Grok 4.7 reaches Kiro in GovCloud at up to 5.4x credit cost

Amazon has added Grok 4.7 to the Kiro IDE and CLI in the AWS GovCloud (US) Regions, according to an announcement dated November 13. The model is experimental, has a 500K context window and costs 2.7x or 5.4x credits per request.

AWS describes Grok 4.7 as the maker’s most capable model for coding and knowledge work. AWS prints the maker’s name as SpaceXSI; other coverage of the September launch uses SpaceXAI, and this article treats them as the same company.

Two credit multipliers split at 200K tokens

The pricing is the most concrete part of the post. Requests of up to 200K tokens use a 2.7x credit multiplier. Anything above that is billed at 5.4x. The multipliers apply to Kiro’s credits, not dollars, and the announcement does not say what the base credit costs, so the real price per task cannot be worked out from the page.

The 500K window is therefore split into two tiers. A long repository session that crosses the 200K line pays double the rate of one that stays under it. Teams that run big refactors will want to watch that threshold in the monitoring and tracking guide AWS links from the post.

AWS labels support as experimental, without saying what that changes in practice. It does not explain whether the label affects reliability, the service terms or how long the model stays available.

To get it, users update the IDE or CLI to the latest version, restart it and pick the model from the selector. Administrators control which models their users see, so an organization can leave Grok 4.7 switched off if its policies require.

What AWS claims about long coding runs

The post says the model is built to stay focused on hard tasks that run for hours and to check its work before moving on. Compared with Grok 4.6, AWS says, it plans better and recovers from errors more reliably in repo-scale coding.

These are statements from AWS and the model’s maker. The announcement gives no benchmark, no test conditions and no numbers for the improvement over Grok 4.6, and nothing here has been independently verified.

A regulated-cloud rollout that trails the public launch by two months

The analysis here is about competition and timing. According to SiliconANGLE, SpaceX launched Grok 4.7 on September 21 at $2 per million input tokens and $6 per million output tokens. The GovCloud arrival in Kiro comes about seven weeks later.

The same report says the benchmark claims at launch came from the company. It notes that one of them, CursorBench, was built by Cursor, which it describes as a SpaceX acquisition. On that test the company said Grok 4.7 beat GPT-5.6 Sol and Fable 5.1, while on EEBench, a chip-design test, it trailed OpenAI’s GPT-6 Astra. A model that leads on some coding tests and trails on others is a mixed pitch against rivals, and the GovCloud post offers no rival comparison at all.

The model has a history in this environment. AWS made Grok 4.6 available on Amazon Bedrock in GovCloud (US) in a post from August. Kiro itself reached GovCloud earlier in the year, and its own blog says content collection for service improvement is off by default, inference stays inside GovCloud and sign-in goes through AWS IAM Identity Center.

That blog also says Kiro had begun pursuing FedRAMP High and DoD Cloud Computing Security Requirements Guide authorization. The page does not say either was granted, and the Grok 4.7 post does not mention compliance status or where the model’s inference runs. For government contractors, those are the questions that decide whether the model can be used, and AWS points readers to its account teams for details.

Who can use it today

The model is limited to customers who use Kiro in the GovCloud (US) Regions, which AWS runs for US government agencies and regulated workloads. Other Kiro users are not covered by this post. For them, the model’s wider availability is a separate matter: one report in the search results says Grok 4.7 reached Amazon Bedrock with the same 500K window, though this article did not verify it.

The practical question for GovCloud teams is cost control. The multipliers are high next to the base rate, and Kiro lets administrators gate the model, so the first adoption signal will be whether organizations leave it on for everyone or restrict it to a few repositories. AWS has not said whether the experimental label will be removed or when.