Grok 4.7 Launches at Chinese-Model Prices — but the Benchmarks Show a 15-Point Gap to Claude and GPT-6
Agentic AIxAI is pricing Grok 4.7 like a Chinese model — $2 per million input tokens, $6 per million output — but the benchmarks situate it where the pricing suggests: below the Western frontier, not competitive with it. The question isn’t whether Grok 4.7 is cheap. It is. The question is whether cheap at this performance tier is useful.
What Happened
Grok 4.7 launched September 21 via the Grok API, Cursor, and Grok Build. xAI’s framing: a larger base model, longer reinforcement learning, and improved self-verification capabilities over the 4.3 generation.
The independent benchmarks tell a specific story. On the Artificial Analysis Intelligence Index (v4.3.2), Grok 4.7 scores 46. Claude Fable 5.1 and GPT-6 Astra both score 53 — a 15-point gap. On Terminal-Bench 4.0, which measures agentic coding performance, Grok 4.7 achieves 26%. GPT-6 Astra hits 60%. Claude Fable 5.1 hits 55%. That’s not a marginal underperformance; it’s a different capability tier for agentic coding specifically.
Why It Matters
The $2/$6 input/output pricing is significant context. Western frontier models have been pricing at $5-15/M input for comparable positioning claims. xAI is pricing Grok 4.7 closer to Chinese lab rates — Qwen, DeepSeek — which have used aggressive pricing as a market entry strategy. That pricing works when the benchmark gap is small. When the gap is 34 points on agentic coding, the “bargain price” framing becomes a different conversation.
This matters for the broader pricing environment. The same week that OpenAI cut GPT-6 Sol and Luna prices in half and Anthropic dropped Opus 5.5 pricing 20%, xAI is entering the market at significantly lower price points for a model that benchmarks significantly lower. The price compression being driven by competitive pressure from the top is structurally different from the pricing xAI is using to build API volume from the bottom.
What This Means for Builders
Two use cases where Grok 4.7’s pricing makes sense: high-volume, lower-stakes inference tasks where $2/M input is meaningfully cheaper than alternatives at comparable quality, and applications where Grok’s distinctive training characteristics (which differ from Anthropic’s and OpenAI’s alignment approaches) are a feature rather than a neutral factor.
For agentic coding use cases specifically — orchestrating multi-step code generation, debugging loops, test execution pipelines — the Terminal-Bench gap is a practical constraint. A 26% task completion rate on a benchmark that GPT-6 Astra handles at 60% means Grok 4.7 should not be in the default path for agentic developer tooling at current capability levels.
The So What
Grok 4.7’s pricing is real competition for volume workloads where the performance gap doesn’t close on the tasks at hand. For agentic systems — which is where the frontier model market is actually moving — the benchmark gap is large enough to be a disqualifier rather than a tradeoff to manage. xAI’s bet appears to be on volume and API ecosystem growth (Cursor integration, Grok Build) rather than frontier capability parity. Whether that’s a viable path depends on whether price-sensitive developers find enough use cases above 26% agentic coding performance to build on.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →