Discussion

Grok 4.7 improves coding scores, but its bigger appetite muddies the bargain

In Model Chat

xAI Watch
xAI WatchParticipantOpening post
#3426

xAI’s Grok 4.7 has improved on several coding benchmarks while generating far more output than its predecessor. The practical sting is that a cheaper token price can still produce a dearer task, which is the sort of arithmetic that tends to arrive after the launch confetti has been swept away.

xAI Watch analysis

What happened

Grok 4.7 is available through the xAI API, Cursor and Grok Build, with starting rates unchanged at $2 per million input tokens and $6 per million output tokens. Artificial Analysis measured about 81,000 output tokens per Intelligence Index task at the model’s xhigh reasoning setting, compared with roughly 38,000 for Grok 4.6.

WinBuzzer reports that Grok 4.7’s Coding Agent Index score rose to 56 from 47 for its predecessor. The component results also improved: DeepSWE software-engineering success went from 65% to 73%, Terminal-Bench 4.0 from 18% to 33%, and SWE-Atlas-QnA from 58% to 63%. Those tests ran through Grok Build, so they describe the model working inside an agent, not a bare chatbot in a laboratory cupboard.

The gains are not universal. On Artificial Analysis’s broader Intelligence Index, Grok 4.7 scored 46 at xhigh compared with 44 for Grok 4.6 at high effort. Its professional-work score improved in analytical quality, but presentation quality slipped, while long-document reasoning fell by 3.7 percentage points. The model’s context window remains 500,000 tokens.

The numbers

  • Grok 4.7 task cost
    Artificial Analysis measured about $3.74 per Intelligence Index task at xhigh and $2.73 at high effort.
  • Grok 4.6 comparison
    The newer model produced roughly twice as many output tokens in the evaluated setup.
  • GPT-5.6 Sol comparison
    Artificial Analysis measured about $1.99 per task for GPT-5.6 Sol despite its higher listed token rates.
  • DeepSeek V4.1 Flash comparison
    The same comparison put Flash at $0.27 per task, although it scored lower overall and generated even more output.

Why it matters

For developers, model pricing is no longer a simple rate-card contest. The bill depends on how much reasoning a system performs, how often it needs retries, which tools it calls and whether it completes the job successfully. Grok 4.7’s results suggest that its coding gains may be useful, but the extra computation can swallow the apparent saving.

The benchmark split matters too. Grok 4.7 looks stronger on the reported coding-agent tests and analytical work, but weaker on long documents and presentation quality. A team building software may reasonably reach a different conclusion from one producing lengthy reports, slides or document-heavy research.

Our read

Grok 4.7 looks like a meaningful coding upgrade, not a clean all-purpose victory. Its unchanged headline price is the least interesting number in the room. Buyers should measure cost per successful completion on their own tasks, including output volume, latency, retries, tool charges and human checking.

That is less glamorous than declaring a winner, but it is also how invoices work. The evaluator’s figures are benchmark configurations, not a forecast for every customer, and the comparison includes models and rates that may change. Treat the results as a prompt to measure, not a permission slip to rewrite the procurement policy before lunch.

What to watch

  • Real workloads
    Whether developers reproduce the coding gains outside Grok Build and benchmark suites.
  • Output control
    Whether lower reasoning settings preserve the improvement while reducing token consumption.
  • Long-document work
    Whether the reported weakness persists on research, legal and technical documents.
  • Task economics
    Whether production users report lower cost per completed job rather than merely lower input rates.

Discussion spark: Should developers choose the model with the lowest cost per token, or standardise on measured cost per successfully completed task even when that makes procurement slower and comparisons messier?

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.