Grok 4.6 reaches frontier parity at a lower price, then doubles the rate where its agents run

xAI released Grok 4.6 on August 12, and the headline is what did not change: the model still lists at $2 per million input tokens and $6 per million output, the same rate as Grok 4.5. What changed is where that model sits. Independent testing now places Grok 4.6 level with OpenAI’s GPT-5.6 Sol near the top of the intelligence chart, which makes it one of the least expensive proprietary models to reach that tier. The number worth reading the rest of the rate card for is 200,000: above that prompt size, xAI’s own pricing doubles.

What shipped

Grok 4.6 is a point release over Grok 4.5, which xAI positions for long-running agents and more ambitious interactive and visual work rather than as a raw capability jump. On Artificial Analysis’s independent Intelligence Index it scores 61, in line with GPT-5.6 Sol and a step behind Anthropic’s Claude Opus 5 at 63 and Claude Fable 5 at 62. xAI reports gains over 4.5 on agentic-coding benchmarks, citing 65.9 percent on DeepSWE and 69.9 percent on CursorBench, and pairs the model with a 500,000-token context window.

Availability was broad on day one. Grok 4.6 is live in Cursor and xAI’s own Grok Build, through the API at console.x.ai, and via partners including OpenRouter, Vercel, and Cloudflare, with a first-week bonus of 2x included usage in Grok Build and Cursor. For a developer, that breadth matters more than the version number: the model can be routed to from the same gateways already carrying GPT and Claude traffic, so testing it against an existing workload is a config change rather than an integration project.

The catch above 200K tokens

xAI’s pricing page splits Grok 4.6 into two bands by prompt size. Below 200,000 tokens, the rate is the advertised $2 input, $0.50 cached input, and $6 output per million. At or above 200,000 tokens, every one of those figures doubles, to $4, $1, and $12. The higher band is not a surcharge on the overflow above the threshold. It reprices the whole request.

That structure sits awkwardly next to the model’s own pitch. Long-running agents are exactly the workloads that accumulate context: a coding agent that keeps a large working set, a research loop that appends every step, a session that carries a long history. Those are the runs most likely to cross 200K tokens, and once they do, the effective rate is $4 and $12, not $2 and $6. The model marketed for long tasks is the one whose price moves when the task gets long. That does not erase the value, but it does mean the headline number describes the short-context case, and the agent case needs its own math.

Where it lands against the frontier

At the sub-200K rate, Grok 4.6’s argument is the one Grok 4.5 made, now against a stronger field. Artificial Analysis puts the $2/$6 card at more than 60 percent below Claude Opus 5 while landing on the same intelligence tier as GPT-5.6 Sol, whose flagship lists far higher per token. That is a real wedge for teams that want frontier-class reasoning without flagship pricing.

The qualifier is that “cheapest” depends on where you draw the line. Among proprietary frontier models, Grok 4.6 is priced aggressively. Against open weights it is not the floor: DeepSeek’s recent V4 line undercuts it on raw token cost and can be self-hosted, at the cost of running the infrastructure yourself. Grok’s specific pitch stays proprietary-frontier-at-lower-cost, with real-time access to X data as the differentiator, rather than the absolute cheapest tokens available.

What it means for developers

The practical read is to route by task and to price the two bands separately. For agentic coding and reasoning that stays under 200K tokens of context, Grok 4.6 is now a credible frontier option at a fraction of GPT-5.6 Sol’s rate, worth benchmarking against whatever model a team currently defaults to. For long-context agent loops, the comparison narrows: at $4 and $12 the gap to the incumbents shrinks, and the decision should be made on the doubled numbers rather than the headline.

The standing caveats from earlier Grok releases still apply. Independent coding evaluations beyond the intelligence index are thin, xAI’s benchmark figures are its own, and the brand and platform coupling with X remains a procurement question for some organizations.

What’s worth watching

The near-term signals are whether independent coding benchmarks confirm the agentic gains xAI claims, and whether the 200K pricing band holds as agent workloads, the use case xAI is selling, push more traffic across it. If rivals answer the $2/$6 tier, the frontier-at-a-discount position gets harder to hold; if xAI relaxes the long-context penalty, the agent pitch and the rate card would finally point the same way. Stackmaven’s next look at Grok will fold independent agent and coding evaluations in alongside the pricing picture, on or around the November follow-up.

Sources cited

1. xAI: Grok 4.6 x.ai
2. xAI docs: Models and pricing docs.x.ai
3. Artificial Analysis: Grok 4.6 benchmarks and analysis artificialanalysis.ai
4. Vercel: Grok 4.6 now available on AI Gateway vercel.com