Grok 4.7 climbs the coding leaderboards at unchanged pricing, and the catch is how many tokens it burns to get there

xAI released Grok 4.7 on September 21, and on paper it is a straightforward step up: a larger base model, trained with longer reinforcement learning, that the company says is better at verifying its own work. The more useful signal comes from outside xAI. Artificial Analysis, which benchmarks models independently, places Grok 4.7 in the top tier for coding and agentic work while flagging a cost that the headline token price does not show.

Where the gains landed

xAI positions 4.7 as its strongest model yet for coding and knowledge work, and the independent numbers broadly agree on direction. Artificial Analysis scores Grok 4.7 at 46 on its Intelligence Index, a two-point gain over Grok 4.6 that puts it among the top four labs but several points behind the leading Claude Fable 5.1 and GPT-6. On the tasks that matter more to developers it ranks fourth on the Coding Agent Index, behind Fable 5.1, GPT-6 Astra, and Claude Opus 5, and it gains real ground on long-running agent work: 1695 Elo on GDPval, up 90 points over 4.6, and 1657 Elo on AA-Briefcase, just behind the two Claude models. xAI’s own tables report 46.3 percent on CursorBench 4.0 and 71 percent on DeepSWE v1.1, and the independent evaluations land in the same neighborhood rather than contradicting them.

The pattern is a model that closed part of the gap to the frontier on exactly the workloads developers are pushing hardest, agentic coding and multi-step tasks, without topping any aggregate leaderboard.

The token bill under the sticker price

Grok 4.7 keeps Grok 4.6’s pricing: $2 per million input tokens and $6 per million output tokens, with cached input at $0.50, and it holds roughly the same output speed at about 188 tokens per second. Those rates sit closer to Chinese open-weight models than to Western frontier pricing, and that is the headline xAI is selling.

The independent data complicates it. Artificial Analysis measured Grok 4.7 spending about 81,000 output tokens per task, substantially more than competing models use to reach similar results. Output tokens are the expensive side of the meter, so a model that burns more of them narrows the effective cost gap its per-token price advertises. For a one-off prompt the sticker price is what you pay; for an agent looping through a long task, the token count is what shows up on the invoice. The practical read is to price Grok 4.7 against a real workload and count tokens, not to assume the cheap rate translates directly into a cheap task.

What it means for developers

For teams choosing a model to sit behind a coding agent, Grok 4.7 is a credible fourth option that is genuinely cheaper per token than Fable 5.1 or GPT-6 and now close enough on agentic coding to be worth testing. It is available through Cursor, xAI’s Grok API, third-party coding harnesses, and gateways including Vercel’s AI Gateway, so slotting it into an existing setup is mostly a config change. The caveat is that the decision is no longer a simple price comparison. The right test is cost per completed task on your own workload, where 4.7’s higher token consumption can eat into the per-token savings, alongside quality on the specific coding and agent flows you run. Teams that already standardized on Claude or GPT for coding will not find a reason here to switch defaults; teams that are cost-sensitive and running their own evals now have a frontier-adjacent option that competes on price.

What to watch

The open questions are whether Grok 4.7’s coding and agent gains reproduce on real repositories rather than benchmark suites, whether the high token consumption holds across workloads or was specific to Artificial Analysis’s task mix, and how the effective cost per task compares once teams measure it against Fable 5.1 and GPT-6 on their own pipelines. The broader pattern is that the model race is fragmenting by workload and price rather than resolving into a single leader, which leaves developers doing more of their own benchmarking. Stackmaven’s follow-up will revisit around December 21.

Sources cited

1. Grok 4.7 (xAI) x.ai
2. Benchmarking Grok 4.7 (Artificial Analysis) artificialanalysis.ai
3. xAI launches Grok 4.7 at bargain prices, but benchmarks reveal a wide gap to Claude and GPT-6 (The Decoder) the-decoder.com