Artificial intelligence news.
Straight from the original announcements.

Grok 4.7 arrives on Amazon Bedrock, with tunable reasoning for long-running agents

October 13, 2026, 17:52By the AI News Bites Editorial TeamSource: AWS
Grok 4.7 arrives on Amazon Bedrock, with tunable reasoning for long-running agents
Image: AWS

SpaceXAI's Grok 4.7 is now available on Amazon Bedrock, and AWS has published a walkthrough of building a tool-calling agent on it. In AWS's test of 72 agent runs, the agent answered 67 correctly, and the results show how the model's reasoning effort setting trades accuracy against token use.

Key points
  • SpaceXAI's Grok 4.7 is now available on Amazon Bedrock via the us.xai.grok-4.7 and global.xai.grok-4.7 inference profiles.
  • SpaceXAI lists a 500K token context window, with text and image input and text output.
  • Grok 4.7 supports four reasoning effort levels: low, medium, high and xhigh.
  • In AWS's 72 agent runs, 67 were correct; xhigh scored 18 of 18, while low scored 15 of 18 using about an eighth of the output tokens.
  • Reasoning tokens count as output tokens, so completion limits must leave room for both reasoning and the answer.

What SpaceXAI says about Grok 4.7

According to SpaceXAI, Grok 4.7 is trained to work longer on more difficult tasks and knows when to check its own work. That matters most for agents, which carry a task through many steps and must verify what they find along the way. SpaceXAI says the model uses a larger base model than Grok 4.6 and was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete.

SpaceXAI's published benchmark results, as reproduced by AWS, show the largest gains over Grok 4.6 in categories that measure sustained work:

  • Terminal-Bench 4.0 (multi-hour terminal work): 20.3% to 37.6%
  • EEBench (electrical engineering): 53.0% to 64.0%
  • Harvey Legal Agent Benchmark: 15.8% to 19.6%
  • CursorBench 4.0 (longer-running coding): 40.4% to 46.3%
  • HealthBench Professional (clinical reasoning): 48.5% to 56.7%
  • AA Briefcase v1.1 (multi-hour office work): 1,546 to 1,657

SpaceXAI lists a 500K token context window, with text and image input and text output. On Amazon Bedrock, the model supports configurable reasoning effort, tool calling and structured output.

How to access it on Bedrock

Grok 4.7 runs on the bedrock-runtime endpoint and is reached through a cross-Region inference profile. Two profiles are offered. The profile us.xai.grok-4.7 routes within the US geography and suits teams with US data residency requirements. The profile global.xai.grok-4.7 routes across commercial AWS Regions and suits teams with no residency constraints that want the widest capacity. Requests name a profile rather than a bare model ID.

The model supports the OpenAI-compatible Chat Completions and Responses APIs, as well as the Amazon Bedrock Converse API. On Converse, reasoning comes back as a reasoningContent block next to the text block. AWS advises setting only maxTokens in inferenceConfig and placing model-specific parameters, such as reasoning effort, in additionalModelRequestFields.

Access requires an AWS account with model access enabled for Grok 4.7 and an IAM identity with bedrock:InvokeModel permission on both the inference profile and the account's default project. For production, AWS recommends generating a short-term bearer token from existing AWS credentials so that no long-lived secret is stored.

One practical warning: reasoning tokens count as output tokens and draw down the completion limit before answer text appears. If the limit is hit during reasoning, the response comes back with no answer text, so agent code should check for empty content. AWS used a limit of 32,768 tokens in its runs.

Testing reasoning effort on search-style tasks

Grok 4.7 offers four effort levels: low, medium, high and xhigh. To see what they change, AWS built a synthetic environment of 20 services exposed through seven tools, with three tasks that require search rather than simple lookup. Each has exactly one correct answer, checked by brute force.

  • two_fault_diagnosis: two services failed silently and 12 others raised alerts. Of 190 possible pairs, only one explains exactly the alerts observed.
  • maintenance_schedule: patch eight services over four nights under ten constraints and a rule that two services from the same team never share a night. There are 2,520 possible schedules and one is valid.
  • migration_budget: choose migrations that maximize risk reduction within a budget, including transitive prerequisites. There are 53 feasible sets and one optimal. Picking the best ratio first lands nine points short of the optimum.

AWS ran every task at every effort level six times, for 72 runs. The agent answered 67 correctly. At xhigh it got all 18 right. At low it got 15 of 18 right while using about an eighth of the output tokens. AWS reports token counts rather than currency, since prices change and the counts are what a user's own costs derive from.

The post also lists practices for trustworthy evaluations: state every rule the grader relies on, prove each task has a single answer by enumerating candidates, and grade a required final line rather than the whole response.

What it means for agent builders

The result gives developers a lever they can set per request. Tasks where a wrong answer is costly may justify xhigh, while cheaper, faster runs at low may be enough for others. AWS stresses that benchmarks only show performance on standard tasks, and recommends measuring on tasks shaped like a team's own workload. For current Regions, service tiers and token prices, AWS points to the Grok 4.7 model card in the Amazon Bedrock User Guide and the Bedrock pricing page.

Related news

AWS shows how to route coding agent requests with NeMo Switchyard
For developersNovember 13, 2026, 23:30

AWS shows how to route coding agent requests with NeMo Switchyard

AWS has published a sample deployment that routes coding agent requests between models on Amazon Bedrock using NVIDIA NeMo Switchyard and LiteLLM. The approach measures…

AWS adds models, agent safeguards and cost controls in October
Products and toolsNovember 11, 2026, 23:26

AWS adds models, agent safeguards and cost controls in October

AWS says its October 2026 releases give AI builders more model choices, stronger controls for production agents and clearer ways to manage costs. The updates span Amazon…

Moonshot AI’s Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock
ModelsNovember 11, 2026, 02:07

Moonshot AI’s Kimi K3.1 Pro Highspeed arrives on Amazon Bedrock

Moonshot AI’s Kimi K3.1 Pro Highspeed is now available on Amazon Bedrock, giving developers access to a post-trained Kimi K3 model with a 1-million-token context window…