01
Intake
Text and image. jpg, jpeg, or png, 20 MiB per image, no limit on how many images. Text and image can arrive in either order.
SpaceXAI · September 21, 2026 · model id grok-4.7
The published specification for SpaceXAI's coding and knowledge work model. Text and image in. Text out. A 500,000-token context window.
The announcement calls it their most capable model for coding and knowledge work, and prints the headline twice as fast, at half the price of comparable models. The same post says the standard model is served at the same price and speed as Grok 4.6.
01 · Architecture
SpaceXAI has not printed a parameter count, a layer count, or an expert layout for grok-4.7. What follows is the structure the September 21 announcement and the model docs do describe: a new base, a longer reinforcement-learning run, a native harness, a new safeguard stack, and a fixed reasoning path into text.
Text and image. jpg, jpeg, or png, 20 MiB per image, no limit on how many images. Text and image can arrive in either order.
A new, larger base model than Grok 4.6. The scale of that base is unpublished.
A longer reinforcement-learning run on a harder mix of tasks, weighted toward problems that take many hours to finish.
The announcement says the model is better at verifying its own work and at managing longer context.
Trained to natively understand the Grok Bot harness, for conversation and general knowledge work.
An entirely new safeguard stack. The post calls this the strongest model they have tested on refusals and jailbreak resistance.
Always on. Settings are low, medium, high, and xhigh. If you leave the setting unset, the default is high. It cannot be turned off.
Text. The Responses API also returns encrypted reasoning on every call. Image generation, video, and voice are other models.
| Field | Official record |
|---|---|
| Parameter count | Not published |
| Layer count | Not published |
| Expert layout | Not published |
| Base relative to Grok 4.6 | New, larger |
| Training emphasis | Longer RL, multi-hour tasks, self-check, longer context, Grok Bot harness |
| Modalities | Text and image in, text out |
| Context | 500,000 tokens |
| Knowledge cutoff | May 2026 |
| Current events | Training data only, unless Web Search or X Search is on |
| Regions on the model page | us-east-1, us-west-2, us-central-1 |
02 · Reasoning
reasoning_effort controls how much the model thinks before it answers. presencePenalty, frequencyPenalty, and stop return an error on reasoning models. Reasoning tokens are billed with the rest of the tokens.
More reasoning tokens, for deeper thinking. The docs point this setting at hard problems, complex math, multi-step logic, and competition-level tasks.
curl https://api.x.ai/v1/responses \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.7",
"reasoning": {"effort": "high"},
"input": "Your task"
}'
On the Responses API, grok-4.7 returns reasoning.encrypted_content on every response, even when include does not list it. Chat Completions has no field for that ciphertext. The docs also expose a summary of the internal reasoning. logprobs and top_logprobs are ignored on grok-4.20 and newer.
03 · Published scores
Columns are Grok 4.7 xHigh, Grok 4.6 High, GPT-5.6 Sol Max, and Fable 5.1 Max. Token prices in that table are $2 and $6, $2 and $6, $4 and $20, $10 and $50. The asterisk on DeepSWE marks a high-effort score.
| Benchmark | Grok 4.7 xHigh | Grok 4.6 High | GPT-5.6 Sol Max | Fable 5.1 Max |
|---|---|---|---|---|
| AA Briefcase v1.1 | 1,657 | 1,546 | 1,487 | 1,678 |
04 · Price
Once the prompt reaches 200,000 tokens, every token in that request bills at the long-context rate. Cached input is a quarter of the input rate at both tiers.
$2 / $6
Input and output per 1M tokens, prompt under 200,000. Cached input is $0.50.
| Prompt | Input | Cached | Output |
|---|---|---|---|
| Under 200k | $2.00 | $0.50 | $6.00 |
| Reaches 200k | $4.00 | $1.00 | $12.00 |
| US endpoint, under 200k | $2.20 | $0.55 | $6.60 |
| US endpoint, above 200k | $4.40 | $1.10 | $13.20 |
The US endpoint is https://us.api.x.ai/v1. Token use there bills at 1.1 times the global rate. The docs say that endpoint currently serves grok-4.7 and grok-4.6 only.
Priority processing is twice the standard rates, and only when the response confirms service_tier priority. It covers Chat Completions and Responses. It does not cover image generation, video generation, or Batch API requests.
Batch API is not supported for grok-4.7.
Published rate limits on the model page: 150 requests per second, 50,000,000 tokens per minute.
The same model, on faster infrastructure, at twice the standard token rates. The pricing page limits it to Cursor and Grok Build. The public xAI API does not serve it. Grok Build's free tier does not include it. The announcement says this variant has twice the output speed at twice the price. Cursor bills its own fast variant through the Cursor plan.
| Prompt | Input | Cached | Output |
|---|---|---|---|
| Below 200k | $4.00 | $1.00 | $12.00 |
| Above 200k | $6.00 | $1.50 | $18.00 |
05 · Safety
The announcement says Grok 4.7 is the strongest model they have tested on refusals and jailbreak resistance. It leads their comparison on LatchBio's biosafety benchmark at 62.4%. On HackerBench v0.3, their benchmark for risky and malicious cyber tasks, 3.3% of risky dual-use prompts were allowed through. Select cybersecurity partners have invite-only access to the red-team capabilities for defense research.
06 · Access
Base URL https://api.x.ai/v1. Model id grok-4.7. The API page shows the xAI SDK, the OpenAI SDK, TypeScript, and curl against that id.
Available today, with a free start at x.ai/build. Install line from the announcement: curl -fsSL https://x.ai/cli/install.sh | bash
Available today. Fast, if you want it, is billed through the Cursor plan.
The announcement also lists third-party coding harnesses, model routers, and cloud platforms. Function calling and structured outputs are supported on the model.
The models page sends code, chat, and general work to Grok 4.7. Images go to Grok Imagine Image 2.0. Video goes to Grok Imagine Video 1.5. Voice goes to the Grok Voice API.
07 · Sources
Every figure on this page comes from these SpaceXAI pages, read September 21, 2026. Third-party parameter counts are left off because they are not on these pages.