Grok 4.6 is xAI's current flagship model, released August 12, 2026, and available through the Grok app, X, and the API under the model name grok-4.6. It keeps Grok 4.5's exact pricing — $2 per million input tokens and $6 per million output tokens under 200K prompt tokens — while jumping five points on the Artificial Analysis Intelligence Index.
Short answer: Grok 4.6 is xAI's flagship as of August 2026, with a 500K-token context window and API pricing unchanged from 4.5 at $2/M input and $6/M output tokens under 200K tokens. It scores 61 on the Artificial Analysis Intelligence Index (rank #6 of 182), roughly tying GPT-5.6 Sol. Cached-input pricing rose from $0.30/M to $0.50/M.
In my testing, I pulled the specs below from xAI’s own launch post and its model docs, then cross-checked the benchmark numbers against Artificial Analysis’s independent test run rather than trusting a single aggregator, because xAI's version churn (4.3, 4.5, 4.6, and an already-teased 4.7) makes it easy to quote figures for the wrong release by accident. Here's what changed from 4.5, what it costs, and the exact steps to start using it.
What you'll need
For the consumer app, a Grok or X account gets you in — the free tier includes basic chat, though new model access rolls out to free accounts on a delay, so don't assume 4.6 shows up in your picker the moment it's announced. For the API path, you need an account at console.x.ai, a payment method on file, and a sense of your token volume: at $2/M input tokens, a 10,000-word document (about 13,000 tokens) costs roughly 3 cents to send once, before any output. I checked x.ai's own pricing page for the current SuperGrok subscription cost while writing this and it loaded fine this time, unlike when I researched 4.5 last month — worth a direct look yourself since consumer tiers shift more often than API rates.
Step-by-step: Grok 4.6
1. Decide between the app and the API
Chatting doesn't need code — open the Grok app or X and pick 4.6 from the model selector. Building something, or feeding it a long document, means the API, which gives you the full 500K-token context window and per-token billing you can predict in advance.
2. Confirm the model picker actually says 4.6
xAI staged the 4.5 rollout across account tiers rather than flipping it on for everyone at once, and 4.6 has followed the same pattern in my testing. If your picker still shows 4.5 or an older 4.3 build, none of the benchmark numbers in this article apply to what you're actually running — check the label, not the release date headline.
3. Get an API key if you're building anything
Sign in at console.x.ai, generate a key, and call the model as grok-4.6. Its knowledge cutoff is still February 1, 2026 — the same as 4.5 — so anything more recent needs a search or browsing tool attached to the request, or it'll answer confidently from stale information.
4. Set the reasoning effort level
Grok 4.6 exposes four effort levels — low, medium, high (the default), and xhigh — the same lever Anthropic and OpenAI ship on their current flagships. In my testing, high made a visible difference on a multi-file coding task versus low, but at a real cost: Artificial Analysis clocked its time-to-first-token at 43.74 seconds on the high setting, well behind Grok 4.5's response times, so a chat-style app that needs snappy replies should default lower.
5. Watch the 200K-token pricing cliff and the cache rate
Input and output pricing double to $4/M and $12/M once a request crosses 200K prompt tokens — the whole request, not just the overage. And if you're relying on prompt caching for a repeated system prompt, note that xAI raised the cached-input rate from 4.5's $0.30/M to $0.50/M with this release; still a real discount off $2/M fresh input, just a smaller one than before.
Example prompts you can copy
These target what xAI says 4.6 is built for — long-running, multi-step agent work — rather than one-shot chat requests:
- "Here's a 15-file codebase [paste or attach]. Find every place a shared config value is read, and list which ones would break if I renamed it." (Tests sustained reasoning across a multi-step task, not a single-file lookup.)
- "Research [topic] using search, then draft a 500-word brief citing at least three distinct sources with dates." (Uses search grounding to get around the February 2026 cutoff.)
- "Here's a rough app idea [description]. Turn it into a working single-page prototype, then list the three things you'd fix before shipping it." (Plays to the "interactive and visual work" focus xAI called out at launch.)
- "Set reasoning effort to low and answer this, then answer it again at high. Tell me what changed." (A quick, cheap way to see whether the effort toggle is worth the latency for your use case.)
- "Summarize this 40-page document, then flag every number in your summary that you're not fully confident about." (Tests both the 500K context window and whether it hedges honestly instead of guessing.)
Common mistakes to avoid
The first mistake I'd flag: assuming a subscription that had 4.5 automatically has 4.6 the day it's announced. The rollout is staged, and I've seen the model picker lag the launch post by more than a week on some accounts. Second, quoting the $2/$6 rate for every request without checking prompt length — it doubles past 200K tokens, and that's easy to miss on a long document you didn't count first. Third, forgetting the cached-input price changed; if you built cost projections around 4.5's $0.30/M cache rate, they're now understating your bill by roughly 67% on cached tokens. Fourth, defaulting to xhigh reasoning effort for routine requests — the 43-second-plus time to first token on high alone is a real tax on anything interactive, and xhigh will run slower still. Fifth, expecting current-events knowledge without attaching a search tool; the cutoff didn't move from 4.5's February 1, 2026 date, so a plain API call will guess rather than admit it doesn't know.
Grok 4.6 vs. the other current flagships
| Model | Input $/M tokens | Output $/M tokens | Context window | AA Intelligence Index |
|---|---|---|---|---|
| Grok 4.6 | $2.00 (under 200K) | $6.00 (under 200K) | 500K tokens | 61 — rank #6 of 182 |
| Claude Opus 5 | $5.00 | $25.00 | 1M tokens | 63 — rank #1 of 182 |
| GPT-5.6 Sol | $5.00 | $30.00 | ~1.05M tokens | 61 — rank #5 of 182 |
| Claude Sonnet 5 | $2.00 | $10.00 | 1M tokens | — |
Prices and scores confirmed against each vendor's own docs and Artificial Analysis on 2026-08-20. Grok 4.6 ties GPT-5.6 Sol's Intelligence Index score at less than half the per-token output cost, and it's the cheapest input rate of the three reasoning-focused flagships — Claude Sonnet 5 matches it on input but isn't positioned as a top-tier reasoner the way the other three are.
Tools that make this easier
If you're deciding whether 4.6 is worth adding next to what you already run, my how to use ChatGPT, how to use Claude AI, and how to use Gemini guides cover each assistant's own setup quirks, so you know what you're actually comparing it against. For coding specifically, best AI tool for code tests Grok's coding output head-to-head with Claude Code and Cursor, and ChatGPT alternatives for coding covers the wider field. I ran this same pricing-and-benchmark check on the previous release in Grok 4.5: pricing, benchmarks, and setup guide and on Anthropic's latest in Claude Opus 5: pricing, benchmarks, setup guide, useful if you're weighing vendors rather than just tracking one lineup. My starter kit for AI is the place to start if none of these accounts exist yet, and how we test AI tools explains how I verify numbers like the ones above before publishing them.
Where this leaves you
Grok 4.6 isn't a repricing move — xAI held the $2/$6 rate steady from 4.5 while pushing the Intelligence Index score up five points to 61, close enough to GPT-5.6 Sol's 61 that the two are effectively tied on reasoning quality, at a fraction of GPT-5.6 Sol's output cost. The catch is latency: a time-to-first-token north of 43 seconds on the default high effort setting is slow for anything a person is waiting on, and the cached-input price increase quietly raises the bill for anyone running the same system prompt at volume. If you're already on Grok, checking your model picker costs nothing. If you're building on the API, the math favors 4.6 for reasoning-heavy batch work and favors a faster model for anything interactive.
Last updated: August 20, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Frequently Asked Questions
Grok 4.6: is it free?
The Grok app has a free tier, but new model access typically rolls out to free accounts after paid tiers, so 4.6 may not appear in a free account's picker right away. The API isn't free at all — it's pay-per-token at $2/M input and $6/M output under 200K tokens, billed to a card on file at console.x.ai.
How long does it take to start using Grok 4.6?
Through the app, a couple of minutes — sign in and check whether the model picker shows 4.6 yet. Through the API, expect 10–15 minutes to create a console.x.ai account, generate a key, and send a first test call with grok-4.6 as the model string.
What is the easiest way to try Grok 4.6?
Open the Grok app or X, confirm the picker shows 4.6, and run a real multi-step task through it — a research brief or a small coding change — rather than a single generic prompt, since that's what xAI built this release for.
Does Grok 4.6 know about current events?
Not on its own. Its training data runs through February 1, 2026, the same cutoff as Grok 4.5, so anything after that needs a search or browsing tool enabled in the app or the API request.
How does Grok 4.6 compare to Claude Opus 5 and GPT-5.6 Sol on price?
It undercuts both on output cost by a wide margin — $6/M versus $25/M for Opus 5 and $30/M for GPT-5.6 Sol — while scoring within two points of Opus 5 and matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The tradeoff is speed: Artificial Analysis measured Grok 4.6 as slower to first token than either competitor at its default reasoning effort.