Grok
Grok 4.5 — xAI
xAI's real-time, no-filter flagship, retooled for coding. 500K token context, three reasoning effort levels, and $2/$6 per million tokens — a price tag that undercuts every other frontier model on this site. Trained on real Cursor developer sessions, and it shows on agentic coding work.
Released
Jul 8, 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 4.5 landed July 8, 2026 and is the first xAI model built specifically for coding and agentic work. It ships configurable reasoning effort (low / medium / high, defaulting to high) and a 500K token context window. Training folded in real developer sessions from Cursor — which SpaceX agreed to buy for $60 billion in June, after merging with xAI back in February.
Read the spec sheet honestly: this is a trade, not a straight upgrade. Context halved from Grok 4.3's 1M to 500K, and native video input is gone. What you get back is a model that inspects repos, edits files, runs commands, and retries after errors — 83.3% on Terminal Bench 2.1 and an Artificial Analysis Intelligence Index of 54, up from 38 on 4.3.
The pitch is price. At $2 / $6 per million tokens with cached input at $0.50, Grok 4.5 is the cheapest way to run a frontier-tier agent right now. Musk's own framing was "roughly comparable to Opus 4.7, but much faster" — aim lower than that and you'll be pleasantly surprised.
02
Key Capabilities
500K Token Context
500,000-token window — half of Grok 4.3's, still enough for most real codebases
Agentic Coding
Inspects repositories, edits files, runs commands, and retries after errors. 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
83.3% on Terminal Bench 2.1 and 64.7% on SWE-bench Pro. 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 Grok 4.5. |
| SuperGrok | $30/mo ($300/yr) | 100 default, 30 Think, 30 DeepSearch requests / 2 hrs. Grok 4.5 arriving in staged rollout. |
| X Premium+ | $40/mo | High usage caps + X social features. Also receiving Grok 4.5 in stages. |
| SuperGrok Heavy | $300/mo | The only consumer plan with guaranteed full Grok 4.5 access today. Largest quotas, Grok 4 Heavy included. |
| API (Grok 4.5) | $2 / $6 per 1M tokens | Input / output. Cached input $0.50/M. Tools $5 per 1K calls; file search $10 per 1K calls. |
| 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
Effort defaults to high, so first-token latency is worse out of the box than on 4.3 — drop to low or medium for chat-speed work
Context dropped from 1M to 500K, and native video input was removed entirely — 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 is a mess: only SuperGrok Heavy ($300/mo) guarantees Grok 4.5 today. Everyone else waits on a staged rollout
Not yet available in the EU, and it still trails Claude Fable 5 on SWE-bench Pro (64.7% vs 80.4%)
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 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 4.5 is still rolling out to SuperGrok and X Premium+ in stages, and isn't available in the EU yet
Milestone
Artificial Analysis Intelligence Index jumped from 38 (Grok 4.3) to 54, landing 4th overall