feat(app): add DeepSeek V4 Pro to the profit estimators with DeepSeek list-price defaults / 利润估算器新增 DeepSeek V4 Pro 并默认采用 DeepSeek 官方定价

[functionstackx]

Summary

Adds DeepSeek V4 Pro (DeepSeek V4 Pro 0813 1.6T) to /profit-estimator and /profit-estimator-per-gigawatt, plus the /zh mirrors, following the pattern from #1001 (GLM 5.2/5.3) and #1002 (MiniMax M3).

• Routes: /profit-estimator/deepseek-v4 and /profit-estimator-per-gigawatt/deepseek-v4 (PROFIT_ESTIMATOR_MODELS in model-routes.ts; the slug is the site-wide deepseek-v4 from the compare registry). Kimi K3 stays the bare-path default, so DeepSeek V4 Pro gets its own slugged, indexable page here even though it is the app-wide default elsewhere. Models outside the allow-list still 404 and stay out of the sitemap.
• Per-model defaults (profitModelDefaults(Model.DeepSeek_V4_Pro)):
  • 24 tok/s/user, the speed DeepSeek's own API serves at.
  • DeepSeek list price $1.32 input / $0.044 cached / $3.96 output per M tok, the peak-hour rate for deepseek-v4-pro-0813 on the DeepSeek pricing page (off-peak is 50% of that: $0.66 / $0.022 / $1.98). Pinned as a list price for the same reason as GLM and MiniMax: the OpenRouter aggregate for deepseek/deepseek-v4-pro-0813 sits below even the off-peak rate (about $0.58 / $0.019 / $1.74 at time of writing), so the catalog would understate what the lab charges. OpenRouter and Custom remain one click away; the caption links the DeepSeek pricing page when the list price is in force.
  • 30% model license fee, the DEFAULT_LAB_CUT_PCT.
• Coverage at 24 tok/s/user (checked against the live agentic rows, P90): B200 (SGLang min 16.8, vLLM min 19.9), B300 (SGLang 22.4, vLLM 8.3), and MI355X (ATOM 10.8, SGLang 12.0, vLLM 4.3, MORI-SGLang disagg 23.5) all have a measured point on both sides of 24, so they are priced. GB200 Dynamo-vLLM (min 40.1), GB300 (Dynamo-SGLang disagg 29.6, Dynamo-TRT disagg 37.2, Dynamo-vLLM disagg 38.8), and H200 Dynamo-SGLang (min 26.5) bottom out above 24 and list under Not priced. This is the existing clamp-skip behaviour, not new logic; a lower-interactivity run on those SKUs will price them without a code change.
• Model switch re-seeds the interactivity target, price source, and license fee in every direction (45/OpenRouter/30% for Kimi, 100/list/30% for GLM, 83/list/20% for MiniMax, 24/list/30% for DeepSeek).
• Docs: docs/tco-calculator.md profit-estimator section.

Tests

• Unit (vitest): profitModelDefaults(Model.DeepSeek_V4_Pro) and its list to revenue-pricing conversion; the fallback case now uses Model.DeepSeek_R1 since DeepSeek V4 Pro has an entry. modelRoutesForTab expects [DeepSeek V4 Pro, Kimi K3, MiniMax M3, GLM 5.2] in MODEL_ROUTES (selector) order, the deepseek-v4 slug, and Kimi K3 as the tab default. bun run typecheck, bun run lint, bun run fmt, the typography check, and the full vitest suite pass locally (the one pre-existing benchmark-transform.test.ts failure reproduces on master in my sandbox and is unrelated).
• E2E (Cypress): fixtures now emit dsv4 agentic rows keyed by ?model=DeepSeek-V4-Pro. Both fixture curves start below 24 tok/s/user, so the new spec asserts all five SKUs (H200 included) are priced there, unlike the 45/83/100 defaults where H200 is skipped. New specs cover the /deepseek-v4 route defaults, caption, and pricing-source link; the 3-option price selector (OpenRouter reads the DeepSeek row at $0.66 / $1.98, not Kimi's); Custom seeding from the list price; Kimi to DeepSeek to Kimi model-switch re-seeding with the expected pathnames; and the /zh mirror label (DeepSeek 官方定价). The Kimi selector assertion now expects four models, and the "404s models the estimator does not serve" checks use deepseek-r1 instead of the never-valid deepseek-v4-pro slug. I could not run Cypress in my sandbox, so please rely on CI for the e2e run.

中文说明

在 /profit-estimator 与 /profit-estimator-per-gigawatt（含 /zh 镜像）新增 DeepSeek V4 Pro（DeepSeek V4 Pro 0813 1.6T），路径为 /deepseek-v4，沿用 #1001（GLM 5.2/5.3）与 #1002（MiniMax M3）的模式。Kimi K3 仍为默认模型，其他模型继续返回 404。

• 按模型的默认值：DeepSeek V4 Pro 默认 24 tok/s/user（DeepSeek 官方 API 的实际服务速度），并采用 DeepSeek 官方定价：输入 $1.32 / 缓存输入 $0.044 / 输出 $3.96（每百万 token），即官方定价页上 deepseek-v4-pro-0813 的高峰时段价格（非高峰为 5 折：$0.66 / $0.022 / $1.98）。不直接读取 OpenRouter 的原因与 GLM、MiniMax 相同：其聚合价（约 $0.58 / $0.019 / $1.74）低于官方价（甚至低于非高峰价），会低估实验室的实际收入；OpenRouter 与自定义仍可一键切换，使用官方定价时图表说明会链接 DeepSeek 定价页。模型许可费保持默认 30%。
• 24 tok/s/user 处的覆盖（对照线上 agentic 数据，P90）：B200、B300、MI355X 均有实测点覆盖，可计价；GB200 Dynamo-vLLM（最低约 40）、GB300（最低约 30）、H200 Dynamo-SGLang（最低约 27）曲线下限高于该点，按现有逻辑列为"未计价"，待更低交互性的运行数据落地后无需改代码即可计价。
• 切换模型时会重置交互性目标、价格来源与许可费。
• 同步更新 docs/tco-calculator.md。

测试：新增 vitest 单元测试（默认值、官方定价换算、路由白名单与 slug）；Cypress fixture 新增 dsv4 agentic 数据并按 ?model= 参数返回，新增 DeepSeek 路由默认值（含 H200 在 24 tok/s/user 下可计价）、价格来源选择器、自定义价格初始化、Kimi → DeepSeek → Kimi 切换重置与中文镜像的 e2e 用例。本地已通过 typecheck / lint / fmt / typography 与完整单元测试；沙箱中无法运行 Cypress，请以 CI 结果为准。

---

Note

Low Risk
Scoped feature addition to profit estimator routing and defaults, with tests and docs; no auth or critical infrastructure changes.

Overview
Adds DeepSeek V4 Pro to both profit estimator tabs (/profit-estimator and /profit-estimator-per-gigawatt, including /zh), via the existing deepseek-v4 slug. Kimi K3 stays the bare-path default; DeepSeek gets its own indexable route even though it is the app-wide default elsewhere.

Per-model behavior follows GLM/MiniMax: profitModelDefaults seeds 24 tok/s/user, DeepSeek peak list pricing ($1.32 / $0.044 cached / $3.96 per M tok with a link to DeepSeek’s pricing docs), and the standard 30% model license fee. Model switches re-seed interactivity, price source, and license fee like the other supported models.

Routing & discovery: PROFIT_ESTIMATOR_MODELS in model-routes.ts now includes DeepSeek V4 Pro; unsupported models still 404. Docs in tco-calculator.md describe coverage notes at 24 tok/s/user (some SKUs may show as not priced until lower-interactivity data exists).

Tests: Vitest covers defaults and route allow-list; Cypress fixtures add dsv4 agentic rows and a dedicated DeepSeek spec (defaults, OpenRouter/custom price flows, Kimi ↔ DeepSeek switching, Chinese labels). Selector expectations move to four models; 404 checks use deepseek-r1 instead of the formerly invalid slug.

^{Reviewed by Cursor Bugbot for commit 4dac848. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Sep 5, 2026 12:51am UTC

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 4dac848. Configure here.}

[cursor]

Wrong Kimi switch pathname assertion

Medium Severity

Switching from DeepSeek back to Kimi K3 asserts the bare /profit-estimator-per-gigawatt path. modelRoutePathnameRewrite keeps the slugged /kimi-k3 URL on an in-page switch; only a fresh visit canonicalizes the default model to the bare path. This new spec will fail in Cypress.

 

^{Reviewed by Cursor Bugbot for commit 4dac848. Configure here.}

[blacksmith-sh]

Found 4 test failures on Blacksmith runners:

Failures

Test	View Logs
Profit Estimator — DeepSeek V4 Pro/ re-seeds the operating point and price source when switching from Kimi K3	View Logs
Profit Estimator — DeepSeek V4 Pro/ re-seeds the operating point and price source when switching from Kimi K3	View Logs
Profit Estimator — MiniMax M3/ re-seeds the operating point and price source when switching from GLM	View Logs
Profit Estimator — MiniMax M3/ re-seeds the operating point and price source when switching from GLM	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}