Group rankings by model line, fetch source URLs, show output cost by jammyfu · Pull Request #3 · jammyfu/TokenSpeedSimulator

jammyfu · 2026-09-13T10:29:58Z

Summary

Follow-up polish on the merged AA ranking demo (PR #2), including later same-PR asks.

Ranking UI is grouped by mainstream lines/families instead of one flat list: OpenAI (GPT), Anthropic (Claude), Google (Gemini), SpaceXAI (Grok), DeepSeek, Alibaba (Qwen), Z AI (GLM), Kimi, MiniMax, plus speed specialists / other.
Each snapshot row has a family field. Global AA rank numbers and median TPS stay. Doubao and ERNIE are omitted because the 2026-09-13 AA public table has no median Tokens/s for those rows (no invented numbers).
Switching A or B to a model in a different line resets that pane’s stream. Same-line SKU / TPS tweaks keep existing text.
Source text is editable on every breakpoint. Paste a public http(s) URL and Fetch readable title + body into the source box (Vite /api/fetch-page + client fallbacks). Fetch updates input only.
Models with an AA-published output $/1M show that price, a $/s · $/min · $/h burn rate, and cumulative spend while streaming. Missing prices render as —. Formula: spend ≈ tokens × output $/token; rate ≈ TPS × output $/token.
One primary Play action. zh/en labels, 100dvh layout, no paid AA API.

How grouping + clear work

src/data/modelFamilies.ts orders household Western + CN lines, then specialists/other.
Desktop: sticky family headers. Mobile: family pills filter the compact chip row.
isSameFamily(prev, next) decides whether streamA.reset() / streamB.reset() runs.

How source / URL fetch works

Controls always expose a source <details>.
/api/fetch-page fetches the URL, strips HTML, rejects localhost/private addresses.

How cost works

Field: outputPricePerMillionUsd (USD per 1M output tokens), set only when an AA model page quotes it (Celeris, Mercury, DeepSeek V4.1 Flash, Claude Opus 5, Grok 4.6, Qwen3.8 Max, GLM-5.3 max, MiniMax-M3).
Rate = active TPS × $/token; spend = tokens emitted × $/token.
Stats show $/s · $/min · $/h plus running spend; compare shows A/B per-second rates. Footer has the formula.

Verification

python3 tools/verify.py passed.
Browser-checked family grouping, same-line keep, cross-line clear (incl. compare B), URL fetch of example.com, ~390px pills + Play, and cost ticking on Celeris-1.

Present AA median TPS rows under household lines (OpenAI, Anthropic, Google, DeepSeek, specialists/other). Switching A or B to a different line resets that pane; same-line SKU changes keep streamed text. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

Restore the source editor on every breakpoint and add a secondary Fetch action that loads a public page into the source box via a same-origin Vite proxy, with client-side fallbacks. Fetching updates input only; output still clears only when the model line changes. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

Stops the token loop from gluing the last word of a page onto the next repeat. Notes browser verification of grouping and line-change clear. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

Group Qwen, GLM, Kimi, MiniMax, and Grok from the AA public table (Doubao/ERNIE omitted — no published median TPS). Show AA output $/1M, burn rate, and cumulative spend when a list price exists. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

chatgpt-codex-connector · 2026-09-13T11:13:35Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-13T11:17:36.009246Z	`c50d402`	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

Single-model stats now read $/s · $/min · $/h from active TPS × AA output list price. Compare mode still shows A/B per-second rates. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c50d402bb0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-09-13T11:17:33Z

+      const upstream = await fetch(target.href, {
+        redirect: 'follow',


Revalidate redirect targets before following them

When the dev or preview server is reachable by an untrusted user, redirect: 'follow' lets an initially public URL redirect to a loopback, metadata, or other private address without passing the new destination through parsePublicHttpUrl. Because the endpoint returns the fetched body, this creates an SSRF path; follow redirects manually and validate every destination before requesting it.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-09-13T11:17:33Z

+    throw new SourceFetchError('blocked');
+  }
+  if (/^\d+\.\d+\.\d+\.\d+$/.test(host) && isPrivateIPv4(host)) {
+    throw new SourceFetchError('blocked');


Reject bracketed private IPv6 destinations

When fetching a literal IPv6 URL such as http://[::1]:8080/, URL.hostname retains the brackets, so the value does not match the ::1 entry in BLOCKED_HOSTS, and the IPv4-only check below also ignores it. The proxy therefore accepts loopback and other private IPv6 destinations and can expose internal HTTP services; normalize IPv6 literals and reject loopback, link-local, and private IPv6 ranges.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-09-13T11:17:33Z

+  const spentA = cumulativeCostUsd(streamA.tokensCount, selected?.outputPricePerMillionUsd);
+  const spentB = cumulativeCostUsd(streamB.tokensCount, compare?.outputPricePerMillionUsd);


Accumulate cost across same-family price changes

When tokens have already streamed and the user switches to another model in the same family, the intentional no-reset behavior preserves tokensCount, but this calculation reprices every previous token using the newly selected model. For example, switching from Celeris-1 to Mercury 2 retroactively changes the displayed total, while switching to an unpriced specialist erases it entirely; track spend incrementally per price segment or reset cost whenever the price changes.

Useful? React with 👍 / 👎.

cursoragent and others added 2 commits September 13, 2026 10:29

cursor Bot changed the title ~~Group rankings by model line and clear output on family change~~ Group rankings by model line, clear on family change, fetch source URL Sep 13, 2026

cursoragent and others added 2 commits September 13, 2026 11:03

fix: keep a trailing newline on fetched source text

cca3ff8

Stops the token loop from gluing the last word of a page onto the next repeat. Notes browser verification of grouping and line-change clear. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

cursor Bot changed the title ~~Group rankings by model line, clear on family change, fetch source URL~~ Group rankings by model line, fetch source URLs, show output cost Sep 13, 2026

jammyfu marked this pull request as ready for review September 13, 2026 11:13

fix: show hourly burn rate next to per-second and per-minute cost

c7dcc50

Single-model stats now read $/s · $/min · $/h from active TPS × AA output list price. Compare mode still shows A/B per-second rates. Co-authored-by: PaintingCoder <jammyfu@users.noreply.github.com>

chatgpt-codex-connector Bot reviewed Sep 13, 2026

View reviewed changes

Navigation Menu

Repository navigation

Group rankings by model line, fetch source URLs, show output cost - #3

Group rankings by model line, fetch source URLs, show output cost#3
jammyfu wants to merge 5 commits into
mainfrom
cursor/family-grouped-rankings-0e4b

jammyfu commented Sep 13, 2026 •

edited by cursor Bot

Loading

Uh oh!

chatgpt-codex-connector Bot commented Sep 13, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		const upstream = await fetch(target.href, {
		redirect: 'follow',

		const spentA = cumulativeCostUsd(streamA.tokensCount, selected?.outputPricePerMillionUsd);
		const spentB = cumulativeCostUsd(streamB.tokensCount, compare?.outputPricePerMillionUsd);

Conversation

jammyfu commented Sep 13, 2026 • edited by cursor Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

How grouping + clear work

How source / URL fetch works

How cost works

Verification

Uh oh!

chatgpt-codex-connector Bot commented Sep 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Sep 13, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

jammyfu commented Sep 13, 2026 •

edited by cursor Bot

Loading

chatgpt-codex-connector Bot commented Sep 13, 2026 •

edited

Loading