@spacexai/grok-4.7 is live on Brainbase

Grok 4.7 is now available on Brainbase as @spacexai/grok-4.7. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

SpaceXAI released Grok 4.7 on September 21 as its most capable model for coding and knowledge work. It is served at the same price and speed as Grok 4.6, keeps the 500,000-token context window, accepts text and image input, and adds a fourth reasoning level, xhigh, above the default high. The knowledge cutoff moves to May 2026.

#What changed from Grok 4.6

Three things, according to SpaceXAI. The base model is larger, with the parameter count reported at 2.1 trillion against 1.5 trillion for Grok 4.6. The reinforcement learning run is longer and weighted toward problems that take many hours to complete, which is where the gains on terminal and office work come from. And the model was trained to understand the Grok Bot harness natively, so it manages long context and verifies its own work more carefully across a multi-step task.

For agents, self-verification is the change that matters. A model that checks its result before returning it takes fewer corrective turns, and each turn it does not take is context it does not consume and a tool call it does not make.

#Grok 4.7 benchmarks

On SpaceXAI's published evaluation set, Grok 4.7 improves on Grok 4.6 across every benchmark reported. The largest gain is on Terminal-Bench 4.0, where it nearly doubles Grok 4.6 and edges GPT-5.6 Sol. It leads all compared models on EEBench and the Harvey legal benchmark, sits just behind Claude Fable 5.1 on AA Briefcase, and trails both frontier models on HealthBench Professional. Select a metric below to compare the reported results.

SpaceXAI ran Grok 4.7 at xhigh effort, Grok 4.6 at high, and the GPT-5.6 Sol and Fable 5.1 comparisons at their maximum settings; the DeepSWE result for Grok 4.7 is at high effort. Independent testing from Artificial Analysis scores Grok 4.7 at 46 on the Intelligence Index, two points above Grok 4.6 on the index's current v4.3 scale, which places SpaceXAI among the top four labs on that leaderboard and moves the model past GPT-5.6 Sol on the Coding Agent Index. Their AA-Briefcase result matches the published 1,657 Elo. The cost of those gains is tokens: at xhigh, Grok 4.7 used about 81,000 output tokens per index task against 38,000 for Grok 4.6, so the same task costs more to run even though the per-token price did not move.

#Grok 4.7 pricing and context

Grok 4.7 keeps Grok 4.6's rate card. Standard pricing applies to prompts under 200,000 tokens; a request at or above that threshold is billed at the long-context rate for all of its tokens, not just the excess. The fast variant serves the same model at twice the output speed for twice the price, through Cursor and Grok Build rather than the public API.

These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage. Two things follow for agents. Output tokens dominate at xhigh, so the $6 output rate and the higher token usage together set the cost of a run. And the 200,000-token threshold doubles the bill for every token in a request that crosses it, which makes context compaction and prompt caching worth setting up before the first long task.

#Run Grok 4.7 on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Grok 4.7 inside the Codex harness on a managed sandbox:

curl https://api.brainbaselabs.com/v2/threads \
  -H "Authorization: Bearer $BRAINBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": {
      "harness": "codex",
      "model": "@spacexai/grok-4.7"
    },
    "input": "Audit this repository, fix the highest-impact reliability issue, and run the tests."
  }'

#Where Grok 4.7 fits

Grok 4.7 is the strongest choice in its price band for terminal-heavy work, engineering and legal tasks, and multi-hour office work that ends in a document or a deck. Its EEBench and Harvey results are the best of the compared models, and its Terminal-Bench score puts it level with GPT-5.6 Sol at a third of the input price.

Where Fable 5.1 leads by a wide margin, on Terminal-Bench and CursorBench, the question is whether the gap is worth five times the input price for your task. Run Grok 4.7 at high rather than xhigh first: Artificial Analysis found the two settings score the same on the Intelligence Index, and high uses far fewer tokens. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model and its effort setting are the only variables.