The landscape at the top of the AI model rankings has been thinning for a while. Through most of 2025, two or three models dominated the capability benchmarks while everyone else competed for the next tier. Grok 4.7 lands firmly in that second tier — and makes a deliberate choice to compete on price rather than chase a top-three ranking it has not yet earned.
Where It Actually Places
On the Artificial Analysis Intelligence Index — one of the more comprehensive capability rankings across multiple task types — Grok 4.7 in its extra-high thinking mode scores fifth overall, behind Fable 5.1, Astra, Claude Opus 5, and Muse Spark 1.3 Max (an open-weights model, meaning one whose underlying weights are publicly released). xAI’s own public prediction heading into launch was that the model would land roughly on par with Opus 5, not 5.1. That turned out accurate.
The benchmark picture is mixed in ways that matter for real use cases. In coding, Grok 4.7 scores 71 percent on DeepSuite — considered the benchmark closest to production reality — putting it level with Fable 5.1 and just behind Astra’s 74 percent. In legal work, it leads all OpenAI, Anthropic, and xAI models at just under 20 percent, though an open-weights model scores more than double that. Where it falls furthest behind is agentic coding — the Terminal Bench 4.0 benchmark, which tests autonomous multi-step tasks (sequences of actions a model executes on its own, without step-by-step human guidance). Grok 4.7 reaches 38 percent there; Fable 5.1 and Astra both sit above 57 percent. For teams building agents or long-horizon coding workflows, that gap is not cosmetic.
One other number worth noting: Grok 4.7’s context window is 500,000 tokens — roughly half of what most frontier-tier models now offer. Context window refers to how much text a model can hold in view at once; a shorter one starts to matter on large codebases, long documents, or any task where a lot of background needs to be in scope simultaneously.
The Case for the Price
At two dollars per million input tokens and six dollars per million output tokens, Grok 4.7 costs the same as its predecessor. Fable 5.1 and Astra run about five times that. For workloads that do not require top-three capability and where inference cost at scale is a real variable, the value equation is genuinely favorable — though it is worth noting that Grok 4.7’s higher-thinking modes consume more tokens than the raw per-token rate suggests, which partially narrows the effective cost advantage.
There is a structural reason that pricing can be sustained for now: xAI over-invested in compute infrastructure before its models were driving enough demand to fill it. Excess GPU capacity allows the company to price aggressively without losing margin. Whether that holds as demand catches up to capacity is an open question.
Reading the Numbers Carefully
Two caveats on the benchmark picture are worth keeping in mind.
First, xAI now owns Cursor, the developer IDE behind the Cursorbench results that show Grok 4.7 performing comparably to Opus 5. That relationship is a reason to weight those particular numbers carefully.
Second, in xAI’s own published comparison on the DeepSuite benchmark — where there is no such conflict — the company’s table quietly omitted Astra, which happens to be the top scorer in that category. Independent analysis included Astra and put it clearly ahead.
A Market Shifting Quietly
The competitive picture extends beyond xAI. Enterprise token consumption is tilting toward open-weights models, which now account for the majority of usage by volume. At the same time, most of the economic value in the market still concentrates at OpenAI and Anthropic. Grok 4.7 occupies an interesting middle position: a closed model priced closer to open-weights rates, betting that cost sensitivity will matter more to enterprise buyers than the gap separating it from the true frontier.
The model worth keeping an eye on in that same conversation may actually be Muse Spark 1.3 Max — an open-source model that places fourth on the intelligence index and leads the legal benchmark overall. It receives little attention relative to its benchmark position, and that is the clearest signal that open-weights development is closing the gap faster than proprietary model rankings tend to suggest.
Grok 4.7 is a real and capable model. It is probably not the right choice for teams running agentic workflows or working at the edge of what models can do. For cost-sensitive workloads in the second-tier capability range, it is a credible option — and the pricing pressure it creates on the broader market may matter more in the medium term than where it ranks today.
Text summarized and optimized using Anthropic’s models and reviewed by a human.