Grok AI: All Models Available: capabilities, context windows, pricing, and when to use each

Graziano Stefanelli
Oct 23, 2025
5 min read

xAI’s Grok stack now spans frontier reasoning, cost-efficient “fast” variants, code-focused models, and legacy VLM/image engines. For builders and researchers, the practical question is no longer “Does Grok do X?” but which Grok to deploy for a given job (reasoning, coding, retrieval, or vision) and at what price point. Below is a consolidated, source-backed map of every generally available Grok model family today—plus historical models you’ll still encounter in docs and repos.

·····

.....

The current catalog (API & product)

Model (alias)	Modality	Context window	Notable features	Throughput (typical)	Pricing (per 1M)*
grok-4-fast-reasoning	Text (reasoning)	2,000,000	Tool use, structured output, function calling	~480 rpm	$0.20 in / $0.50 out
grok-4-fast-non-reasoning	Text	2,000,000	Cost-efficient drafts; no reasoning tokens	~480 rpm	$0.20 in / $0.50 out
grok-4-0709 (aka grok-4)	Text (frontier)	256,000	Highest quality “full Grok 4”	~480 rpm	$3 in / $15 out
grok-3	Text	131,072	Strong enterprise tasks; extraction & coding	~600 rpm	$3 in / $15 out
grok-3-mini	Text	131,072	Lower cost, faster than Grok-3	~480 rpm	$0.30 in / $0.50 out
grok-code-fast-1	Text (code)	256,000	Agentic coding; tool use optimized	~480 rpm	$0.20 in / $1.50 out
grok-2-vision-1212 (us-east-1 / eu-west-1)	Text+Image in	32,768	Multimodal understanding of images	up to 600 rpm (region-dependent)	$2 in / $10 out
grok-2-image-1212	Image out	—	Image generation	~300 rpm	$0.07 / image

*Pricing, context, and rate limits are taken from xAI’s Models & Pricing documentation; regional variants and rps/rpm differ by endpoint.

Key clarifications from the docs: Grok-4 is a reasoning model (no separate “non-reasoning mode”); some completion parameters (e.g., stop, penalties) are not supported on reasoning models. Also, real-time knowledge requires tools—by default models do not have current-events awareness without Live Search.

·····

.....

Grok 4: the frontier tier (and why “Fast” matters)

Grok-4-0709 (aka grok-4) is the highest-quality baseline for complex synthesis, analysis, and instruction following. It carries the premium price point and a 256k context. For maximum quality, use this.
Grok-4-Fast (reasoning / non-reasoning) stretches the context to 2M tokens while radically cutting price and latency. It’s designed for cost-efficient reasoning and high-throughput automations where you need long context + tool use, not necessarily the absolute peak quality of full Grok-4.

xAI positioned Grok 4 publicly as its new flagship in July 2025 (API + consumer surfaces). Premium tiers on X and SuperGrok also advertise Grok-4 access.

·····

.....

Grok 3 family: strong generalist + a smaller “mini”

xAI introduced Grok-3 and Grok-3 mini as beta reasoning models with RL-enhanced “Think” modes; they became widely available across API and cloud hosts (including Azure AI Foundry). Grok-3 targets enterprise-style tasks—data extraction, programming, summarization—and sits at the same list price as Grok-4 (standard) but with a 131k window; mini drops cost for speed.

Even after Grok-4’s launch, grok-3 remains visible in the models table for compatibility and cost/perf tradeoffs. If you need reliable JSON extraction at moderate cost and don’t need the newest tool stack, Grok-3 is a sensible default.

·····

.....

Grok Code Fast 1: agentic coding (new)

Announced late August, grok-code-fast-1 is a “lightning-fast” reasoning model built for agentic coding: tool-augmented code generation, refactors, and environment-aware planning. It pairs a 256k window with low input cost and higher output cost (reflecting longer completions for code). For developer copilots and pipeline automation, it’s the value workhorse.

·····

.....

Vision & image models: Grok-2 VLMs remain in catalog

For multimodal pipelines, the current docs still list grok-2-vision-1212 (image understanding with 32k context) and grok-2-image-1212 (image generation at $0.07/image). These are stable, production VLMs useful when you need image inputs/outputs alongside Grok-3/4 text models.

Tip: If you need PDF page or chart analysis, route pages as images to the VLM or use Grok-4 with tools that extract page text + figures, then fuse results. Tools are billed separately (see below).

·····

.....

Tools, Live Search, and costs that show up on your bill

Grok’s agentic tool calling adds separate metering for server-side tools (Web Search, X Search, Code Execution, etc.). You’re charged per 1,000 tool calls in addition to tokens; Live Search has its own per-source pricing and is being superseded by the agentic tool-calling API later this year. If you rely on real-time grounding, budget for token + tool costs.

Tool pricing examples (from docs): Web Search and X Search $10 / 1k calls; Live Search $25 / 1k sources (deprecating by Dec 15, 2025).

·····

.....

Historical & legacy models you’ll still see referenced

Model	What it was	Why you’ll still see it
Grok-1 (open weights, MoE 314B)	Base model weights and architecture released openly in Mar 2024.	For research baselines, forks, and MoE studies.
Grok-1.5	Long-context (128k) upgrade over Grok-1 with stronger math/code.	Shows up in older benchmarks and migration notes.
Grok-1.5V	First multimodal preview (documents, diagrams, screenshots).	Early VLM milestone before Grok-2 Vision.
Grok-2 / Grok-2 mini	2024 beta generation that leapfrogged 1.5; underpinned later vision & image SKUs.	Useful for understanding the performance lineage.

xAI’s public timeline shows a steady march: Grok-1 → 1.5 → 1.5V → 2 / 2 mini → 3 / 3 mini → 4, with 4-Fast and Code Fast rounding out today’s production catalog.

·····

.....

Choosing the right Grok (quick guide)

Your need	Pick this	Why
Highest-quality general reasoning & writing	grok-4-0709	Best-available quality; robust tools; 256k context.
Mass-scale automations & long inputs	grok-4-fast-reasoning	2M context, far lower cost; reasoning + tools.
Low-latency drafts (no reasoning tokens)	grok-4-fast-non-reasoning	Cheapest Grok-4 variant for quick copy/transform.
Stable enterprise extractor	grok-3	Proven JSON/CSV extraction and coding help.
Budget generalist	grok-3-mini	Good enough for many tasks at minimal cost.
Image understanding	grok-2-vision-1212	Multimodal input with regional endpoints.
Image generation	grok-2-image-1212	Simple, inexpensive image synthesis.
Agentic coding	grok-code-fast-1	Purpose-built for code + tools at speed.

·····

.....

Availability notes (consumer vs. API)

Consumer surfaces (Grok web / X Premium+ / SuperGrok): Grok-4 runs the user-facing assistant; Heavy tier advertises a higher-capacity Grok-4 variant for power users. Feature availability (e.g., feed steering, search grounding) is evolving across X apps.
API & cloud hosts: Grok-3/Grok-3 mini shipped broadly (including Azure AI Foundry), then Grok-4 and Grok-4 Fast moved into the public docs and consoles with explicit pricing and limits. Always check regional endpoints and rpm/rps on your account.

·····

.....

Practical tips before you switch models

Reasoning vs. non-reasoning: Reasoning models reject certain legacy params and meter reasoning tokens separately; plan prompts and budgets accordingly.
Real-time grounding: Grok models have no built-in current events—enable Web/X search tools or pass context yourself.
Long context strategy: Even with 2M tokens, retrieval + section prompts remain faster and cheaper than raw dumps. The docs support cached prompt tokens for repeat runs.
Migration hygiene: xAI maintains model aliases (e.g., grok-4, -latest, -<date>). Pin exact versions for reproducibility; use -latest only if you accept silent upgrades.

·····

.....

Bottom line

Today’s Grok lineup spans premium Grok-4, ultra-long-context Grok-4-Fast (reasoning & non-reasoning), workhorse Grok-3 / 3-mini, the grok-code-fast-1 coder, and Grok-2 VLMs for image understanding/generation. Pair them with agentic tools when you need real-time grounding. If you want top quality, pick Grok-4; if you want scale and cost efficiency with giant prompts, pick Grok-4-Fast; for stable enterprise extraction, Grok-3 still earns its keep. All of it is documented—with prices, limits, and tool costs—right in xAI’s Models & Pricing page.

.....

DATA STUDIOS

.....