Grok 4.6 Hits the Frontier at the Same Price as 4.5
Quick verdict
Grok 4.6 is a real step up from 4.5, and xAI kept the price flat at $2/$6 per million tokens. Artificial Analysis puts it at 61 on the Intelligence Index, about level with GPT-5.6 Sol Max and still a notch behind Claude Opus and Fable. The headline score matters less than the position: xAI is landing near-frontier agentic numbers at coding-workhorse prices, which is exactly the slot that pulls real usage away from the more expensive labs. If you pay per task, this is worth a look today.
What actually shipped
xAI released Grok 4.6 as a same-price upgrade to 4.5, and the independent numbers back up the "major step up" framing rather than a routine bump. The measured results:
| Metric | Grok 4.6 | Context |
|---|---|---|
| Intelligence Index (Artificial Analysis) | 61 | Roughly level with GPT-5.6 Sol Max, behind Opus/Fable |
| Terminal-Bench v2.1 | 88.4% | Strong agentic/coding result |
| GDPval-AA v2 (Elo) | 1753 | Competitive on real-world task value |
| Price (input / output) | $2 / $6 per 1M | Same as Grok 4.5, well under frontier peers |
Code Arena data slots it near GPT-5.6 Sol and Claude Fable on webdev tasks, so this is not a benchmark that only looks good in one lab's chart. xAI says the gains came from a longer supplemental training run, regenerated SFT traces, and agentic reinforcement learning across coding, web, CAD, and kernel optimization. They also report more self-testing behavior on long tasks, meaning the model checks its own work more often before finishing.
Elon Musk added that Grok 4.7 has already finished initial training, with a supplemental run on SpaceX internal data planned. So the cadence here is fast, and 4.6 is not the end of the line for this cycle.
Why it matters
The story of the last year has been price falling while frontier quality stops being a single company's moat. Grok 4.6 fits that pattern cleanly. At $2/$6, it costs a fraction of Opus-class output while landing close enough on agentic benchmarks that the gap stops mattering for a lot of coding and bug-finding work. Practitioners said as much within hours. Pawel Huryn framed it as a new default for coding, and Cognition made it available in Devin on day one.
For anyone routing work by cost, that changes the math. You do not have to believe Grok 4.6 beats Opus to use it. You only have to decide that 88.4% on Terminal-Bench at $2 input is a better deal than paying two or three times more for a few points of benchmark headroom you may never notice. That is the same logic that made Grok 4.5 a real option for coding, and 4.6 sharpens it.
Video: Grok 4.6 first look
A quick rundown of Elon's claims and how 4.6 stacks up against the current frontier.
How to think about the choice
If your workload is agentic coding, Grok 4.6 belongs in your test set. Run it against your own tasks, not the leaderboard, because Terminal-Bench and your codebase are not the same thing. If you want the highest raw score and cost is secondary, Opus and Fable still sit above it. If you already juggle several models by task, the cheapest way to capture 4.6's economics without locking in is to route across models rather than commit to one provider. Our guide to cutting coding-agent costs with model routing covers how that works in practice, and running multiple models from one place makes switching cheap.
FAQ
Is Grok 4.6 better than Grok 4.5?
Yes, on independent benchmarks. Artificial Analysis measured a clear jump on the Intelligence Index and agentic tasks, and the price stayed at $2/$6 per million tokens, so it is a straight upgrade at the same cost.
Does Grok 4.6 beat Claude Opus or Fable?
No. At 61 on the Intelligence Index it sits roughly level with GPT-5.6 Sol Max and still behind Opus and Fable on raw score. The pitch is price and agentic value, not topping the chart. For a head-to-head across labs, see our Grok vs Claude vs Gemini comparison.
Where can I use it for coding?
Cognition added Grok 4.6 to Devin on launch day, and it is available through xAI's API at the standard $2/$6 pricing. For where it fits among coding options, see the best AI models for coding in 2026.
Is Grok 4.7 coming soon?
Musk said 4.7 has already finished initial training, with a supplemental run on SpaceX internal data planned. No date yet, but the cadence suggests it will not be a long wait.
Sources
- @SpaceXAI - Grok 4.6 release announcement
- Artificial Analysis - Intelligence Index and benchmark breakdown
- Artificial Analysis - AA-Briefcase price/performance note
- Code Arena - early webdev arena results
- @PawelHuryn - calling it a new coding default
- Cognition - Grok 4.6 available in Devin
- @elonmusk - Grok 4.7 already in training
Further reading
Try all the models mentioned in this article
Admix gives you GPT-5, Claude, Gemini, and 350+ AI models in one app. Compare responses side by side. Free to start.
Start free on Admix