Grok 4.7 is live on Autohive: what it costs, what it's good at, and what still beats it

Key facts

xAI shipped Grok 4.7 on 21 September 2026, and it’s now available in every Autohive workspace, where it shows up as Grok 4.7 under the xAI provider with the public API ID grok-4.7. It reads both text and images as input and writes text as output. The context window holds 500,000 tokens, and Autohive supports up to 128,000 output tokens per response. Pricing starts at $2 per million input tokens and $6 per million output tokens for standard requests. Because this is a generally available model rather than a preview or beta, you can switch an existing agent over to it without any migration work.

The rest of this post covers what changed since the last version, how the model tests against its own claims and against independent benchmarks, what it actually costs to run in practice, and where it fits best among the agents you’re already building.

What changed from Grok 4.6

The headline change is agentic coding, since Grok 4.7 plans, calls tools, and works through multi-step tasks with less hand-holding than Grok 4.6 needed. It writes code, reviews pull requests, debugs, and runs longer tool chains more reliably than the previous version did.

The per-token list price didn’t move, so Grok 4.7 costs the same per token as Grok 4.6. That looks like a free upgrade, but there’s a catch worth understanding before you switch anything over. Because Grok 4.7 tends to think harder and produce more tokens per task, your actual bill per job can rise even though the price on each token stayed flat. The math on that trade-off is below.

The 500,000 token context window carries over, along with the tiered pricing that doubles rates on very large requests. If you were already banking the cached-input savings xAI added to Grok 4.5, those still apply.

Specs and reasoning behavior

xAI’s own API lets developers pick a reasoning effort of low, medium, high, or xhigh, which controls how much internal computation the model spends before it answers. Autohive manages reasoning for you and doesn’t expose a manual effort slider right now, which keeps setup simple but also means you can’t dial the effort down by hand to save tokens. Because of that, it’s worth watching your run costs closely for the first few weeks after you switch.

xAI’s own platform adds function calling, web search, X search, and code execution, which sit outside what Autohive currently exposes. Inside Autohive, you get tool use through the integrations and custom actions you already build, plus reasoning, streaming, and vision, so most agentic workflows still have everything they need.

xAI benchmark claims (vendor-run)

These numbers come from xAI’s own launch post, and because xAI ran every test itself, including the runs for the competitor models, it makes sense to treat them as vendor claims rather than neutral, third-party results.

Grok 4.7 doesn’t win everything, even on xAI’s own table. GPT-5.6 Sol leads on DeepSWE, and Claude Fable 5.1 leads on CursorBench, Terminal-Bench, and HealthBench, so rivals beat Grok 4.7 on several tasks even in the tests xAI chose to run.

A few caveats matter here, and they change how much weight the numbers above should carry. xAI ran the competitor models itself, effort settings weren’t matched across models, and Grok 4.7 usually ran at xhigh, its most expensive setting. xAI also didn’t publish sample sizes or error bars for any of these results. Given all that, it makes more sense to read this table as a general direction than as a strict scoreboard.

Independent benchmark evidence

Artificial Analysis, an independent group, tested Grok 4.7 on its own hardware and its own prompts, which gives a cleaner read on real-world performance than numbers a vendor publishes about itself.

The Coding Agent Index jumped 9 points, which backs up xAI’s main claim that the coding gains are real rather than marketing. The broader Intelligence Index rose only 2 points, a smaller step but still a genuine one.

The token count, though, is the number worth watching most closely. Grok 4.7 used about 81,000 output tokens per task in that suite, more than double the roughly 38,000 tokens Grok 4.6 needed to finish the same kind of work. Artificial Analysis also clocked about 7.1 minutes per task in its benchmark run, but that figure describes a heavy, multi-step benchmark task rather than normal chat, so don’t read it as the delay you’ll feel in a quick conversation.

Taken together, independent testing shows a real improvement over 4.6, strongest in agentic coding, though Grok 4.7 still sits behind several rivals on the wider rankings.

Grok 4.7 pricing

Grok 4.7 uses tiered pricing based on the size of your prompt, and all the figures below are in USD per 1 million tokens.

The higher rate applies to the whole request once your prompt crosses 200,000 tokens, not just the tokens above the line. Confirmed by xAI’s models documentation and release notes.

Cost examples in plain English

Same token price, higher task cost. Grok 4.7 charges the same $6 per million output tokens as Grok 4.6, but independent testing showed it produces about 81,000 output tokens per task versus roughly 38,000 for 4.6. At $6 per million, that works out to roughly $0.49 per task for 4.7 against about $0.23 for 4.6. In other words, the price per token didn’t change, yet the cost per finished job about doubled anyway, which is a reminder that low per-token pricing doesn’t guarantee a lower bill per task.

The 200,000 token cliff. A request with 199,000 prompt tokens is billed at $2 per million input, but push it to 205,000 tokens and the entire request jumps to $4 per million input, plus $12 output, because the higher rate applies to the whole prompt rather than just the tokens above the line. That means a small increase in prompt size can nearly double the price of a single call, so if your prompts regularly sit near 200,000 tokens, it’s worth trimming them or budgeting for the step up.

Grok 4.7 vs GPT-5.6, Claude, and Gemini

Grok 4.7 is one of the cheapest options here on per-token output, undercutting every other model in the table at the standard tier, which makes it a reasonable default for high-volume agentic work. For the full tier breakdown on the OpenAI side, including how the Terra and Luna variants are priced, see Autohive’s GPT-5.6 Sol, Terra, and Luna pricing breakdown. What you give up with Grok 4.7, though, is context: its 500K window is smaller than the 1M and 1.05M windows of its rivals, and its pricing doubles on large requests while Claude’s stays flat no matter how big the prompt gets. Benchmark comparisons across vendors are also rarely apples-to-apples, since each lab tests under its own settings and effort levels.

Best Autohive use cases for Grok 4.7

Grok 4.7 is a strong pick whenever the work in front of your agent is agentic and code-heavy:

• Agentic coding and debugging across several steps
• Pull request review and complex refactoring
• Multi-step tool workflows that chain integrations and custom actions
• Analysis of large documents that still fit under the 500K window, including reference files stored in your agent’s Content Hub, which is where you keep source material for an agent to pull from during a run
• Image, dashboard, and UI review using vision
• Direct internal analysis where you want plain answers
• Research that benefits from X search on xAI’s platform

If you’re wiring any of these into a live agent for the first time, the guides on building a custom agent, setting up multi-agent teams, and creating a workflow walk through the setup in more detail.

When another model is a better choice

Grok 4.7 shouldn’t be your default for every agent, and there are a few situations where a different model will serve you better:

• You need a 1M+ context window or flat long-context pricing. Claude holds a steady rate across its full window, which suits very large jobs better. See the Claude Opus 5 release.
• You want the highest-end coding or computer-use work. Current OpenAI and Anthropic options lead Grok on some of those tasks, even in xAI’s own numbers.
• The task is routine. A cheaper worker model handles simple, repetitive jobs at a fraction of the cost. Our post on why you shouldn’t run every agent on one model covers the trade-off, and the model selection guide walks through matching a model to a job.

Grok 4.7 also takes longer to produce a first token because it reasons before replying, it burns more tokens per task than 4.6, and it applies looser content filtering than some rivals, so it’s worth weighing all three factors before you switch a production agent over.

How to enable Grok 4.7 in Autohive

New to Autohive? Create a new agent first, then come back and point it at Grok 4.7 using the steps below.

Switching an existing agent takes no code and no migration script:

1. Open your Workspace.
2. Click Edit agent on the agent you want to update.
3. Go to Model configuration.
4. Under Model type and name, choose xAI, then Grok 4.7.
5. Click Save.

After you save, test your key prompts and action chains against the new model, then check real run costs on a few live jobs before you trust it with anything important. Because Grok 4.7 can use noticeably more tokens per task than the version it replaces, checking your actual spend rather than assuming the list price tells the whole story is the step that matters most.

FAQ

Is Grok 4.7 more expensive than Grok 4.6? Not per token, since the list price is identical. But because Grok 4.7 tends to produce more tokens per task, your total cost per job can land higher even at the same rate.

What is the context window on Autohive? 500,000 tokens for input, with up to 128,000 output tokens per response.

Can I set the reasoning effort myself? Not in Autohive today. The xAI API offers low, medium, high, and xhigh effort levels, but Autohive handles reasoning internally and doesn’t show a manual slider.

Is Grok 4.7 a preview or a final release? Generally available, not a preview or beta.

Why does a request suddenly cost more once my prompt gets large? Once your prompt hits 200,000 tokens, the rate doubles for the entire request, to $4 input, $1 cached input, and $12 output per million. Keep prompts under that line if you want the lower rate.

Do I need to rebuild my agent to use Grok 4.7? No. Change the model in your agent settings and save. There’s no migration step, just testing your prompts and checking costs afterward.

How does Grok 4.7 compare to GPT-5.6 Sol and Claude on price? Grok 4.7 is cheaper per output token at the standard tier than every model in our table. The trade-off is a smaller context window and pricing that doubles on large requests.

The bottom line

Grok 4.7 is a real step up for agentic coding, backed by independent testing, and it comes at a low per-token price. The catch is token burn, because it thinks more and writes more per task, so you’ll want to watch your real per-job costs, mind the 200,000 token pricing cliff, and keep a bigger-context or cheaper model on hand for the jobs that need one. It’s live in your workspace right now, so switch an agent over, run a few real tasks through it, and check the bill before you roll it out everywhere else.