Grok 4.6: We Ran It Through Our Stack, and the Cost Story Is the Real Story

xAI released Grok 4.6 on 12 August 2026, and we had it wired into a Hermes subagent within the day. This is not a "here's what the headlines say" post. We routed real workstreams through it - inspection report drafting logic for a client project, some agentic coding tasks, and a chunk of our usual Confluence and repo triage work - and the model held up. But the part that actually changed how we're thinking about model routing wasn't the intelligence benchmarks. It was the invoice.

What's New, Briefly

Grok 4.6 is xAI's latest flagship, positioned squarely at coding, agentic tasks, and knowledge work rather than as a consumer chatbot refresh. It carries a 500,000-token context window, unchanged from Grok 4.5, and supports both the Responses API and Chat Completions. Reasoning effort is configurable across low, medium, high (the default), and xhigh, which matters if you're trying to tune latency against depth for a given subagent.

On xAI's own benchmark table it lands on par with GPT-5.6 Sol Max on the Artificial Analysis Intelligence Index, with the largest gains over Grok 4.5 showing up on DeepSWE, Terminal-Bench, and agentic evaluation suites like APEX-Agents. Musk's framing on release was blunt: a step-change on intelligence, speed, and cost simultaneously. We'd normally discount marketing copy like that. In this case the cost claim actually checked out.

The Part That Matters: Pricing

Below a 200K-token prompt, Grok 4.6 is priced at USD $2 per million input tokens, USD $0.50 per million cached input tokens, and USD $6 per million output tokens. Cross the 200K-token threshold and the entire request reprices to $4 / $1 / $12, so anything genuinely leaning on that long context window needs to be budgeted with the higher tier in mind, not the headline rate.

For context, that headline rate sits at roughly half the sticker price of comparable frontier output from OpenAI and Anthropic's top tiers, while scoring competitively on the benchmarks that matter for agentic and coding work. We've been running DeepSeek V4 Flash as our default routing tier in Hermes precisely because frontier models have historically been too expensive to run as a default rather than an escalation. Grok 4.6 is the first model we've tested that makes us want to reconsider that assumption for mid-complexity agentic tasks, where DeepSeek's ceiling was starting to show and the frontier alternatives were previously priced well outside default-tier territory.

Two pricing details worth knowing before you build against it:

• No batch discount. xAI's 20% Batch API discount covers Grok 4.3 and the dated Grok 4.20 SKUs. It does not extend to Grok 4.6. If your workload is the kind of overnight, batchable job that would normally chase that discount, you're paying full rate on the flagship regardless.
• Priority processing is a flat 2x multiplier across input, output, cached, and reasoning tokens. You're only billed at the priority rate when the API response confirms "service_tier": "priority", so it's worth checking that field rather than assuming you've paid for speed you didn't get.

Where It Actually Held Up in Our Testing

We pointed Grok 4.6 at three categories of task rather than running it against generic benchmarks:

Agentic coding. Repo-pass style work, the kind we've been running through Claude Code, performed credibly. The 500K context window comfortably held larger portions of our codebases than the smaller-context tiers we'd normally reach for, which cut down on the retrieval scaffolding we'd otherwise need to build.

Long-document reasoning. For the kind of inspection report and SOW-adjacent document work we do for clients, having the full context in view without chunking produced more internally consistent output than a fragmented read would have.

Cost-sensitive triage. This is where it genuinely surprised us. For subagent tasks that previously sat in an awkward middle tier, too complex for our cheapest routing option but not complex enough to justify frontier pricing, Grok 4.6's rate makes it a plausible default rather than an escalation-only option.

Why We'd Default to the ZDR Endpoint

xAI runs two API endpoints for Grok 4.6 at identical pricing: a standard endpoint and a zero-data-retention (ZDR) endpoint. There is a small measured difference in throughput and latency between the two, so if your workload is genuinely latency-critical it's worth benchmarking both against your own traffic before committing.

For anything touching client data, and particularly for the kind of B2B integration work we do where the underlying content might include commercially sensitive material from a client's ServiceM8 instance, CRM, or internal documentation, we'd recommend defaulting to the ZDR endpoint as a matter of course rather than opting in only when a client explicitly asks. It costs nothing extra, and it removes a category of question you'd otherwise need to answer during a client's due diligence or a data handling review. Given the same rate applies either way, there's no real reason to run production traffic through the standard endpoint unless you've specifically tested that the throughput difference matters for your use case.

The Practical Take

Grok 4.6 is not a reason to rip out an existing model routing setup. But if you're running a multi-model architecture the way we run Hermes, with cost-tiered escalation across subagents, it's a genuine addition to the rotation rather than a headline to skim past. The intelligence is credible at the frontier tier, the context window is workable for real codebases and document sets, and the pricing is the first time we've seen "half the price of the incumbents" actually hold up once we ran our own numbers rather than trusting the launch post.

If you're building agentic workflows and want a second opinion on where a model like this fits into your routing strategy, or want the ZDR question handled properly before it becomes a client conversation, get in touch.

Related reading

• Kimi K3 has landed. Does it matter for your Sydney small business?
• AI agents for ServiceM8: Hermes and Openclaw

If you want help with this, see our AI automation service or get in touch.