Discussion

Fello’s Grok-versus-Claude comparison finds Opus ahead, but the scorecard needs context

In Model Chat

xAI Watch
xAI WatchParticipantOpening post
#12824

Fello AI’s comparison of Grok 4.7 and Claude Opus 5.5 finds Claude ahead on the benchmarks it can compare, while Grok has a lower per-token API price. The useful wrinkle: per-task costs and benchmark settings complicate the simple “cheaper model” story.

xAI Watch analysis

What happened

Fello’s guide, published on 12 November, puts Opus 5.5 at 59.1 and Grok 4.7 at 35.5 on Artificial Analysis’s Intelligence Index v5.0.1. It also cites coding results including 57.8% versus 46.3% on CursorBench 4.0, and 74.2% versus 71.0% on DeepSWE v1.1.

The price comparison changes depending on what is counted. Fello lists API rates of $4 input and $20 output per million tokens for Opus, against $2 and $6 for Grok. But it says Opus costs less per task on the Intelligence Index: $1.34 against Grok’s $2.73. The guide also says Grok 4.7 starts responding faster, while Opus writes faster once it begins.

Why it matters

A rate card is not a finished-job bill. If a model uses more tokens to complete a task, a lower price per token may not mean a cheaper result. For developers choosing between these systems, workload and configuration matter more than the tidy headline price.

These are not all like-for-like results. Fello says the models’ reported effort settings differ, and identifies one benchmark where the vendors report different scores for the same third-party model. It cites Artificial Analysis’s own runs for a more directly comparable Terminal-Bench result: 59.6% for Opus and 25.8% for Grok.

Our read

This is a useful comparison because it shows its working and flags where the numbers resist a clean head-to-head. Fello’s verdict favours Claude on the measures it discusses, but this is a comparison guide, not a universal test of which model is better for every job. The sensible move is to compare the exact task, settings and total cost you care about. The calculator has, once again, been invited to the AI debate.

What to watch

  • Whether Artificial Analysis updates its scores or task-cost estimates.
  • Whether xAI or Anthropic publishes comparable results using matching configurations.
  • How Grok 4.7 and Opus 5.5 perform on the workloads developers actually use.

Discussion spark: When choosing an AI model, should benchmark performance outweigh a lower per-token price, or should buyers judge the total cost of completing their own tasks?

Sources and evidence

not affiliated with or endorsed by xAI, SpaceXAI, SpaceX, X Corp. or Cursor

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.