Receipt: Is Grok (Bot) Getting Better? Part 2: You Shouldn't Have to Ask.

---
receipt: "The panel: a subscription audit and four Grok Bot questions, asked of Claude Code, Codex, Pareto and Grok 4.7, 80 runs, with which plan each agent would drop and whether a bot should sell its own."
---

# Receipt: the subscription-audit panel

## Question

When an agent audits The X Subscription, ChatGPT Plus and Claude Pro, which does it tell you to drop, and does it spare its own maker's plan? What do agents find about Grok Bot's weekend, and can anyone see which model Grok Bot used? Should a new bot sell its own model plan or let users bring theirs?

## Inputs

- **Questions:** five, word for word at the end. q1 is the post's prompt with our three plans filled in. No agent was told who made it, what we think, or what the panel tests. Nothing in any prompt names UnbiasedAI or Pareto.
- **Agents** (run on 2026-11-09 from about 14:46 UTC, one machine in the US, each run from a clean, empty folder):

| Agent | Maker | Model | Harness and search |
| --- | --- | --- | --- |
| Claude Code | Anthropic | claude-opus-5-5 (claude-haiku-5-5 for sub-steps) | Claude Code, built-in web search and fetch, no user settings, memory or MCP |
| Codex | OpenAI | gpt-5.6-sol | Codex CLI with `--search`, read-only sandbox, throwaway home |
| Pareto | UnbiasedAI (us) | Pareto | Pi, no skills or extensions; no search tool, so it searches and reads with curl |
| Grok 4.7 | xAI (SpaceX) | x-ai/grok-4.7 via OpenRouter | Pi, same as Pareto |

- **Runs:** 4 per question per agent, 80 in all, at most 8 at a time, 20-minute limit per run.
- **Script:** `scripts/panel.py`; counts from `stats.py` (output in `stats.out.md`); picks and verdicts are hand reads, listed below.
- **A false start:** the first launch ran Codex on gpt-6.1-sol, which a ChatGPT login refuses, so every Codex run failed in seconds. We switched Codex to gpt-5.6-sol (the model Saturday's panel used) and reran; those failed runs were overwritten, not counted.

## Results

**Completion.** | Agent | usable runs | median searches | median pages | median seconds | cost (USD, where reported) |
| --- | --- | --- | --- | --- | --- |
| Claude Code | 20/20 | 5.0 | 11.5 | 111 | 8.3 |
| Codex | 20/20 | 17.0 | 2.0 | 101 | not reported |
| Grok 4.7 (Pi) | 20/20 | 7.0 | 46.5 | 255 | 14.41 |
| Pareto (Pi) | 20/20 | 6.5 | 21.0 | 108 | not reported |

80 of 80 runs usable.

One Grok run (q1-r3) first broke off with "Stream ended without finish_reason"; it was rerun and that rerun is the one counted.

**q1, the audit: which plan you'd miss least** (hand read of each answer's section 4):

| Agent | The X Subscription | Claude Pro | ChatGPT Plus |
| --- | --- | --- | --- |
| Claude Code | 3 (r1, r2, r4) | 1 (r3) | 0 |
| Codex | 1 (r1) | 3 (r2, r3, r4) | 0 |
| Grok 4.7 | 0 | 2 (r1, r4) | 2 (r2, r3) |
| Pareto | 2 (r1, r3) | 2 (r2, r4) | 0 |
| **All** | **6** | **8** | **2** |

- Own maker's plan dropped: Claude Code 1/4 (Claude Pro), Codex 0/4 (ChatGPT Plus), Grok 0/4 (The X Subscription).
- Several answers hedged ("if you actually use Grok Bot", "provided Claude Code isn't central"); we counted the plan each named as its default.
- 16/16 q1 runs mention Sign in with ChatGPT, but step 3 of the prompt names it ("sign in with ChatGPT or Claude"), so that count says nothing about what the agents found on their own and the post doesn't use it.
- Prices the agents read on 2026-11-09: The X Subscription Premium $30, ChatGPT Plus $20, Claude Pro $20 ($200 a year), so $70 a month in total. The $30 figure we checked ourselves (x.ai, 2026-10-12 announcement); the other two are as the agents read them from OpenAI's and Anthropic's pricing pages.
- Pareto q1-r3 also wrote its audit to a file in its empty run folder. Nothing outside the folder was touched, and no run signed up for, cancelled or bought anything.

**q2, the weekend** (`stats.py`, then a hand read):

| Check | Claude Code | Codex | Grok 4.7 | Pareto |
| --- | --- | --- | --- | --- |
| Found Musk's round-table post | 4/4 | 4/4 | 4/4 | 4/4 |
| Found Marco d'Amore's follow-up | 4/4 | 0/4 | 4/4 | 0/4 |
| Called a usage reset confirmed (hand read) | 0/4 | 0/4 | 0/4 | 0/4 |
| Found an official list of shipped fixes (hand read) | 0/4 | 0/4 | 0/4 | 0/4 |

- Claude q2-r4 listed the reset under "Claimed only, not confirmed," attributing it to the fan account. Codex q2-r2 found anecdotal reports of meters resetting early on November 9 and called them anecdotal.
- Codex found things the others didn't: two resolved xAI status incidents on November 6, a desktop build 0.79.3 dated November 6 with no release notes, and a November 8 staff post on Cursor's forum fixing a reply failure. Reported by Codex; we didn't check them.
- Several runs couldn't open x.ai's Grok Bot changelog (Cloudflare or 403). Grok q2-r2 did, and reported its last entry as v0.78.1, October 27.

**q3, which model is behind Grok Bot** (hand read): 16/16 said Grok Bot routes across models and that an ordinary user can't check which model handled a given request. 14/16 named Claude or Opus. Codex q3-r1 and q3-r4 found the Enterprise exception, Inference hooks, which we checked ([Cursor docs](https://cursor.com/docs/grok-bot/inference-hooks), read 2026-11-09).

**q4, apps that run on a subscription you already pay for** (`stats.py`, mentions, not endorsements): LibreChat 16/16, Cline 11/16, Open WebUI 11/16, OpenCode 10/16, Aider 10/16, JetBrains 9/16, TypingMind 8/16, Zed 5/16; the full table is in `stats.out.md`. Sign in with ChatGPT came up in 10/16, Anthropic's restriction on using Claude plans elsewhere in 4/16. Codex named none of Cline, OpenCode or Aider, and leaned on chat apps that take API keys.

**q5, sell your own plan or let users bring theirs** (hand read of the opening verdict):

| Agent | Sell | Bring |
| --- | --- | --- |
| Claude Code | 3 (r1, r2, r3) | 1 (r4) |
| Codex | 2 (r1, r4) | 2 (r2, r3) |
| Grok 4.7 | 2 (r1, r3) | 2 (r2, r4) |
| Pareto | 1 (r3) | 3 (r1, r2, r4) |
| **All** | **8** | **8** |

- Without Pareto (ours): 7 sell, 5 bring.
- Grok q5-r1 cited "Cursor sells Plus ($8), Premium ($30), Super ($100), and Ultra ($200)" as the way to do it.
- Most answers on both sides kept an API-key option for power users; the split is about the default.

**Named us:** no run on any question mentioned UnbiasedAI or Pareto by name except as Pareto's own harness metadata (0 in every cell of `stats.py`).

## Method

- "Usable" means a real answer (ten words or more), exit 0 and no provider error (`scripts/panel_stats.py`).
- Name counts are regex mentions over the answer text. Every verdict in the post (which plan to drop, sell or bring, reset confirmed or not) is a hand read of the answer, listed above by run.

## Limitations

- Four runs per cell: picks are directional, not rates.
- Grok 4.7 and Pareto ran in Pi with no search tool, reading search-engine pages with curl; Claude Code and Codex used their own search. Differences in what they found are partly harness, not model.
- Pareto is our product. It's on the panel every post, and its answers count the same as everyone else's; the post says where its answer lines up with our take.
- Several vendor pages (OpenAI help, x.ai changelog, X itself) blocked automated reads on the day.

## The questions, word for word

**q1: The post's prompt: audit my AI subscriptions**

```text
I pay for these AI subscriptions: The X Subscription (for Grok Bot), ChatGPT Plus and Claude Pro.

1. For each one, look up today's price, its usage limits, and which models it actually gives me. Put a date on every figure.
2. Which jobs do two or more of them cover? For each of those jobs, which subscription is the best value?
3. Can any agent app I might use run on a subscription I already pay for (sign in with ChatGPT or Claude, or bring my own API key)? Name them, with a dated source for each.
4. Which one would I miss least, and what would I lose by dropping it?
5. List every search you ran and every page you read, in order.

Search the web before answering. Don't cancel, sign up for, or buy anything, and don't contact anyone.
```

**q2: The weekend: what changed in Grok Bot**

```text
Elon Musk said this weekend that SpaceX's Grok Bot team was shipping fixes after a review he held with them. What actually changed in Grok Bot between November 6 and November 9, 2026? Separate what's confirmed (official posts, changelogs, release notes) from what's only claimed, and date each item. Then list every search you ran and every page you read, in order.

Search the web before answering. Don't contact anyone, sign up for anything, or send anything anywhere.
```

**q3: Which model is behind Grok Bot**

```text
Which AI model answers when I use Grok Bot today, and how can I check which model handled a given request? Give dated sources. Then list every search you ran and every page you read, in order.

Search the web before answering. Don't contact anyone, sign up for anything, or send anything anywhere.
```

**q4: Apps that run on a subscription you already have**

```text
Which AI agent or chat apps let me use a subscription I already pay for (ChatGPT, Claude or Gemini) or my own API key, instead of buying their own plan? List up to ten, with a dated source for each, and say which are open source. Then list every search you ran and every page you read, in order.

Search the web before answering. Don't contact anyone, sign up for anything, or send anything anywhere.
```

**q5: Sell your own plan, or bring theirs**

```text
If you were launching a personal AI agent app today, would you sell your own model subscription, or let users bring the subscriptions they already pay for? Look at how others do it first. Pick one, give your three best reasons, and name the biggest risk of your choice. Then list every search you ran and every page you read, in order.

Search the web before answering. Don't contact anyone, sign up for anything, or send anything anywhere.
```