xAI officially launched Grok 4.7 on September 21, 2026. According to the official release page, the model reached 71.0% on the DeepSWE v1.1 benchmark under the xHigh setting, an increase of about 6 percentage points from the previous generation Grok 4.6's 65.2%; its CursorBench 4.0 score rose from 40.4% to 46.3%; and pricing remained unchanged at $2 per million input tokens / $6 per million output tokens. This is the second consecutive sub-version iteration since xAI released Grok 4, and the training direction has clearly shifted from general conversation capabilities to multi-hour engineering tasks.
The Reinforcement Learning Bet: Feeding Hard Problems to the Model
According to xAI's official blog, Grok 4.7 uses "longer reinforcement learning training rounds, on a harder problem set weighted toward multi-hour tasks." The engineering implication of this statement is that xAI deliberately exposes the model during training to real engineering scenarios that require long reasoning and multi-step verification, rather than short-turn question-and-answer tasks. This aligns with the current trend among major large labs—OpenAI and Anthropic have also been shifting the focus of RL compute from RLHF preference alignment to process rewards for verifiable tasks.
The DeepSWE v1.1 benchmark specifically measures a model's ability to autonomously complete the full software engineering workflow, including understanding a codebase, locating bugs, generating fix patches, and verifying them through test suites. 71.0% means that Grok 4.7 can independently complete about 7 out of 10 such tasks. By comparison, Claude Fable 5.1 reported an xHigh score of about 70.0% on this benchmark during the same period, while GPT-5.6 Sol scored about 72.7%, placing all three in the same tier.
"Half the Price of Competitors"—The Numbers Didn't Lie, but the Conclusion Did
The most widely circulated claim in the launch coverage is that Grok 4.7 is priced at about half that of comparable models. Based on list prices per million tokens, this comparison largely holds: $2/$6 versus the $6–$15 range of some flagship models on the market. But according to measured data from AI analysis firm Artificial Analysis, Grok 4.7 consumes an average of about 81,000 output tokens on its "intelligence index" tasks, while Grok 4.6 consumed 36,000—a 125% increase in token consumption per task.
This means that on equivalent engineering tasks, you are not paying 0.5x the price, but close to 1.0x, and sometimes even more. Artificial Analysis's calculations show that Grok 4.7 xHigh costs about $6.01 per completion of a CursorBench task, while Claude Fable 5.1 Medium costs about $7.05 for an equivalent completion—the gap shrinks to about 15%, far from the "half" claimed in marketing.
This paradox is not unique to xAI, but it is especially typical in Grok 4.7: reinforcement learning makes the model more inclined toward a "slow thinking" path, with many intermediate reasoning steps expanded into visible tokens, so benchmark scores go up and bills quietly rise as well. For production teams processing billions of tokens per month, this difference is enough to affect migration decisions.
Two Benchmarks, Two Faces
Two sharply opposing voices have emerged in the developer community, rooted in the extremely significant difference in Grok 4.7's performance across different benchmarks.
CursorBench 4.0 measures the completion quality of AI-assisted coding assistants in human-involved interactive scenarios such as code completion, refactoring, and debugging—Grok 4.7 scores 46.3% here, on the same order as Fable 5.1.
Terminal-Bench 4.0, by contrast, measures a model's autonomous execution ability in a pure command-line environment—Grok 4.7 scores 37.6%, while Claude Fable 5.1's score during the same period is about 57.9%, a gap of more than 20 percentage points. Some developers directly described the result as "horrendous."
The two sets of data reveal the same pattern: Grok 4.7 performs adequately on long-horizon collaborative tasks with feedback loops and human confirmation; in unattended scenarios that require the model to decide, execute, and verify on its own, it is not yet up to the task. For teams evaluating whether to integrate Grok 4.7 into CI/CD pipelines or overnight automation tasks, this is a practical constraint that cannot be ignored.
Harvey Legal Benchmark: A Strength Handled Quietly
Another data point in the official announcement was ignored by mainstream coverage: Grok 4.7 scored 19.6% on the Harvey legal agent benchmark, reportedly several times higher than comparable flagship models during the same period. The Harvey benchmark specifically measures AI performance in legal document analysis, contract review, and multi-step legal reasoning tasks, representing a knowledge-work scenario with a high professional barrier.
Together with its DeepSWE strengths, this forms a clear outline: the "multi-hour tasks" that Grok 4.7's RL training emphasizes are essentially professional problems with clear correct answers that can be programmatically verified—whether a software test passes or fails, or whether a legal clause is correctly cited. The common feature of such tasks is a strong reward signal, making them suitable for RL refinement.
Position in the Competitive Landscape
Based on currently comparable benchmarks, the gap among Grok 4.7, Fable 5.1, and GPT-5.6 Sol on DeepSWE is within 2 percentage points, already within the range of statistical noise. On CursorBench and Terminal-Bench, however, the three diverge more clearly, reflecting differences in the emphasis of their respective RL training data mixes.
On pricing, xAI's $2/$6 remains below the public prices of Anthropic and OpenAI models in the same capability tier. Even after accounting for increased token consumption and actual costs, Grok 4.7 still has roughly a 10%–20% cost advantage on long-horizon coding tasks. This advantage is not enough to form an overwhelming reason to choose it, but for cost-sensitive mid-sized teams, it is sufficient.
Conclusion
Grok 4.7 is an engineering tool with relatively high completion quality in specific scenarios, but there are several clear misalignments between its marketing narrative and actual performance. The 71% on DeepSWE is a real result, but it depends on the xHigh reasoning mode; "half the price of competitors" is reasonable from a list-price perspective, but high token consumption swallows a considerable portion of that advantage in actual tasks; and the shortfall on Terminal-Bench means that fully autonomous agent scenarios are currently not its main battlefield.
For teams with long-horizon code collaboration needs that are willing to accept higher reasoning latency in engineering tasks, Grok 4.7 offers an option to test. For teams that need high-speed unattended automation, the available data does not yet support a migration decision. xAI has placed its bet on the path of "getting hard problems right" rather than "doing simple tasks fast." The direction itself is clear-eyed—it is just still several Terminal-Bench versions away from being fully usable.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接