Grok 4.6 just got a second front door. A week after SpaceXAI shipped Grok 4.6 at $2/$6 per million tokens through its own API, Cursor, and Grok Build, the model landed as a selectable foundation model inside Amazon Bedrock — AWS's managed model marketplace, the same place enterprise customers already reach Claude, Llama, and Mistral. Nothing about the model changed. What changed is who can call it, and how.
That distinction is the whole story, and it's worth being precise about it before the "Grok is now on AWS" headline gets simplified into something it isn't.
TL;DR — what people are actually asking
| Question | Direct answer |
|---|---|
| Is this a new model? | No — same Grok 4.6 weights from the August 12 launch |
| Does it cost more on Bedrock? | No — same $2/M input, $6/M output list price |
| Why use Bedrock over the direct API? | AWS IAM auth, VPC-isolated traffic via PrivateLink, one consolidated bill |
| Do I need an xAI account? | No — Bedrock access is granted entirely inside the AWS console |
| Is Claude on Bedrock too? | Yes, and has been for longer — Grok joins an existing multi-model catalog |
| Does this replace the direct xAI API? | No — both exist in parallel; Bedrock is an additional access path |
| Is this useful if I'm not on AWS? | Not really — the direct API stays simpler if you have no other AWS dependency |
What "launching on Bedrock" actually means
Amazon Bedrock is not a hosting service that runs its own copy of a model — it's a managed API layer AWS operates in front of foundation models from multiple providers. When a model "launches on Bedrock," the provider (here, SpaceXAI) makes its weights callable through Bedrock's standard InvokeModel and Converse APIs, and AWS handles the surrounding plumbing: authentication, request routing, logging, and billing.
For a reader deciding whether this matters, the mechanical shift is:
- Before: call
api.x.aiwith an xAI-issued API key, billed directly by SpaceXAI. - Now, in addition: call Bedrock's endpoint in your chosen AWS region, authenticated with an AWS IAM role or policy instead of an xAI key, billed as AWS usage alongside every other Bedrock model you already run.
Both paths keep working. Bedrock is additive distribution, not a migration you're forced into — the same pattern SpaceXAI already used when Grok 4.6 landed simultaneously in Cursor, OpenRouter, Vercel, and Cloudflare at launch. Bedrock is simply the next partner route, aimed specifically at AWS-native enterprise buyers rather than individual developers.
Pricing: does Bedrock cost more or less?
List price is unchanged. Grok 4.6 on Bedrock runs at the same $2 per million input tokens / $6 per million output tokens SpaceXAI set for the direct API at launch, with the faster variant still priced at roughly 2x ($4/$12). AWS's standard model for third-party foundation models is to pass through the provider's own rate card rather than add a markup on top of it — the same arrangement Claude, Llama, and Mistral already have on Bedrock.
What can add cost is usage outside the raw token price: PrivateLink VPC endpoints, cross-region inference, CloudWatch logging retention, and Bedrock Guardrails all meter separately if you turn them on. None of that is mandatory — a team that just wants IAM-based auth and unified billing, with no VPC isolation, pays token price and nothing else.
What enterprise guarantees does Bedrock add that the direct API doesn't?
This is the actual reason to care about the launch, and it's worth being specific rather than gesturing at "enterprise-grade":
| Capability | Direct xAI API | Grok 4.6 on Bedrock |
|---|---|---|
| Authentication | Standalone xAI API key | AWS IAM roles/policies — same identity system as the rest of your AWS account |
| Network isolation | Public internet, unless xAI offers its own private link | AWS PrivateLink — traffic stays inside your VPC, never crosses the public internet |
| Data residency | Governed by xAI's own terms | Processed and encrypted at rest in your chosen AWS region |
| Compliance posture | Whatever xAI documents independently | Inherits Bedrock's existing SOC 2, ISO 27001, HIPAA-eligible, and FedRAMP Moderate scope |
| Training on your data | Governed by xAI's own policy | Bedrock's standing commitment: prompts and completions aren't used to train AWS or third-party models |
| Billing | Separate xAI invoice | One AWS bill, consolidated with every other Bedrock model and AWS service |
None of these are exotic Bedrock-only features — they're the same reasons enterprises already reach for Bedrock over calling Anthropic, Meta, or Mistral direct. Grok 4.6's arrival just means that calculus now applies to xAI's flagship model too. A regulated bank or healthcare team that couldn't get an independent security review of xAI done in time can now put Grok 4.6 behind the exact same IAM policies and audit trail it already uses for Claude.
Is this the same Grok 4.6 people compared against Fable 5 and Sol?
Yes — same benchmarks apply. At launch, Grok 4.6 High tied GPT-5.6 Sol Max at 61 on the Artificial Analysis Intelligence Index and trailed Fable 5 Max's 62, while leading GDPVal-AA v2 and AA-Briefcase. Bedrock doesn't run a separate fine-tune or a quantized variant — it's the identical model card. If you were holding off adopting Grok 4.6 because of those eval gaps against Fable 5 on coding-heavy work, Bedrock access doesn't change that decision — it only changes how you'd call it if you do adopt it.
Grok joins an existing multi-model catalog — it's not the first
The framing worth correcting: this isn't Bedrock's first frontier model from outside Amazon. Claude has been on Bedrock for years, and Anthropic's enterprise plugins already lean on that Bedrock relationship for financial-services customers who need the same IAM/VPC guarantees described above. Llama and Mistral sit on Bedrock too. What Grok 4.6's launch actually does is turn Bedrock into a genuine cross-lab marketplace — Anthropic, Meta, Mistral, and now SpaceXAI's Grok, side by side under one identity and billing system — rather than "Claude plus open-weight options."
For a team standardizing on AWS, that's the practical upside: switching your default model from Claude to Grok 4.6 for a specific workload, or running both behind a single router, no longer means standing up a second vendor relationship. It's a model swap inside infrastructure you already operate — closer in spirit to how Cursor Router already selects across Claude, GPT, and Grok models based on task type, just at the cloud-platform layer instead of the IDE layer.
Honest limitations
- No published Bedrock-specific benchmark. SpaceXAI's own eval table was published for the direct API; nothing suggests Bedrock changes model behavior, but no separate Bedrock-hosted benchmark run has been published to confirm identical latency or output at scale.
- Region availability isn't universal on day one. New Bedrock model launches typically roll out to a handful of AWS regions first, not the full regional footprint — check the Bedrock model catalog in the console for your specific region before planning around it.
- If you're not on AWS already, this adds nothing. The entire value proposition is "you already have AWS IAM, VPC, and billing infrastructure." A team with no other AWS dependency gains no benefit from routing through Bedrock instead of the direct API, and picks up Bedrock's own API shape (
Converse/InvokeModel) to learn instead. - Guardrails and logging cost extra. The compliance features listed above are opt-in, metered add-ons (PrivateLink endpoints, Bedrock Guardrails, CloudWatch log retention) — the base token price alone doesn't include them.
What to actually do this week
If you're already running Claude or other models on Bedrock and want to test Grok 4.6 without touching your xAI relationship at all: request model access in the Bedrock console, point an existing Converse API call at the Grok 4.6 model ID, and compare output against your current default on a task you already have eval coverage for. If you're not on AWS, the direct xAI API remains the simpler path — Bedrock buys you nothing without the surrounding AWS account to plug into.
Related on explainx.ai
- Grok 4.6 launch — official evals and day-one Cursor access
- SpaceX closes Cursor acquisition — the Grok/SpaceXAI branding explained
- Fable 5 vs Grok 4.6 vs GPT-5.6 Sol vs Qwen3-8-Max — full comparison
- Anthropic's Claude financial-services agents and plugins — Bedrock enterprise access in practice
- Perplexity adds Grok 4.6 — verified cost comparison against Fable 5
- Cursor Router — automatic model selection across Claude, GPT, and Grok
- AWS Certified Generative AI Developer exam — what it covers
- GitHub Copilot adds Grok 4.6 across CLI, IDE, and cloud
Primary sources: Amazon Bedrock model catalog documentation (aws.amazon.com/bedrock) and SpaceXAI's Grok 4.6 announcement (x.ai/news/grok-4-6).
Facts, pricing, and availability are accurate as of August 19, 2026. Bedrock regional rollout and any Bedrock-specific pricing changes may evolve after publication — check the AWS Bedrock console for current model and region availability. Follow @explainx_ai for updates.