xAI Ships Grok 4.6, Matching GPT-5.6 on Benchmarks at Half the Price of Rival Frontier Models

xAI released Grok 4.6 on August 12, and the number that matters is 61 — its score on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max at a fraction of the price. Standard pricing is $2 per million input tokens and $6 per million output tokens; requests above 200,000 tokens jump to $4/$12. A faster variant costs double. Either way, it undercuts both GPT-5.6 and Claude Opus on cost while matching them on the composite benchmark xAI’s rivals use to sell frontier positioning.

The bigger jump is on SWE-bench Verified, where Grok 4.6 hits 95.60% — up roughly 9 points from Grok 4.5 and the strongest verified coding result xAI has published. xAI built the release around long-running agentic work: staying coherent across many-step tasks like researching a topic end to end, navigating an unfamiliar codebase, or turning a rough idea into a working application without losing the thread. The model accepts text and image input, carries a 500,000-token context window, and ships with function calling, structured outputs, web search, X search, and code execution baked into the API.

That context window matters more than the headline benchmark for anyone building agents. Most coding-agent failures aren’t reasoning failures — they’re context failures, where the model loses track of a file it edited three tool calls ago. 500K tokens gives Grok 4.6 room to hold a mid-size codebase plus its own scratchpad without aggressive summarization, which is the same bet Anthropic made with Claude’s extended context and OpenAI made with GPT-5.6’s long-context tier.

The pricing move is the more consequential story for the market. Grok 4.6 landing at $2/$6 while matching a $2/$6-and-up field on intelligence score continues a price war that’s been running all summer — OpenAI cut GPT-5.6 Luna’s pricing 80% to $0.20/$1.20 per million tokens in early August, and DeepSeek shipped V4-Flash at $0.14/$0.28 the same week. Frontier capability is no longer the differentiator it was a year ago; every lab in the top tier now clusters within a few points of each other on Artificial Analysis, which means price and agent reliability are becoming the actual battleground.

For developers already building on Grok, 4.6 is a straightforward upgrade — same API surface, meaningfully better SWE-bench, no reason to hold back. For teams choosing a model cold, it’s now genuinely hard to justify picking a more expensive frontier model on capability grounds alone. The gap that used to exist between “the smart model” and “the cheap model” has mostly closed.

Sources

Related reading