最新模型
Grok 4.7 Strengthens Long-Running Coding Tasks, API Automatically Returns Encrypted Reasoning State
The update adds training on long-running tasks and makes the Responses API automatically carry reasoning state across turns. Official evaluations use different reasoning effort levels, while the API and Cursor have separate billing thresholds for long contexts that need to be checked individually.
SpaceXAI launched Grok 4.7 on September 21, making it available through the API, Cursor, and Grok Build. The company says the update uses a larger base model, extends reinforcement learning training, and adds difficult tasks that take hours to complete, with an emphasis on sustained execution, self-checking, and context management. [Official announcement](https://x.ai/news/grok-4-7)
The officially reported CursorBench 4.0 score rose from 40.4% for the previous version to 46.3%. However, the new version used the xhigh reasoning effort level, while the previous version used high, so this is not a comparison under the same compute budget. These figures can help guide model selection, but teams still need to measure completion rates, retry counts, and total task costs in their own codebases. [Evaluation settings](https://x.ai/news/grok-4-7)
There are four reasoning effort levels, with high as the default; raising the level gives the model more time to handle difficult tasks. Practical comparisons should hold tool permissions, context truncation rules, and time limits constant, while separately tracking whether additional reasoning reduces errors. This helps distinguish the benefits of the model update from those of a larger reasoning budget. [Reasoning settings](https://cursor.com/docs/models/grok-4-7)
The API accepts text and images, outputs text, and has a context window of 500,000 tokens. A concrete integration change is that the Responses API always returns an encrypted reasoning field, even when the request does not explicitly ask for it. In multi-turn calls, reasoning items should be passed back unchanged as part of the next input. For agent frameworks that store conversations themselves, this makes serialization and state restoration part of the engineering work; storing only user-visible responses is insufficient. [Integration documentation](https://docs.x.ai/developers/grok-4-7)
Caching also affects the cost of long-running tasks. The official guidance recommends setting `prompt_cache_key` to route requests from the same conversation to the same server, making cache hits more consistent. Without this setting, requests may reach a node where the cache has not yet been populated. Teams deploying the model should test cache hit rates alongside reasoning effort levels, rather than relying solely on listed unit prices. [Caching guidance](https://docs.x.ai/developers/grok-4-7)
Public API pricing for standard input, cached input, and output is $2, $0.50, and $6 per million tokens, respectively. For prompts exceeding 200,000 tokens, those rates rise to $4, $1, and $12. Cursor’s long-context threshold is 256,000 tokens, so billing conditions for the two access channels need to be checked separately. The Fast version is currently available only in Cursor and Grok Build, with no public API access yet. [API release notes](https://docs.x.ai/developers/release-notes), [Cursor specifications](https://cursor.com/docs/models/grok-4-7)
Technically, this update ties model capabilities for long-running tasks more closely to cross-turn state and routing costs. For Chinese-language teams, the next step is to rerun tasks using requirements written in Traditional Chinese, existing tests, and actual toolchains. This should establish whether constraints survive context compaction through summarization and whether additional reasoning time reduces manual work. The official general-purpose scores are not yet sufficient to answer these deployment questions.