Grok 4.7 on Bedrock: 72 Agent Runs Show What Reasoning Effort Buys (and Costs)

If you run agents on Grok 4.7, the most useful new data point this week comes from AWS, not from SpaceXAI. In an October 13, 2026 post on the AWS Machine Learning blog, AWS engineers built a tool-calling agent on Grok 4.7 on Amazon Bedrock and ran it 72 times across four reasoning effort levels. The agent answered 67 of 72 correctly. At xhigh it got all 18 right; at low it got 15 of 18 using about one eighth of the output tokens. This post walks through the data, the setup, and a practical way to choose an effort level.

AWS announced the model's Bedrock availability with its four reasoning effort levels on September 28, 2026, and Unite.AI reported the 500K token context window.

For the model itself, see our coverage of Grok 4.7 on Amazon Bedrock and the Grok 4.7 training data and parameter claims.

Quick reference

What does SpaceXAI say Grok 4.7 is for?

AWS summarizes SpaceXAI's claims: Grok 4.7 uses a larger base model than Grok 4.6, was trained with a longer reinforcement learning run on a harder mix weighted toward many-hour tasks, and is better at verifying its own work and at managing long context. The vendor-published table that AWS reproduces, from SpaceXAI's launch post, shows the biggest gains on sustained-work categories:

These are SpaceXAI's published numbers at launch, not AWS measurements. For how Grok 4.6's own launch claims looked, see our Grok 4.6 launch and evals post.

How was the test built?

A calendar grid being filled with puzzle pieces, representing the maintenance scheduling task in the Grok 4.7 agent evaluation

AWS created a synthetic environment of 20 services exposed through seven tools, with three tasks designed so that a quick heuristic is not enough:

AWS lists four evaluation practices worth copying whatever model you use:

1. State every rule the grader relies on. Otherwise you measure the prompt, not the model.
2. Prove each task has exactly one answer. Each environment is small enough to enumerate by brute force; AWS shows asserts that exactly one schedule is valid and that removing the team rule makes more than one valid.
3. Grade a required final line, not the whole response. Each prompt asks for a FINAL: line, and format failures are tracked separately from wrong answers. All 72 responses included a gradeable final line.
4. Constrain tool parameters with enums so every call targets a real resource.

The system prompt is short and asks the agent to check its answer against every rule before replying.

What were the results?

Three findings stand out.

• xhigh was perfect. All 18 xhigh runs were correct; medium and high each got 17 of 18. With six runs per cell, AWS says the differences are indicative, not statistically conclusive, though the direction was consistent across all three tasks.
• Effort changes reasoning, not exploration. Tool calls were nearly flat, so the agent gathered the same facts at every level. What changed was how much it thought about them.
• Low is a bargain, with a catch. Low reached 15 of 18 at roughly an eighth of the output tokens and a median of 18 seconds. Where answers get reviewed downstream, that is efficient.

AWS also notes where effort mattered most. Constraint satisfaction (scheduling) saw the clearest xhigh advantage. Optimization and hypothesis testing were solved at medium and above in all 18 runs. The pattern: extra effort helps most when the agent has to weigh several candidate answers against each other rather than follow one path.

How do you call it?

Grok 4.7 runs on the bedrock-runtime endpoint through a cross-Region inference profile. Use us.xai.grok-4.7 to keep processing in the US, or global.xai.grok-4.7 for the widest capacity. It supports the OpenAI-compatible Chat Completions and Responses APIs and the Converse API, where reasoning comes back as a reasoningContent block. The AWS post's sample uses the OpenAI client with a short-term bearer token generated from IAM credentials via aws-bedrock-token-generator, which avoids storing a long-lived key.

Gotchas from the post:

• IAM needs two resources. bedrock:InvokeModel on the inference profile and on your account's default project.
• Reasoning tokens count as output tokens. They draw down the maximum output limit before answer text appears. If the cap is hit during reasoning you get no answer text, so check for empty content. AWS used 32,768 tokens.
• Confirm effort levels at runtime. Levels can change between versions; the post shows a loop that probes which values are accepted.
• Retries are part of the design. Retry only 408, 429 and 5xx errors, with exponential backoff and jitter, and honor Retry-After.

How should you choose a reasoning effort?

An eye above a row of bars, representing monitoring output tokens and tool calls per agent run at each reasoning effort level

AWS's practical rules, with our notes:

1. Match effort to the task. Search, optimization and hypothesis testing benefit from medium or higher. Single-path tasks, or ones reviewed downstream, can run at low.
2. Set effort per request, not per agent. Effort is a request parameter, so an agent can plan and verify at high and format tool results at low. This is the most transferable idea in the post.
3. Measure on your own tasks, with repeats. One run per configuration is one observation.
4. Monitor reasoning volume. In production you lack a known answer, but you have token counts. Every correct run at medium and above used at least 1,739 output tokens, so a run with zero tool calls or far fewer tokens than the lowest correct baseline for its level is worth flagging. AWS gives a flag_for_review function using a fifth of the lowest correct run as the threshold.

The same effort-versus-cost thinking applies to other labs' models. Our Opus 5.5 task-cost post covers effort settings and caching for Claude, and the four-way Gemini 4 Argon comparison puts Grok 4.7 next to its rivals.

What this means for what you build or pay

• Do not default to the maximum. Seven-plus times the output tokens bought three more correct answers out of 18 here. Whether that is worth it depends on the cost of a wrong answer in your workflow.
• Latency matters for agents. Median time rose from 18s to 106s. For interactive features, low or medium may be the only viable choice.
• Budget in tokens, then convert. AWS reports tokens because rates change. Apply your inference profile and service tier from the Bedrock pricing page; the Global and US profiles are priced differently, and the Flex and Priority tiers apply multipliers.
• Route by urgency. Service tiers are set per request, so an overnight analysis can use Flex and an interactive call Priority.
• Secure your agent. An agent that calls tools in your cloud account needs guardrails; see our coverage of the AgentCore hijack research, and consider AgentBeam, the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

Limitations AWS lists

• Six runs per task and level: indicative, not conclusive, with no confidence interval.
• Synthetic environment: no real-tool latency, partial results or errors.
• Minutes, not hours: the "long-horizon" label is a scaled-down proxy.
• One model version, one Region, one prompt per task.
• Token counts, not currency.

We would add one more: AWS is a distributor of the model it tested, which is a normal conflict of interest for a vendor tutorial. The code and method are shown, so you can re-run it on your own tasks.

What we are watching next

• Independent replications on other models using the same one-answer-per-task design.
• Whether Bedrock adds a max level or changes effort defaults.
• Real long-horizon (multi-hour) agent results, where the effort trade-off could look different.

Related reading

• Grok 4.7 on Amazon Bedrock
• Grok 4.7 training data and 2.1 trillion parameter claims
• Gemini 4 Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
• Opus 5.5 task cost, caching and effort
• AgentCorruption: hijacking Bedrock AgentCore agents
• Grok 4.6 launch and evals

Details are accurate as of October 13, 2026. All measurements are from AWS's own post; benchmark scores in the first table are SpaceXAI's published launch numbers.