← Back to feed

Grok 4.6 is now available via Google Enterprise Agent Platform, xAI announced on August 21. Enterprise developers can access it through Model Garden — Google’s managed model catalog for agent workloads — at $2 per million input tokens and $6 per million output tokens. Cached input drops to $0.50 per million.

The model brings a 500K token context window and four configurable reasoning effort levels: low, medium, high, and xhigh. xAI positioned this as purpose-built for long-running agents and demanding interactive and visual tasks.

The Context

The same pricing appeared two weeks earlier when Grok 4.6 hit GitHub Copilot on August 14. What’s different here is the infrastructure layer. GitHub Copilot targets developers in an IDE. Google Enterprise Agent Platform is a runtime environment — orchestration, tool access, and deployment pipelines for production agents inside enterprise Google Cloud accounts. The Model Garden listing puts Grok 4.6 alongside Gemini, Claude, and other catalog models as a selectable runtime for customers already committed to GCP.

For xAI, this is the second major cloud distribution agreement within a week, and the first with Google. The practical significance: enterprises that have standardized on Google Cloud infrastructure can now deploy Grok 4.6 without leaving their existing stack.

Where Grok 4.6 Sits

On capability benchmarks, Grok 4.6 scores 95.6% on SWE-bench Verified under the neutral bash-only harness, ranking fourth globally — behind Claude Opus 5 (97.0%), DeepSeek V4 Pro 0813 (96.4%), and GPT-5.6 Sol (96.2%). Artificial Analysis puts its Intelligence Index at 61, matching GPT-5.6 Sol.

At $6/M output, Grok 4.6 sits well below Claude Opus 5 ($75/M) and GPT-5.6 Sol ($30/M), and within a dollar of DeepSeek V4 Pro’s API pricing. For enterprise workloads where output volume is the cost driver, the combination of frontier-tier capability and mid-range pricing makes it a plausible default selection in agent pipelines.

Reasoning Effort

The four-level reasoning configuration (low through xhigh) gives buyers direct control over latency and cost per call. Low-effort calls run faster and cheaper; xhigh pushes toward maximum capability at higher token and time cost. This is the same architecture pattern Anthropic uses with extended thinking tiers and that OpenAI uses with high/medium/low effort on its reasoning models.

For production agent systems that mix reasoning requirements across task types — a common pattern in enterprise agentic workflows — this configurability matters operationally. A triage step can run on low; a multi-step reasoning chain can escalate to xhigh without switching models.

Distribution Pattern

xAI has now placed Grok 4.6 into two of the largest enterprise developer platforms within a week of each other. Both Microsoft’s GitHub Copilot and Google’s Enterprise Agent Platform represent distribution at scale that the xAI API alone cannot reach. The pace suggests a deliberate co-launch strategy: announce a model, then push it into partner platforms before the news cycle closes.

Whether that pattern continues — and what model follows Grok 4.6 — is the relevant next data point.