[twtrubiks]

Depends on #3357. This branch is stacked on it, so the first two commits belong to #3357. Only the last commit (feat(runtime): add OpenAI, xAI, and DashScope pricing) is new here. I'll rebase onto main once #3357 lands.

Summary

Models from the OpenAI-compatible providers (OpenAI, xAI, DashScope) have no entry in pricing_for_model, so they fall back to default_sonnet_tier() ($3 / $15 per MTok after #3357) and are marked pricing=estimated-default. For many of them that is far off. For example, gpt-5.4-nano is estimated at 15x its input price, and qwen-plus at about 26x.

This adds a table of published rates (input / output / cache write / cache read per MTok), keyed by model id. An id also matches its dated snapshots and -latest alias (gpt-4.1-2025-04-14, qwen-plus-latest), and provider prefixes such as openai/ and qwen/ are stripped. Matching is strict, so o1 does not match o1-mini and qwen-plus does not match qwen-plus-character.

Provider	Models	Source
OpenAI	gpt-6-*, gpt-6.1-sol, gpt-5.6-*, gpt-5.5, gpt-5.4 (+ mini/nano), gpt-5.2, gpt-5.1, gpt-5 (+ mini/nano), gpt-4.1 (+ mini/nano), gpt-4o (+ mini, + the 2024-05-13 snapshot), o1, o3, o3-mini, o4-mini	Standard tier
xAI	grok-4.7, grok-4.6, grok-4.5, grok-4.3, grok-4.20*, grok-build, grok-code-fast*; retired grok-3, grok-4/grok-4-0709, grok-4-fast-*, grok-4-1-fast-*	Models, May 15 retirement
DashScope	qwen-max, qwen-plus, qwen-flash, kimi-k2.5	Model Studio pricing, context cache

Pricing choices:

• Tiers: long-context (OpenAI >272K, xAI >=200K) and DashScope input-length tiers use the base rate, because usage is aggregated across requests. This is the same approach fix(runtime): update model pricing table to current rates #3357 takes for Haiku 5.5.
• Cache: providers without a separate cache-write charge use the input rate for cache writes. Models without a cached-input discount (e.g. gpt-4o-2024-05-13) use the input rate for cache reads. The OpenAI-compatible client already maps prompt_tokens_details.cached_tokens to cache reads.
• xAI retirements: the slugs retired on 2026-05-15 still resolve, but xAI bills them at grok-4.3 rates ($1.25 / $2.50), so they are priced that way. grok-code-fast-* is an alias of grok-build-0.1.
• DashScope region: prices depend on the deployment region, which is not visible from the model id. This uses the China (Beijing) rates, since claw's default DASHSCOPE_BASE_URL is the Beijing endpoint. Users of the international (Singapore) endpoint pay roughly 3–4x more. Output uses the non-thinking rate because claw does not enable Qwen thinking mode. Cache hits use the 20% implicit-cache rate.

Anti-slop triage

• Classification: actionable-fix
• Evidence: official pricing pages linked above; new prices_openai_compatible_provider_models test covering snapshots, prefixes, retired xAI slugs, and ids that must stay unpriced.
• Non-destructive review result: merge candidate (after fix(runtime): update model pricing table to current rates #3357)

Verification

• cargo fmt --all --check
• cargo test -p runtime --lib usage, cargo test -p api: pass
• cargo clippy -p runtime --all-targets: no new warnings in the touched file.
• Targeted tests/docs checks ran, or the gap is explicitly recorded.
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Compatibility and known gaps

• Output change: for the listed models, /cost, usage summaries, and the estimated_cost_usd JSON / telemetry fields now show provider rates instead of the default, and drop the pricing=estimated-default marker. Field names and formats are unchanged.
• Left unpriced (still the default):
  • Responses-only *-pro models, which claw's Chat Completions client cannot call.
  • Ids not on the current pricing pages: o1-mini, grok-3-mini, grok-2, qwen-turbo, qwq-plus, qwen3-max, qwen3-coder-plus.
  • Other DashScope variants (qwen3.x-*, dated kimi-*), Groq, and local models.
• I don't have API keys for these providers, so the rates were checked against the official docs and unit tests rather than live bills.
• Provider prices change often. The rates are kept in one table so updates touch a single place.

Resolution gate

• If this PR resolves an issue, the issue number and fix evidence are linked. (No existing issue; evidence is in this PR.)
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

[XinwangTao]

This fills a real gap — without entries in pricing_for_model, OpenAI/xAI/DashScope models fell back to the default tier and produced badly skewed estimates (gpt-5.4-nano at 15x, qwen-plus at ~26x input price). The strict matching (no o1→o1-mini, no qwen-plus→qwen-plus-character) and the dated-snapshot/-latest alias handling are the right details. Consistent with #3357's approach on long-context tiers.

[XinwangTao]

这个补充很有价值：之前 OpenAI 兼容提供商的模型全部回退到估算默认价（gpt-5.4-nano 差 15 倍、qwen-plus 差约 26 倍），有了精确定价表后成本预估会准确很多。严格按模型 ID 匹配、避免 o1 误配 o1-mini 的处理很稳妥；对已退役但仍在计费的 xAI 老模型按 grok-4.3 费率定价也符合实际。分支叠加在 #3357 上没问题，等 #3357 合并后 rebase 即可，支持合并。

[XinwangTao]

补充这些模型定价很及时：之前 default tier 沿用旧 Opus 4 的 $15/$75，Sonnet 会被高估 5 倍，DeepSeek 系列因为没有独立条目更是直接落到默认档、估算偏差 10 倍以上，直接影响 /cost 和 estimated_cost_usd。补齐 OpenAI、xAI、DashScope 和 DeepSeek 后成本统计才可靠，family_version() 用版本号区分 tier 的思路也清楚。

[XinwangTao]

补充这些模型定价很及时：之前 default tier 沿用旧 Opus 4 的 $15/$75，Sonnet 会被高估 5 倍，DeepSeek 系列因为没有独立条目更是直接落到默认档、估算偏差 10 倍以上，直接影响 /cost 和 estimated_cost_usd。补齐 OpenAI、xAI、DashScope 和 DeepSeek 后成本统计才可靠，family_version() 用版本号区分 tier 的思路也清楚。