fix(pricing): claude-opus-4-7 is $5/$25, not $15/$75 (3x over)

[roy-tong]

Title: llm-judge pricing table lists claude-opus-4-7 at $15/$75 — actual rate is $5/$25 (3x overestimate, breaks judge cost cap)

Commit: 2f727e9a26d2a2f0edef013dc0bf09efafd82978

src/eval/llm-judge/pricing.ts:19 prices claude-opus-4-7 at $15 in / $75 out per 1M tokens, and docs/llm-as-judge.md repeats the same numbers. Anthropic has priced the Opus tier at $5/$25 since Opus 4.5; Opus 4.7 kept that rate (multiple public price trackers confirm, and Anthropic's model docs list $5/$25). $15/$75 is the pre-4.5 Opus rate.

Because this table backs the pessimistic pre-flight cap in evaluator.ts (estimateCostUsd first attempt + retry vs maxCostUsdPerEval, default $0.25), the 3x rate makes claude-opus-4-7 judge evals fail with CostCapError on inputs that would actually cost well under the cap:

• 30k-token accuracy template, maxOutputTokens 512: current table → $0.4884 (rejected); at $5/$25 → $0.1628 (passes). In practice any opus-4-7 judge prompt over ~7k tokens is rejected at the default cap.
• Stored/returned cost_usd is likewise 3x overstated for opus-4-7 evals (10k in + 1k out: $0.225 reported vs $0.075 actual).

Expected: inputUsdPer1M: 5, outputUsdPer1M: 25 for claude-opus-4-7, plus the same fix in the docs table. The file's own header notes these rates drive cost-cap decisions, so the correction is user-facing.

Minimal change: one line in src/eval/llm-judge/pricing.ts + the docs/llm-as-judge.md row (and any pinned test fixtures that assert 15/75). Happy to send a PR if that's easier to review.

元信息

• 语言：issue 用英文（跟随仓库）。
• 停止条件：维护者合并/拒绝即停；若指出 4.7 有特殊定价渠道（如 Bedrock 区域溢价）则核对后撤回。
• 不确定性说明：opus-4.7 的 $5/$25 经三个独立价格源 + 两个独立代码库交叉确认；未直接打开 Anthropic 官方页（区域重定向），如需更强证据链可在主线程用可访问网络环境补一次官方页截图。

[vercel]

@roy-tong is attempting to deploy a commit to the irparent's projects Team on Vercel.

A member of the Team first needs to authorize it.