xAI's Grok 4.6 Lands on Amazon Bedrock With a 500K Context Window

Amazon has made xAI's Grok 4.6 available in Amazon Bedrock, describing it as a frontier model built for long-running agents, coding, and knowledge work. It is xAI's second model in the Bedrock catalog, following Grok 4.3. AWS dates the Bedrock launch to August 18, 2026, and says the model is reachable through both the bedrock-mantle and bedrock-runtime endpoints.

What is actually on offer

Grok 4.6 takes text and image input and returns text only. AWS is explicit that there is no audio, speech, video, embedding, or image generation support, which rules it out for multimodal output work but leaves it squarely aimed at agents and code.

The headline specification is a 500K-token context window paired with configurable reasoning effort at four levels: low, medium, high, and xhigh. That second detail is the more useful one operationally. Reasoning models have a habit of burning tokens on problems that did not need them, and an effort dial gives teams a way to tune cost per call rather than choosing between a reasoning model and a cheap one.

Supported features include tool calling, prompt caching, response streaming, structured outputs, and guardrails. On the API side, Grok 4.6 works with the Chat Completions API, the Responses API, the Converse API including streaming, and OpenAI-compatible endpoints. It does not support the Invoke API, which is worth checking against existing code, since older Bedrock integrations are frequently written against Invoke.

AWS says xAI built Grok 4.6 on Grok 4.5 with a particular focus on long-running agents. The company describes extended supplemental training beyond Grok 4.5 using curated model-generated data for reasoning and technical concepts, plus agentic reinforcement learning across knowledge work, coding, kernel optimization, web development, and CAD. xAI says the model shows increased self-testing and verification on longer trajectories, and that improved safeguards are backed by what it calls its widest-ever suite of pre-deployment testing.

Self-verification on long trajectories is the claim that matters for agent builders. The dominant failure mode in multi-step agents is not a single wrong answer but a wrong answer that the agent accepts and then builds ten more steps on top of. A model that checks its own work more often is attacking the right problem, though neither AWS nor xAI publishes a metric for it here.

The benchmark table

AWS publishes a set of scores for Grok 4.6 High as of August 2026: AA Intelligence Index 61, GDPVal-AA v2 1753, CursorBench v3.2 69.9%, DeepSWE v1.1 65.9%, FrontierCode v1.1 (Extended) 61.3%, APEX-Agents 57.5%, Terminal-Bench v3.0 26%, APEX-SWE 56.4%, AA-Briefcase 1577, and Harvey LAB (Vals) 15.8%.

Read those with care. The post does not put competing models in the same table, so the numbers establish where Grok 4.6 sits on each benchmark but not how it ranks. The spread is also instructive: strong results on code-editing benchmarks like CursorBench and DeepSWE sit alongside 26% on Terminal-Bench v3.0 and 15.8% on the Harvey legal benchmark. Agentic and domain-specialist work remains much harder than producing a correct patch, and a model marketed for long-running agents scoring in the mid-20s on a terminal benchmark is the kind of detail that should shape expectations before a pilot, not after.

Note also that these are scores for the High reasoning setting. Teams running at low or medium effort to control cost should not assume the same results.

Two endpoints, two deployment stories

The endpoint split is the part most likely to trip people up. Through bedrock-mantle, the model ID is xai.grok-4.6 and in-region availability is US West (Oregon) only. Through bedrock-runtime, the IDs are us.xai.grok-4.6 for the US geographic profile or global.xai.grok-4.6 for the global profile, and AWS offers no in-region inference at all. The US geo profile covers us-east-1, us-east-2, us-west-1, and us-west-2. The global profile spans more than 30 regions across the US, Canada, Europe, Asia Pacific, the Middle East, Africa, and South America.

Feature support differs between the two. AWS lists bedrock-mantle as supporting client-side tool calling, reasoning, structured outputs, prompt caching, response streaming, projects, and abuse detection. bedrock-runtime supports reasoning, prompt caching, response streaming, invocation logs, projects in default-only mode, and tool calling via Converse or the OpenAI tools parameter.

For anyone with data residency obligations, that asymmetry is the decision point. Structured outputs on mantle but a single in-region option, versus wide regional reach on runtime with a narrower feature set, is a real trade-off rather than a formality.

Authentication follows the current Bedrock pattern: long-term Bedrock API keys for exploration, short-term bearer tokens via aws-bedrock-token-generator for production, or AWS SigV4 through boto3 and IAM credentials. The OpenAI SDK works against both endpoint types. Required IAM permissions include bedrock:InvokeModel and bedrock:CallWithBearerToken.

Pricing and what it implies

AWS notes that xAI lists Grok 4.6 pricing starting at $2 per million input tokens and $6 per million output tokens. Bedrock layers three service tiers on top: Standard at a 1.0x multiplier by default, Priority at 1.75x, and Flex at 0.5x. Cross-region rates differ by geography. Prompt caching is billed at roughly 25% of the standard input rate, and reasoning tokens are billed separately.

The Flex tier at half price is the interesting line item for agent workloads. Long-running agents are often latency-tolerant in a way that chat is not, and a 50% discount for accepting variable scheduling changes the arithmetic on batch-style agent runs considerably.

The reasoning-tokens-billed-separately note deserves attention too. With a four-level effort dial and separate billing for reasoning output, the cost of a single request can vary by a wide margin depending on a setting that is easy to leave at a default. Any team putting this into production should instrument reasoning token spend per endpoint before it scales.

What to watch

• Whether AWS brings bedrock-mantle in-region availability beyond US West (Oregon), which currently constrains the feature-complete path.
• Real cost per completed agent task at each reasoning effort level, since reasoning tokens bill separately and the effort dial has four settings.
• Independent benchmark comparisons placing Grok 4.6 against peer models, which this announcement does not provide.
• Whether the Flex tier's 0.5x multiplier holds up for long-running agents in practice, or whether scheduling variability erodes the saving.
• Migration friction for existing Bedrock integrations written against the Invoke API, which Grok 4.6 does not support.