Grok
Grok 4.6 — xAI
xAI's real-time, no-filter flagship, retooled for coding. 500K token context and $2/$6 per million tokens — still the cheapest frontier model on this site. Grok 4.6 keeps every spec from 4.5 and spends the upgrade entirely on capability.
Released
Aug 12, 2026
Context
500K tokens
Starting Price
$30/mo (SuperGrok)
API Pricing
$2 / $6
01
Overview
Grok is xAI's flagship AI assistant, built with a uniquely aggressive, unfiltered personality. Unlike other AI models that lead with corporate disclaimers, Grok gives you direct answers — sometimes with a sense of humor. Access it at grok.com or through the X (formerly Twitter) platform.
Grok 4.6 landed August 12, 2026, four weeks after 4.5 and on the same 1.5T-parameter foundation. The spec sheet is unchanged — 500K context, $2/$6, text and image in, text out — and the entire upgrade went into post-training. Its lineage still runs through Cursor, which SpaceX agreed to buy for $60 billion in June after merging with xAI in February, and whose developer sessions trained the 4.5 line.
Read the spec sheet honestly: the 4.x line traded breadth for coding. Context halved from Grok 4.3's 1M to 500K back at 4.5, and native video input is still gone. What you get back is a model that inspects repos, edits files, runs commands, and retries after errors. The Artificial Analysis Intelligence Index has climbed from 38 on 4.3 to 61 on 4.6.
The pitch is still price. At $2 / $6 per million tokens with cached input at $0.50, Grok 4.6 is the cheapest way to run a frontier-tier agent. Don't read the benchmarks as a sweep, though: 4.6 leads GDPval-AA and legal work, while Claude Fable 5 takes most coding rows and GPT-5.6 Sol takes Terminal-Bench. It trades leads rather than dominating — and Artificial Analysis clocks it at 59 tokens/sec against a 78 median, so it isn't the fast one either.
02
Key Capabilities
500K Token Context
500,000-token window — half of Grok 4.3's, unchanged from 4.5, still enough for most real codebases
Agentic Coding
Inspects repositories, edits files, runs commands, and retries after errors. The 4.x line was trained on real Cursor developer sessions
Three Reasoning Levels
Configurable effort: low / medium / high, defaulting to high. Dial down for latency-sensitive work
Real-Time X/Twitter Data
Native access to live public posts and trends
DeepSearch
Extended research mode scanning multiple web and X sources with citations
Strong Terminal Benchmarks
Leads GDPval-AA v2 (1753 Elo), AA-Briefcase and Harvey LAB; 69.9% CursorBench v3.2, 65.9% DeepSWE v1.1. Function calling, web search, X search, and code execution built in
Custom Voices & Voice Library
Clone a voice from a short recording and manage it via the xAI console, now paired with the grok-voice-think-fast-2.0 speech-to-speech stack
Aggressive Pricing
$2 / $6 per million tokens, $0.50 cached input. The cheapest frontier-tier model per token by a clear margin
03
Pricing
| Tier | Price | Details |
|---|---|---|
| Free | $0 | ~10 prompts / 2 hours on Grok 4 and older. No 4.5, no DeepSearch, no image generation. |
| SuperGrok Lite | $10/mo | Entry paid tier. Grok Imagine plus one AI agent. No 4.5 or 4.6. |
| SuperGrok | $30/mo ($300/yr) | 100 default, 30 Think, 30 DeepSearch requests / 2 hrs. Running Grok 4.5 — 4.6 is API-only so far. |
| X Premium+ | $40/mo | High usage caps + X social features. Also on 4.5, not 4.6. |
| SuperGrok Heavy | $300/mo | Largest quotas and Grok 4 Heavy. Still 4.5 like every other consumer tier. |
| API (Grok 4.6) | $2 / $6 per 1M tokens | Input / output under 200K tokens. Cached input $0.50/M. Also on Cursor, Grok Build, OpenRouter. |
| API (>200K tokens) | $4 / $12 per 1M tokens | Requests past 200K total tokens double to extended-context rates. |
04
Best Use Cases
Real-time news and social media analysis
Financial forecasting and market analysis
Agentic coding on a budget — repo edits, command runs, retries
Complex multi-perspective problem analysis
Tasks benefiting from live X/Twitter data
Getting unfiltered, direct answers without corporate hedging
05
Pro Prompts
Real-time Analysis
Give me the unvarnished latest on [current event]. Skip the caveats; I want signal, not noise.Financial Forecasting
Analyze the current market sentiment around [stock/sector] using real-time X data. Identify contrarian signals most analysts are missing.Research Deep Dive
Use DeepSearch to compile everything published in the last 48 hours about [topic]. Rank sources by credibility and highlight conflicting claims.Agentic Tool Calling
Set reasoning effort to high. Use whatever tools you have to check current pricing on [product] across [retailers], compare specs, and produce a buy/wait recommendation with citations.Repo Work
Reasoning effort: high. Work through this repo end to end: find every place [pattern] is used, change it to [new pattern], run the test suite, and fix what breaks. Report only what you changed and what still fails.06
Limitations
Slow for its class — about 59 tokens/sec against a 78 median, and somewhat verbose with it
Context is 500K, down from 4.3's 1M, and native video input was removed at 4.5 — if you built on either, 4.3 is still the model you want
The model itself can't generate images — that's Grok Imagine, and it still lags Nano Banana for typography-heavy work
Requests past 200K total tokens double to $4/$12 — the cheap headline rate only holds on shorter contexts
Consumer access lags the API badly: 4.6 launched on developer surfaces only, so grok.com, the mobile apps and Grok-in-X are all still serving 4.5
Claude Fable 5 still takes most coding benchmarks outright — 4.6 wins knowledge work and legal, not the coding sweep
07
Multimodal Capabilities
Input
- Text
- Images (native vision encoder)
- Video — removed in 4.5 (4.3 only)
Output
- Text (no token output limit)
- Image generation (paid only, via Grok Imagine)
- Voice (grok-voice-think-fast-2.0 + Custom Voices)
Medical document analysis via photo upload
Can process circuit diagrams, design briefs, and visual inputs alongside text
08
Recent Updates
Aug 12, 2026
Grok 4.6 shipped — same 500K context, same $2/$6, same 1.5T-parameter base as 4.5. The whole upgrade is post-training. Developer surfaces only: API, Cursor, Grok Build and partners; consumer apps deferred.
Aug 5, 2026
grok-voice-think-fast-2.0 with speech-to-speech began routing through grok-voice-latest
Jul 8, 2026
Grok 4.5 announced and shipped — 500K context, three reasoning levels, $2/$6 pricing, built for coding and agents
Jun 16, 2026
SpaceX agreed to acquire Cursor for $60B in stock — the largest venture-backed startup acquisition on record. Grok 4.5's training used real Cursor session data
Feb 2026
SpaceX merged with xAI — Grok now ships under the SpaceXAI banner
Heads-Up
Grok 4.6 is API-only for now — grok.com, the mobile apps and Grok-in-X are all still on 4.5. Grok 4.7 is rumoured for early September; it has no model ID yet, so it isn't on this page.
Milestone
Artificial Analysis Intelligence Index has gone 38 (4.3) → 56 (4.5) → 61 (4.6), now level with GPT-5.6 Sol