Grok 4.7: Complete Guide to Pricing, Benchmarks & API (2026)

SpaceXAI shipped Grok 4.7 on 21 September 2026, a month after Elon Musk's first target date and one day before Anthropic launched Claude Opus 5.5. The company calls it "our most capable model for coding and knowledge work". It's a new, larger base model with a longer reinforcement-learning run, not another tune of Grok 4.6. It's live on the xAI API, in Cursor, in Grok Build, in GitHub Copilot and on the major model routers.

The headline numbers depend on who's doing the measuring. xAI's own benchmark table shows big jumps, including +17 points on Terminal-Bench 4.0. Artificial Analysis's independent Intelligence Index moves by only two points, and the model now uses roughly twice as many output tokens per task. This guide covers what Grok 4.7 is, what it costs, where you can run it, how its benchmarks hold up, and how to call it from code.

What is Grok 4.7?

Grok 4.7 is a closed, API-only reasoning model from SpaceXAI. That's the name xAI took after SpaceX acquired it in early 2026 and it rebranded in July. The developer platform is still docs.x.ai, the endpoint is still api.x.ai, and model IDs still start with grok-. xAI's launch post says the model "works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date".

Under the hood, xAI says it trained a larger base model and ran a longer RL phase weighted toward hard, multi-hour tasks. The practical results are better self-verification, better long-context management and a model that keeps going on agentic work instead of stopping early. Earlier reporting blamed the delay on RL that penalised response length too hard, so the model gave up on tasks too soon.

A note on parameter count: several outlets reported Grok 4.7 as a 2.1-trillion-parameter model trained partly on SpaceX engineering data. Neither figure appears in xAI's launch post or model documentation. Both come from Musk's posts in the run-up to launch. Treat them as reported, not confirmed.

There is no Grok 4.7 Heavy, mini or dedicated code variant. SpaceXAI's coding-agent model is a separate product, grok-build-0.1. The one sibling SKU is Grok 4.7 Fast, covered in the pricing section below.

How much does Grok 4.7 cost?

List pricing is unchanged from Grok 4.6. There are two tiers, split at a 200k-token prompt.

Source: xAI model docs. Three details matter in practice:

• The 200k cliff. Once a prompt reaches 200k tokens, every rate doubles. Third-party pricing guides report that the higher rate then applies to the whole request, not just the tokens over the threshold. If you run long-context agents, compacting history to stay below 200k halves your bill.
• Caching needs a hint. xAI recommends setting prompt_cache_key on the Responses API, or the x-grok-conv-id header on Chat Completions, so a conversation keeps hitting the same server. Without it, cache hits are unreliable, and cached input is 75% cheaper than fresh input.
• Grok 4.7 Fast isn't on the public API. The docs say Fast is "the same model served on faster infrastructure, billed at twice the standard token rates. It is available only in Cursor and Grok Build, and it is not included in Grok Build's free tier." You can't call it by model ID.

For comparison, OpenAI lists GPT-6 Astra at $10 / $50 and Anthropic's new Claude Opus 5.5 at $4 / $20. Grok 4.7 is the cheapest frontier-class option on list price by a wide margin. The next section explains why list price understates what you'll actually pay. For a deeper cost walkthrough of the same pricing structure, see our Grok 4.6 pricing and API costs breakdown. Every rate there carries over to 4.7.

Why does Grok 4.7 cost more than its list price suggests?

Because it uses far more tokens. Artificial Analysis measured Grok 4.7 at xhigh using about 81k output tokens per Intelligence Index task. OfficeChai reports roughly 36k for Grok 4.6, so 4.7 uses about 2.25× as many. Running the whole index took around 240M tokens, against a median of about 88M for comparable models.

At $6 per million output tokens, 81k tokens comes to about $0.49 of output per hard task, before input. At the same per-task token count, Grok 4.6 would cost about $0.22. So Grok 4.7's price per task went up even though its price per token didn't. It's also slow: Artificial Analysis puts it at about 7.1 minutes per Intelligence Index task. Its live throughput figures vary widely between measurement windows, so check the current model page rather than trusting a single number.

The fix is to choose effort levels on purpose. high is the default, and xAI describes low as best for latency-sensitive agentic use. Save xhigh for tasks where a correct answer is worth several minutes and tens of thousands of tokens.

Where is Grok 4.7 available?

SpaceXAI made it available across its own products and the major developer channels on day one. That's easier now that SpaceXAI owns Cursor: the $60B acquisition closed on 14 August 2026.

Sources: xAI launch post, GitHub changelog, OpenRouter. Copilot Business and Enterprise admins must enable the model in their Copilot model policy settings before it shows up.

If you want to try Grok 4.7 as a terminal agent, Grok Build is the most direct route: 4.7 is its default model. Our Grok Build CLI install guide covers setup, and Grok Build vs Claude Code vs Codex CLI compares it with the alternatives.

How good is Grok 4.7 on benchmarks?

There are two sets of numbers, and they tell different stories.

xAI's published benchmarks

xAI's launch post compares Grok 4.7 with Grok 4.6, OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5.1. These are vendor-reported scores.

* High-effort score. Source: x.ai/news/grok-4-7. Some secondary coverage rounds Terminal-Bench to 38.0%. xAI's own table says 37.6%. Anthropic's Opus 5.5 system card lists Fable 5.1 at 55.8% on Terminal-Bench 4.0 (and GPT-6 Astra at 57.9%), so treat the Fable 5.1 column as xAI's figures.

Taken at face value, Grok 4.7 beats Grok 4.6 on every row, with the biggest jump on Terminal-Bench (+17.3 points). But even in xAI's own chart, Fable 5.1 leads on four of seven benchmarks. Grok 4.7 wins EEBench and Harvey Legal outright, and on Harvey the whole field scores low. xAI has also moved away from the older public suites. It reports no SWE-bench Verified, GPQA, AIME, HLE or ARC-AGI scores for 4.7, so you can't line it up against older launch posts.

On safety, xAI says Grok 4.7 ships with an entirely new safeguard stack. On HackerBench v0.3 it lets 3.3% of risky dual-use prompts through, and it scores 62.4% on the LatchBio biosafety benchmark.

Independent results (Artificial Analysis)

Source: Artificial Analysis, "Benchmarking Grok 4.7". On the Intelligence Index, Grok 4.7's 46 sits behind Fable 5.1 (53), GPT-6 Astra (53), Claude Opus 5 (51), Fable 5 (50), Muse Spark 1.3 (48) and GPT-5.6 Sol (47), per OfficeChai's summary of the leaderboard. The Coding Agent Index is the brighter result: paired with Grok Build, 4.7 comes 4th among native harnesses and passes GPT-5.6 Sol.

Put the two sets together and the picture is clear. Grok 4.7 is a real step up for agentic coding and professional knowledge work. It's a small step on general reasoning. It isn't the new frontier leader, and a slice of its gains comes from spending more tokens. For a row-by-row look at what changed, read Grok 4.7 vs Grok 4.6. For context on the previous release, see Grok 4.6 benchmarks explained.

How do you use the Grok 4.7 API?

The xAI API is OpenAI-compatible. Point the OpenAI SDK at https://api.x.ai/v1 with an xAI key from the console. xAI's recommended surface is the Responses API, where reasoning effort is a nested object.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["XAI_API_KEY"],
    base_url="https://api.x.ai/v1",
)

response = client.responses.create(
    model="grok-4.7",
    reasoning={"effort": "high"},          # low | medium | high (default) | xhigh
    prompt_cache_key="repo-refactor-42",   # keeps cache hits on the same server
    input="Find the race condition in this Go worker pool and propose a fix:\n\n" + open("pool.go").read(),
)

print(response.output_text)
print(response.usage)  # watch output_tokens: 4.7 is verbose

Here is the same call with curl:

curl https://api.x.ai/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -d '{
    "model": "grok-4.7",
    "reasoning": {"effort": "medium"},
    "input": "Write a SQL migration that adds a soft-delete column to orders."
  }'

If your codebase already uses Chat Completions, it works too. Pass reasoning_effort as described in xAI's reasoning docs, and use the x-grok-conv-id header for caching:

completion = client.chat.completions.create(
    model="grok-4.7",
    reasoning_effort="low",
    extra_headers={"x-grok-conv-id": "support-bot-session-981"},
    messages=[{"role": "user", "content": "Summarise this stack trace in two lines: ..."}],
)
print(completion.choices[0].message.content)

A few practical notes:

• Reasoning is encrypted. On the Responses API, Grok 4.7 always returns reasoning.encrypted_content, even if you don't ask for it. You get readable summaries, not the raw chain of thought. Pass the encrypted blob back on later turns if you want the model to keep its reasoning context.
• Data residency. Use https://us.api.x.ai/v1 to keep inference in the US, at a 10% token premium.
• Through OpenRouter. Use x-ai/grok-4.7 against https://openrouter.ai/api/v1. Our Claude Code + OpenRouter guide shows how to route a coding agent through it.
• Cap your spend. Given the token use, set max_output_tokens on batch jobs and log usage for every call during evaluation.

What about Grok 4.8?

Grok 4.7 had a messy launch. On 24 July, Musk said it was about four weeks out. The date slipped repeatedly through August and early September. Then on 13 September, Musk announced Grok 4.8, a roughly 2.5-trillion-parameter model trained on SpaceXAI's new C++ stack, before 4.7 had even shipped. That led to reports that 4.7 had been shelved. It hadn't: 4.7 launched a week later.

As of 22 September there's no Grok 4.8 model page, API ID, price or release date. Musk said pre-training was finishing and RL would follow. SpaceXAI's recent releases have come about a month apart (4.6 on 12 August, 4.7 on 21 September), but Musk's stated timelines have slipped before, so plan around Grok 4.7 for now. If the 4.6-to-4.7 pattern holds, a 4.8 release would likely keep the same API surface, and moving to it would mean changing one model string.

Who should use Grok 4.7?

Grok 4.7 is a good fit if:

• You already work in Cursor or Grok Build. It's the house model there, the harness is tuned for it, and the Fast variant is only available there. The Coding Agent Index jump (+9) is the most convincing number in the release.
• You want frontier-adjacent quality at mid-tier prices. At $2 / $6 it costs a fifth of GPT-6 Astra's input rate and about an eighth of its output rate, even after you account for the extra tokens.
• Your work is in domains where it leads. Electrical engineering (EEBench) and legal-agent tasks are its strongest results relative to the field.
• You're on Grok 4.6 today. It's the same price and the same API. Swap the model string, then re-baseline cost per task, because output tokens will go up.

Look elsewhere if:

• You need the strongest general reasoning. Fable 5.1, GPT-6 Astra and Claude Opus 5 all score higher on Artificial Analysis's index. See our Opus 5.5 vs GPT-6 Astra vs Grok 4.7 comparison.
• You need more than 500k tokens of context. Grok 4.3 offers 1M and GPT-6 Astra about 1.05M. Anthropic's Claude Opus 5.5 offers 1M.
• Latency matters. At high and xhigh it's slow and verbose. Developers on Hacker News also describe Grok's output as terse and full of coined terms, so it's a weaker choice for user-facing prose.
• You need very long unattended runs. xAI built 4.7 to work for hours, but independent coverage (The New Stack) reports it still fails most long-horizon tasks. Keep a human checkpoint in the loop.

The best way to decide is to run your own evals. Take 20–50 real tasks from your backlog, run them at medium and high, and compare pass rate and total cost against your current model. Grok 4.7's value depends almost entirely on how many tokens it spends on your workload.

FAQ

What is the Grok 4.7 model ID?

On the xAI API it's grok-4.7, with no aliases. On OpenRouter it's x-ai/grok-4.7. Grok 4.7 Fast has no public model ID. It's only available inside Cursor and Grok Build.

How much does Grok 4.7 cost per million tokens?

$2 input, $0.50 cached input and $6 output for prompts under 200k tokens. From 200k tokens up, the rates double to $4, $1 and $12. The US-only endpoint adds 10%, and the Fast variant in Cursor and Grok Build costs twice the standard rate. List pricing is identical to Grok 4.6.

What is Grok 4.7's context window?

500,000 tokens, the same as Grok 4.6. It accepts text and images as input and outputs text. Its knowledge cutoff is May 2026.

Is Grok 4.7 better than Grok 4.6?

Yes, on every benchmark xAI published, most dramatically Terminal-Bench 4.0 (37.6% vs 20.3%) and EEBench (64.0% vs 53.0%). Independently, Artificial Analysis's Coding Agent Index rose 9 points but its Intelligence Index rose only 2, and 4.7 uses about 2.25× the output tokens. So expect better results at a higher cost per task.

Is Grok 4.7 open source?

No. Grok 4.7 is proprietary and available only through the API and SpaceXAI's products. SpaceXAI hasn't published its weights or its parameter count. The 2.1-trillion figure in some coverage comes from Musk's posts, not official specifications.

When is Grok 4.8 coming out?

There's no release date. Musk announced Grok 4.8 on 13 September 2026 as a roughly 2.5-trillion-parameter model and said reinforcement learning would follow pre-training. As of 22 September there's no model page, API ID or price.

Sources

• SpaceXAI: Grok 4.7 launch post
• SpaceXAI Docs: Grok 4.7 and model specifications
• SpaceXAI Docs: Reasoning
• Artificial Analysis: Benchmarking Grok 4.7
• GitHub Changelog: Grok 4.7 in Copilot
• OfficeChai: Intelligence Index leaderboard and token usage
• CellCog: Grok 4.8 announcement

If your team is building products on Grok 4.7 or other frontier models and needs engineers who already work this way, Codersera can help you hire vetted remote developers who fit your stack.