Grok 4.7 is out at $2/$6 per million tokens - cheap, but well behind on agent tasks

xAI released Grok 4.7 on 21 September, priced at $2 per million input tokens and $6 per million output tokens, according to The Decoder. That is the same list price as Grok 4.6, which shipped on 12 August. The model was late: Elon Musk had said on 2 September it would arrive in ten days, then on 11 September said it needed a few more days of tuning.

It is available through the Grok API, Cursor and Grok Build.

What is actually new

Less than the price suggests. The Decoder reports that Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index (v4.3.2), against 53 each for Claude Fable 5.1 and GPT-6 Astra. On Terminal-Bench 4.0, an agentic coding test, it scores 26%. GPT-6 Astra gets 60% and Claude Fable 5.1 gets 55%. The Decoder says it also trails DeepSeek V4.1 Flash on that test.

Two things to be careful about. First, do not compare that 46 with the 61 that Grok 4.6 was reported to score earlier. Those numbers come from different versions of the index, and the scale changed. Second, context window, cache pricing and parameter count were not stated in the coverage we found. Musk described a 2.1 trillion parameter model in July, but that is a founder claim, not a published spec, and we are not repeating it as fact. We could not load xAI's own news page (it returned a 403), so this post relies on The Decoder's report of the launch.

What it means for a business owner

The price is the story, and it is a narrow one. At $2 and $6, Grok 4.7 costs less than the frontier models it is being compared with. The Decoder's own read is that these rates sit closer to Chinese models than Western frontier ones, and that the benchmark gap is probably the reason.

That points to where a model like this fits. High-volume, low-stakes work is the natural home: sorting inbound email, tagging support tickets, pulling fields out of invoices, drafting first-pass replies that a person reviews. In those jobs a 7-point gap on a general intelligence index rarely decides whether the workflow succeeds. The cost per thousand runs does.

It is a poor fit for anything where the model has to take many steps on its own and get each one right. A 26% score on an agentic coding benchmark, against 55% to 60% for the leaders, is the number to take seriously. Agent workflows fail by compounding: a model that is wrong slightly more often at each step is wrong much more often across ten steps. If your automation books meetings, updates records in a CRM, or touches money, cheap tokens will not make up for extra retries and manual cleanup.

The honest caveat

Cheap per token is not the same as cheap per completed task. If a weaker model needs two attempts, or a human has to fix one output in five, the saving shrinks or disappears. Benchmarks also measure coding and reasoning tasks, not your invoices. A model that scores badly on Terminal-Bench could still do fine on a narrow extraction job, and a top scorer could still stumble on your particular data.

There is also the release pattern. Grok 4.7 slipped its date, and coverage from mid-September reported that xAI had already announced a larger Grok 4.8 without a release date. If you build on a model that is being superseded on paper before it has settled, expect to swap it again.

What to do about it

Do not switch a working system because a new price appeared. Instead, pick one high-volume, low-risk task you already run, and test it properly.

1. Collect 100 real examples of that task, including the messy ones. 2. Run them through your current model and through Grok 4.7. 3. Score both by a person, on accuracy and on how many outputs needed fixing. 4. Divide total cost by correct outputs, not by tokens.

If the cheaper model holds up on your data, you have saved real money on that one task. If it does not, you spent an afternoon and learned it before a client did. Keep the model choice in a config file so changing it later is a one-line edit, not a rebuild.

That last point is the durable lesson. Model prices and rankings will keep moving. Automation that is built so the model can be swapped is the only kind that stays cheap.