Grok 4.6 Matches GPT-5.6 Sol at Half the Price and Ships Straight Into Cursor. Here's How I'd Actually Route Between the Frontier Models Now.
Grok 4.6 Matches GPT-5.6 Sol at Half the Price and Ships Straight Into Cursor. Here's How I'd Actually Route Between the Frontier Models Now.
On August 12, SpaceXAI (the company most of us still call xAI out of habit) shipped Grok 4.6 at $2 per million input tokens and $6 per million output tokens, scored it at 61 on the Artificial Analysis Intelligence Index, and put it live in Cursor the same day. That score ties GPT-5.6 Sol exactly and sits one point behind Claude Fable 5's 62. Four labs are now clustered inside a two-point band on the same composite benchmark, and the model that's tied for second place on quality costs roughly half what the others charge. If your app still hard-codes one vendor's API, this is the week that decision starts costing you money you don't need to leave on the table.
What actually shipped
Grok 4.6 is xAI's follow-up to Grok 4.5, and the company's own announcement is specific about what it's for: long-running agent tasks, multi-step coding work, and "more ambitious interactive and visual work" than the previous model handled well. That's not vague marketing filler, it's a description of exactly the workload a solo operator's AI features usually hit: an agent that has to hold context across a dozen tool calls without losing the thread. On xAI's published eval table, Grok 4.6 posted 61 on the Artificial Analysis Intelligence Index (a composite of nine benchmarks including GDPVal-AA, DeepSWE, CursorBench, and FrontierCode), matching GPT-5.6 Sol's 61 and landing one point under Claude Fable 5's 62. It also beat both rivals outright on a couple of individual sub-benchmarks, like CursorBench 3.2, where it posted 69.9% against GPT-5.6 Sol's 67.2%.
Pricing is $2 per million input tokens and $6 per million output tokens for prompts under 200,000 tokens (it steps up to $4/$12 above that threshold), with a faster variant available at double the base rate. It's live today in Cursor, in xAI's own Grok Build tool, through the xAI API, and via OpenRouter, Vercel, and Cloudflare. For the first week, xAI is doubling included usage inside Cursor and Grok Build, which is a straightforward move to get developers to actually try it during a launch window instead of just reading about it.
One correction worth flagging up front, because it's an easy detail to get wrong: xAI the corporate entity really did rebrand. SpaceX acquired xAI in February 2026, took SpaceX public in June at a valuation north of $1.8 trillion, and folded the AI company in as a subsidiary renamed SpaceXAI on July 6. The Grok brand itself didn't change: the chatbot, the apps, SuperGrok, and the developer API all still say "Grok." So "Grok 4.6" is still the right name for the model, it's just now made by SpaceXAI instead of xAI, the same way X is now a subsidiary of SpaceXAI rather than the other way around.
Why the price matters more than the benchmark
A one-point gap on a composite index is close enough to be statistical noise on any single task. What isn't noise is that Grok 4.6 is priced at roughly half of what GPT-5.6 Sol and Claude Fable 5 cost per token, while landing in the same tier on the benchmark that's supposed to summarize "how smart is this thing." For a solo operator, that's the number that actually shows up on your invoice at the end of the month. If you're running an agent feature that burns tokens on tool calls and context re-reads (which most agentic coding and research workflows do), swapping the model underneath doesn't change what the feature does. It changes your margin on that feature by half, on day one, for work that already tests as comparable quality.
I route a decent chunk of my own agentic coding tasks through Claude right now, mostly because I built the harness around it first and haven't had a reason to move. But when I'm doing something narrower, like batch-summarizing scraped content or running a classification pass across a few hundred rows, I've already been sending that to whichever model is cheapest that clears my accuracy bar, because there's no reason to pay frontier rates for a task that doesn't need frontier reasoning. Grok 4.6 at $2/$6 just added a genuinely competitive third option to that rotation, not a token-desperate discount option you route to reluctantly.
The routing reality nobody's default setup accounts for
Here's the part that actually changes how you should build. With four models this close (GPT-5.6 Sol, Claude Fable 5, Grok 4.6, and whatever open-weight model is leading that particular week), "which model is smartest" stopped being the interesting question months ago. The interesting question is per-task cost and latency, because the quality gap between the top four is now small enough that it rarely decides the outcome on its own.
That argues for treating model choice as a routing decision, not a single hard-coded vendor call baked into your codebase. At minimum, that means an abstraction layer: one function that takes a task type and returns a model, instead of import { openai } from ... scattered across six files. At the more deliberate end, it means an actual router, something that sends cheap, high-volume, low-stakes calls to whichever model is currently cheapest at acceptable quality, and reserves the expensive models for the calls where getting it wrong is actually costly. You don't need a sophisticated ML-based router for this. A config table mapping task type to model, updated by hand every time a new release changes the math, gets you most of the value for an afternoon of work.
The failure mode I keep seeing in other people's codebases (and, until recently, my own) is the opposite: a single model string set once during the MVP build and never revisited, because revisiting it means finding every call site and re-testing each one. That's the tax you pay for not building the seam up front, and it's exactly what turns a genuinely good pricing move like this one into something you read about instead of something you benefit from.
What I'd actually do
If you're running any AI feature with real volume, spend the afternoon building the abstraction layer if you don't have one, and then run your own eval, not xAI's benchmark, against Grok 4.6 for whatever task currently costs you the most in tokens. Pull twenty to fifty real examples from your production traffic, run them through both your current model and Grok 4.6, and compare outputs by hand. If the quality holds on your workload, not the aggregate benchmark, but your specific prompts and your specific edge cases, route that task to Grok 4.6 and keep the difference. If it doesn't hold, you've lost an afternoon and learned something real about where the benchmark score and your reality diverge. Either outcome is more useful than trusting a launch-day chart.
The honest counter-take: benchmark parity is not behavioral parity, and this is exactly where that gap tends to bite. An Artificial Analysis Intelligence Index score of 61 versus 61 tells you almost nothing about which model calls your specific tools more reliably, handles your particular JSON schema without drifting, or holds up on the one weird edge case your domain throws at it constantly. Tool-calling reliability, in particular, is notorious for varying between models in ways that don't show up on general benchmarks at all. I'd also flag that a week-one launch promotion (double usage, aggressive marketing framing, Elon Musk calling it a "banger" on X) is not the same as a stable, independently verified read on the model's real-world performance. Give it a couple of weeks for CursorBench-style community numbers to settle before you bet a production workload on it. But the underlying argument holds regardless of how Grok 4.6 specifically performs on your tasks: with four labs this close on capability, the operators who win the pricing war are the ones who can route to whichever model is cheapest without a weekend refactor, not the ones still calling one vendor's SDK directly from a dozen files.
Author
Lukas
@lukcombinator