Grok 4.7
xAI's flagship text model: the same per-token price as Grok 4.6 and a bigger base model, but it burns so many more tokens that a typical task costs twice as much.
Quality
Modality
multimodal
Context
500K tokens
Access
closed
Fabian's Take
"I'm enjoying Grok in Grok Bot. It's quite capable, and nothing else gives you such an easy read on what people on X are saying about a topic. But 4.7 looks like a bad release. On paper it's cheaper per token than Claude or GPT, and that's true, but in some cases it uses more than twice the tokens they do, so it ends up costing more."
Grok 4.7 arrived on 21 September 2026, replacing Grok 4.6 as xAI’s flagship text model. xAI built it on a larger base model and says it spends longer working through hard problems and double-checks its own answers more often. The per-token price didn’t move: $2 per million input tokens and $6 per million output, with the same 500K context window.
The price per token is the wrong number
This is the release where the token price and the real price come apart. Grok 4.7 writes a lot. Artificial Analysis measured about 81,000 output tokens per task on its Intelligence Index at the highest reasoning setting, against about 36,000 for Grok 4.6 and 27,000 for GPT-6 Astra. Same price per token, more than twice the tokens, and its cost per task roughly doubled, from $1.86 to $3.74. That’s more than GPT-6 Astra costs per task on the same test, a model whose tokens cost five to eight times as much.
Its benchmark scores did go up a little (46 against 44 on that index). The trouble is you pay for that twice over, in money and in waiting: at around seven minutes per task, it’s slow in practice even though it generates tokens quickly.
Where it still makes sense
I use Grok through Grok Bot, and there it’s a capable, pleasant model to work with. The other reason to keep it around is X: nothing else gives you that easy a read on what people are actually saying about a topic right now.
What I wouldn’t do is pick it for its token price. If you’re paying per token for agent work, measure cost per finished task. In some cases it uses more than twice the tokens that Claude Opus 5.5 or GPT-6 Sol need for the same job, which wipes out the cheaper token price.
The 200K cliff
Below 200K prompt tokens you pay $2 and $6 per million. Once a prompt reaches 200K, the whole request reprices to $4 and $12, not just the tokens above the line. Combine that with how much it writes and a long agent run gets expensive fast.
The Verdict
Best for: Running inside Grok Bot, and reading what people on X are saying about a topic right now.
Pros
- Capable all-rounder, and good to work with inside Grok Bot
- Easy access to what people on X are saying about a topic right now
- 500K context and the same per-token price as Grok 4.6
Cons
- Uses more than twice the output tokens of Grok 4.6 per task, so a typical task costs about twice as much
- By Artificial Analysis's numbers, dearer per task than GPT-6 Astra despite the much lower token price
- Crossing 200K prompt tokens doubles the price of the entire request, not just the excess
Specs
- Pricing $2/M input, $6/M output below 200K prompt tokens · $4/$12 above · cached input $0.50/M
- Cost Tier moderate
- ⚡Speed Tier moderate
- License Proprietary
