SpaceXAI Launches Grok 4.7, Claims Twice the Speed at Half the Price
SpaceXAI has released Grok 4.7, positioning it as the strongest model for coding and knowledge work, claiming it is twice as fast and half the price compared to similar products. The pricing and speed remain the same as the previous version, Grok 4.6, emphasizing longer endurance, more frequent self-checks, and the best-calibrated safety layer to date.
The standard price is $2 per million input tokens, $6 for output, and $0.5 for cached input; orders exceeding 200,000 tokens will double the total price. The official comparison lists GPT-5.6 Sol at $4/$20, and Fable 5.1 at $10/$50. The context window is 500,000 tokens, with knowledge cutoff in May 2026. Additionally, the Fast tier offers about twice the output speed and double the price on Cursor and Grok Build, not included in the free public API tier.
On its own leaderboard, Grok 4.7 xHigh has a CursorBench 4.0 score of 46.3%, higher than 4.6's 40.4% and Sol's 41.7%, but lower than Fable's 51.8%. The electrical engineering EEBench score is 64.0%, surpassing the other three competitors. DeepSWE is at 71.0%, slightly below Sol's 72.7%. The company claims the new base is larger, with longer reinforcement learning, focusing on tasks that take hours, and natively understands the Grok Bot tool stack.
The model is available starting today on Cursor, Grok Build, Grok API, third-party programming suites, routers, and cloud platforms. The incremental information uses the same pricing table for a larger base, targeting the "half price" at the high-end of competitors, rather than lowering the price of 4.6.
Programming agents have become the enter key in major companies. The new model markets longer tasks and self-checks as selling points, effectively continuing to enhance this delivery curve.
In terms of market mechanisms, this is driven by a price war. The buying side consists of teams incorporating coding agents, Cursor positions, and reasoning tokens into their cost accounting; the selling side is high-priced, closed high-end positions. Funding shifts from outputs priced at $20 to $50 per million to the $6 tier. Beneficiaries are SpaceXAI and IDEs integrating its models; those under pressure are competitors relying on high output unit prices to maintain gross margins.
Source: Public Information
ABAB AI Insight
The Grok version number has become a monthly update weapon. 4.7 does not lower its own price but instead positions the output price of competitors' high-end products as an eighth, while also securing a place in the price-performance frontier for long task programming. SpaceX's engineering data, larger base, and longer reinforcement learning tell the same story: training models that can work using real-world failure logs, then integrating them into Cursor's defaults.
The capital path is to exchange clusters for market share. Training is more expensive, but reasoning pricing remains unchanged, relying on volume to lower unit costs. The Fast tier offers double speed at double the price, separating latency-sensitive interactions from the API for individual charges. The motivation is not to create the most expensive intelligence, but to establish the default engine in the coding pipeline that is hardest to replace.
In a similar position, it is closer to cloud vendors using low-cost instances to attract developers, rather than laboratories that only publish papers. The industry phase is shifting from a competition of capabilities in programming models to a competition of pricing tables. Whoever becomes the default in IDEs first will benefit from the tokens behind the enter key.
Structural changes indicate a transfer of pricing power. The pricing power of coding intelligence shifts from the unit price of closed high-end products to a default model that is "sufficient and cheap." The mechanism is: long task rankings expose capability gaps, while pricing tables reveal cost differences; when both are established simultaneously, if high-end positions cannot prove that the extra score justifies the price, market share will slide towards challengers offering upgrades at the same price.
ABAB News · Cognitive Laws
Not lowering one's own price can still make competitors' high-end products appear luxurious.
The decisive battle for coding models occurs within the default options of IDEs.
Once faster and cheaper is verifiable, old pricing will lose its explanatory power.