From 9aaca572c5b42751d4f43cd1cc05ae4e2876cfb9 Mon Sep 17 00:00:00 2001 From: argszero Date: Sun, 18 Oct 2026 11:22:01 +0800 Subject: [PATCH] docs(config): state why Claude's rows have no long tiers, and stop implying a pending gap MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The catalogue comment listed the models whose long-context tiers are deliberately empty and said, in one breath, that every unfilled row is a "vendor has not published the number" case still waiting to be filled. Two parts of that are wrong: - Claude is not an unpublished number. Its pricing page has a "Long context pricing" section stating that Claude 4.6 and later (except Haiku 5.5) include the full 1M context window at standard pricing, and that a 900k-token request costs the same per token as a 9k one. So claude-opus-5 and claude-sonnet-5.5 correctly have no tiers at all -- reading the old note as "go find the numbers" invites someone to invent tiers that the vendor never charges for. - The "一律" universal does not hold: the reasons differ per family. For qwen3.7-flash the tier table simply has no cache-hit column, so adding tiers would zero that price; for the OpenAI rows the short/long split exists but its boundary is not pinned down. The comment now keeps the filled roster, gives each unfilled family its own reason, and records the rule that applies if Claude Haiku 5.5 is ever added (tiers start at prompt > 100,000 tokens, i.e. 100001). Comment-only change: 1 file, +9/-3, no behaviour difference. --- config/config.example.toml | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/config/config.example.toml b/config/config.example.toml index 4dc36e8..c2ae488 100644 --- a/config/config.example.toml +++ b/config/config.example.toml @@ -365,9 +365,15 @@ endpoints = [ # output_per_m = 16.80 # ⛔ 档位数据只填官网确实给了的：本仓 config 是唯一真源（src/db.rs::seed_models），填错即错账。 # ⚠️ 已填长档的（官网核实，见各行注释）：minimax-m3（>512k）、gemini-3.1-pro、grok-4.6、 -# grok-4.7（后三家都是 >200k）。未填的一律是「官网**没给这个数**」，不是漏了： -# qwen3.7-flash（分档只公布输入/输出两列、缓存命中价未公布 ⇒ 补档会把命中价打成 0 少收）、 -# OpenAI 各家（分了短/长两段，但**没公布分界 token 数**）、Claude（价格页无长档一段）—— 待补。 +# grok-4.7（后三家都是 >200k）。未填的**原因各不相同**，⛔ 不要一律当成「待补」： +# · qwen3.7-flash：分档表只公布输入/输出两列、缓存命中价未公布 ⇒ 补档会把命中价打成 0 少收； +# · OpenAI 各家：分了短/长两段，但**分界 token 数未钉死** ⇒ 不填； +# · Claude：**不是**「官网没给数」，而是官网明说**没有长档溢价** —— pricing 页 Long context +# pricing 一节（2026-10-18 核实，https://platform.claude.com/docs/en/about-claude/pricing）： +# 4.6 及以后（Haiku 5.5 除外）全长 1M 上下文按标准价计，900k 的请求与 9k 同价 ⇒ +# claude-opus-5 / claude-sonnet-5.5 **本来就不该有档**；将来若上 Haiku 5.5，它的档界是 +# prompt > 100,000 tokens（含缓存读/写），届时按 from_tokens = 100001 填； +# · 其余未填行：现核实的官网页只有一组价（没有长档表）⇒ 无档可填，不是漏填。 # 高峰时段价（rant 2026-08-20T11:58:40）：peak_input_per_m / peak_output_per_m / # peak_cache_hit_input_per_m 三个可选字段，缺省 0 = 不启用高峰计费（沿用空闲价）。 # 命中高峰时段（北京时间 9-12、14-18，**周一至周五**且不在 [billing] holidays 内）