zero values.

This doesn't affect Anthropic's actual billing — Anthropic charges based on their own rates regardless of what openclaw computes. It only affects openclaw's internal cost reporting.

Fix

Add the correct per-token rates for claude-opus-4-7 to the pricing table:

• Input: (current Opus 4.7 input rate per million tokens, consult Anthropic's published pricing)
• Output: (current Opus 4.7 output rate)
• Cache read: 0.1× input
• Cache creation 5m: 1.25× input
• Cache creation 1h: 2× input

Also recommend adding a defensive fallback: when computing cost for a model id not in the price table, log a warning rather than silently returning near-zero values. This makes future model additions more obvious when they're missed.

Environment

• openclaw 2026.4.20-beta.2
• Claude Opus 4.7 (claude-opus-4-7)
• Authentication: Anthropic API key (not subscription)

Scope

Trivial fix — adding price-table entries and the fallback warning should be under 20 LOC. Happy to submit the PR once Anthropic's current Opus 4.7 pricing is confirmed (I don't want to guess numbers from the outside and introduce another wrong entry).

Spotted while investigating cache cost behavior, with diagnostic assistance from Claude.