Grok 4.7 improves long-running work, but low token prices do not guarantee a low task cost

What changed in Grok 4.7

SpaceXAI says Grok 4.7 uses a larger base model than Grok 4.6 and a longer reinforcement-learning run weighted toward tasks that take hours. The company also says it improved self-verification, long-context management, and native understanding of the Grok Bot harness. These are vendor descriptions of the training and intended behavior, not independent proof that every agent workflow will improve.

The public model ID is grok-4.7. It accepts text and image input, returns text, supports function calling, web search, X search, and code execution, and offers low, medium, high, and xhigh reasoning effort. High is the default. The model is available through the xAI API and several gateways, while GitHub began rolling it out to paid Copilot plans on September 21.

Read the benchmark claims carefully

The vendor-reported table gives Grok 4.7 46.3% on CursorBench 4.0 and 38.0% on Terminal-Bench 4.0 at xhigh effort. In the same table, Claude Fable 5.1 Max scores 51.8% and 57.9% respectively. Grok 4.7 leads some listed engineering and legal evaluations, but the table mixes models, harnesses, and effort settings, so it does not establish one universal ranking.

Artificial Analysis independently scored Grok 4.7 at 46 on its Intelligence Index, two points above Grok 4.6. Grok 4.7 with Grok Build scored 56 on its Coding Agent Index, up nine points from the prior model and fourth among the tested native harnesses. The same evaluation reports approximately 81,000 output tokens per task on the Intelligence Index at xhigh, compared with 36,000 for Grok 4.6 at high effort. Better task scores can therefore come with a larger reasoning bill.

Where it may fit

• Repository work where an agent must inspect code, run tools, revise a patch, and verify the result over a long session.
• Professional document workflows that create presentations, spreadsheets, or reports and can be graded against a clear brief.
• Applications that need a large text-and-image context window and can use prompt caching across a continuing conversation.
• Teams already using Grok 4.6 that can run a controlled upgrade test without changing prompts, tools, graders, or timeouts.

Test it against completed-task cost

Build a set of at least 30 accepted tasks from your own logs and run each task several times through Grok 4.7 and your current route. Hold the harness, tool permissions, prompts, timeouts, and grader constant. Record acceptance rate, retries, input and output tokens, tool charges, latency percentiles, and human correction time.

Use the reasoning level as an experimental variable. Start with high, then reserve xhigh for tasks where it improves acceptance enough to justify the added tokens and latency. Set a prompt cache key for repeated conversations, pass encrypted reasoning items back unchanged in Responses API loops, and compact context before a long session becomes mostly history.

Pricing and operational limits

The public API starts at $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. Tool calls are billed separately, and reasoning tokens contribute to usage. A low rate per token can still produce an expensive accepted result when a model reasons at length, retries, or makes many server-side tool calls.

Grok 4.7 Fast is the same model on faster infrastructure at higher rates. Fast is not available through the public xAI API; xAI documents it for Cursor and Grok Build only. Availability, provider routing, cache behavior, and regional endpoint premiums can also change the effective cost and latency.

What remains uncertain

SpaceXAI has not disclosed the parameter count. Claims that Grok 4.7 has a specific number of parameters should be treated as unverified unless the company publishes that figure.

Benchmark success does not prove correctness on a production repository or business process. Keep deterministic checks around tool arguments and generated files, require tests for code changes, and retain human review for consequential legal, clinical, security, or financial work.