Table of Contents
🌏 中文版
On March 17, 2024, xAI dropped 314B-parameter Grok-1 on GitHub under Apache 2.0 — while the industry debated whether to open-source at all, xAI open-sourced its largest MoE. That was Grok's starting point and, to date, its only open weight. Two and a half years later the same family spans four lines: the reasoning Grok 4.6 (500K, $2/$6), the radically cheaper Grok 4 Fast (2M), Imagine for image/video, and Grok Build living in the terminal. This post traces Grok from open-source bet to full product suite.
For how to read the benchmarks cited here, see the AI Model Evaluation Sources Guide. This is the eleventh family deep dive in the AI Model Landscape Overview series.
Family Timeline
| Version | Release | Context | Key fact |
|---|---|---|---|
| xAI founded | 2023-07-12 | — | Led by Elon Musk, anchored by the Memphis Colossus supercomputer |
| Grok debut | 2023-11-03 | 8K | First model, Announcing Grok |
| Grok-1 open release | 2024-03-17 | 8K | 314B MoE (8 experts, 2 per token), 64 layers, Apache 2.0 (xai-org/grok-1 / HF) |
| Grok-1.5 | 2024-03-28 | 128K | 16× context, perfect NIAH retrieval, MATH 50.6% / GSM8K 90% / HumanEval 74.1% |
| Grok-1.5V | 2024-04-12 | 128K | First multimodal (vision), RealWorldQA |
| Grok-2 / mini | 2024-08-13 | — | Grok-2 Beta, LMSYS sus-column-r beating Claude 3.5 Sonnet / GPT-4-Turbo; enterprise API the same month |
| Grok 3 / mini (Think) | 2025-02-19 | 1M | Age of Reasoning Agents, 10× compute on Colossus, Elo 1402 (Chatbot Arena), 1M context (8× prior) |
| Grok 4 / Heavy | 2025-07-09 | 256K | Grok 4, 256K, RL on 200K-GPU Colossus, Heavy via parallel test-time compute |
| Grok Code Fast 1 | 2025-08-28 | 256K | Speedy economical reasoning, built for terminal/editors, $1.00/$2.00 |
| Grok 4 Fast | 2025-09-19 | 2M | 2M context, 40% fewer thinking tokens, 98% price delta, SOTA price-to-intelligence (Artificial Analysis) |
| Grok 4.3 / 4.20 | 2026-03 | 1M | Still-active docs line, $1.25/$2.50, 20% batch discount |
| Grok 4.5 / 4.6 | 2026 (current docs) | 500K | Current docs leaders, $2.00/$6.00 (2× above 200K → $4/$12), four reasoning levels low/medium/high/xhigh |
| Grok Imagine | 2026 | — | Image $0.02/$0.04/$0.05, video $0.05–$0.08/sec (pricing) |
| Grok Build | 2026-08-12 | — | TUI + headless + ACP, 26k stars, Apache 2.0 |
Two and a half years, 13 milestones. The through-line is "catch up with compute and distribution" — Colossus doubling each generation, context from 8K to 2M, and distribution via X / grok.com / Tesla / Bedrock to put the model where users already are.
Three Sub-Lines: Reasoning, Image, Terminal
Grok is often treated as "a chat model" but now spans three sub-lines targeting different rivals:
Reasoning/chat (Grok 1 → 4.6 + 4 Fast): xAI's main battle. Grok 3 scaled reasoning with RL at unprecedented scale (10× compute on Colossus, seconds-to-minutes thinking, backtracking); Grok 4 pushed Heavy (parallel test-time compute) for leaderboard wins; Grok 4 Fast trades 2M context and 40% fewer thinking tokens for a 98% price cut — the "price-to-intelligence king."
Image/video (Imagine): grok-imagine-image ($0.02), image-quality ($0.05), image-2.0 ($0.04), video ($0.05/sec) and video 1.5 ($0.08/sec). The AI Model Landscape Overview image ELO places Grok Imagine 2.0 in the top five, but xAI's own announcement and leaderboard details were Cloudflare-blocked during verification — cite LMArena/Artificial Analysis with dates.
Terminal coding (Build / Code Fast): Grok Build is xAI's coding agent harness — fullscreen TUI, mouse-interactive, extensible — with three run modes: interactive, headless (grok -p "prompt" --output-format streaming-json for CI/scripting), and embedded in editors via ACP. The harness is Apache 2.0; the model behind it is Grok 4.6 (grok-build-0.1 / grok-code-fast-1, 256K, $1/$2).
Architecture: MoE Start, Colossus and RL Scaling
Grok-1's MoE Starting Point
Grok-1's Model Specifications (model.py) specify: 8 experts, 2 active per token, 64 layers, 48 Q-heads / 8 KV-heads, embedding 6,144, SentencePiece 131,072, RoPE, 8-bit/activation sharding. At 314B total it was among the largest open MoEs at the time, but 8K context already looks cramped next to the later 128K/1M.
Colossus and Reasoning Scaling
xAI's narrative spine is compute. Grok 3 claims RL at "unprecedented scale" with 10× prior compute on Colossus; Grok 4 adds 200K-GPU Colossus, 6× compute efficiency, 10× Grok 3 reasoning compute. This explains Grok's iteration speed — not an architectural ambush but pushing RL and test-time compute to the limit (Heavy trades parallel compute for scores).
Pricing and Performance: How to Choose
API Pricing (per million tokens, from docs.x.ai)
| Model | Context | Input (<200K) | Cached | Output (<200K) | Input (≥200K) | Output (≥200K) | Reasoning levels |
|---|---|---|---|---|---|---|---|
| Grok 4.6 | 500K | $2.00 | $0.50 | $6.00 | $4.00 | $12.00 | low/medium/high/xhigh (default high) |
| Grok 4.5 | 500K | $2.00 | $0.30 | $6.00 | $4.00 | $12.00 | Same |
| Grok 4.3 / 4.20 (reasoning/non-reasoning/multi-agent) | 1M | $1.25 | $0.20 | $2.50 | $2.50 | $5.00 | — (20% batch discount) |
| Grok Build 0.1 (Code Fast 1) | 256K | $1.00 | $0.20 | $2.00 | $2.00 | $4.00 | reasoning |
| Grok 4 Fast | 2M | — (anchored to 4.5/4.6, 98% delta) | — | — | — | — | Unified reasoning/non-reasoning |
Two details: above 200K tokens the price doubles (both input and output); 20% batch discount applies only to 4.3/4.20. Long-context workloads must be priced at the doubled rate.
Performance Position (xAI official releases; third-party boards cited with dates)
| Metric | Grok 4 (official) | Grok 3 (official) | Reference (in-series) |
|---|---|---|---|
| HLE (Humanity's Last Exam) | 44.4% Heavy / 38.6% (full set, with Python+search) | — | DeepSeek V4 Pro undisclosed; see Claude/GPT family posts |
| ARC-AGI v2 | 15.9% (vs Opus 8.6%) | — | — |
| GPQA | 87.5–88.4% | 84.6% | Gemini 3.1 Pro 94.1% (science reasoning still leads) |
| LiveCodeBench | 79.3% | 79.4% | DeepSeek V4 Pro 96.4% (different board: SWE-bench Verified) |
| AIME'25 | — | 93.3% cons@64 | — |
Grok 4 Fast's official comparison (Grok 4 Fast): GPQA Diamond 85.7% vs Grok 4 87.5%, AIME'25 92.0% vs 91.7%, HLE 20.0% vs 25.4%, LiveCodeBench 80.0% vs 79.0% — near-flagship scores for 2M context and a 98% price cut.
The site's daily Grok 4.6 note (Grok 4.6 daily): GDPVal-AA v2 1753 Elo (field high) but hard-core coding (DeepSWE, Terminal-Bench) still behind GPT-5.6 Sol Max — consistent with the family post's "strong distribution, still chasing on hard engineering" framing.
How to Choose
- Strongest reasoning with 500K → Grok 4.6 (high/xhigh), watch the >200K doubling.
- Value with 2M long docs → Grok 4 Fast — 40% fewer thinking tokens, 98% cheaper, keeping 85%+ GPQA.
- Terminal coding agent → Grok Build (Apache 2.0 harness) + grok-build-0.1 / 4.6 API, with headless/ACP.
- Image/video generation → Imagine image $0.02–$0.05 / video $0.05–$0.08/sec by quality tier.
- Local self-host → Only Grok-1 (314B, Apache 2.0); everything after is closed API.
Licensing Traps: Open Only Once
Grok has the sharpest licensing cliff in the series:
- Grok-1: Apache 2.0, commercially usable, self-hostable, modifiable — but 8K context and 2024 architecture, no longer frontier.
- Grok 1.5 / 2 / 3 / 4 / 4 Fast / 4.5 / 4.6: Closed API, no weight downloads. License is xAI Terms of Service.
- Grok Build harness: Apache 2.0 — harness open, model closed. You can self-host the harness, not the model.
- Grok Imagine: API-billed, no weights.
Comparison: Qwen (mostly Apache 2.0), DeepSeek (MIT), Gemma (Apache 2.0, 2B–27B), Llama 4 (Community License, downloadable), Mistral (Apache/Modified MIT) all offer commercially self-hostable weights; only Grok-1 is self-hostable in the Grok family, two generations behind the current frontier.
Sub-Lines and Ecosystem
| Sub-line | Representative | Positioning |
|---|---|---|
| Flagship reasoning | Grok 4.6 500K | Strongest reasoning, four tunable levels, $2/$6 |
| Long value | Grok 4 Fast 2M | 2M, 98% price delta, SOTA price-to-intelligence |
| 1M flagship | Grok 4.3 / 4.20 1M | 1M, $1.25/$2.50, 20% batch discount |
| Light reasoning | Grok Code Fast 1 / Build 0.1 256K | Terminal/editor coding, $1/$2 |
| Image | Imagine image / image-quality / 2.0 | $0.02/$0.05/$0.04 per image |
| Video | Imagine video / 1.5 | $0.05/$0.08 per second |
| Terminal agent | Grok Build | TUI + headless + ACP, Apache 2.0 harness |
| Open weight | Grok-1 314B | Only open weight, Apache 2.0, 8K |
Two observations:
Distribution is Grok's most underrated lever. Built into X, on grok.com, in Tesla vehicles, on Amazon Bedrock (community-reported as live Aug 2026; not listed on xAI's site at writing — verify against AWS announcements) — Grok's reach does not rely on leaderboards but on the Elon distribution system. That is why xAI can trade "faster iteration, lower price" for catch-up.
Build's openness is harness openness. xai-org/grok-build (26k stars) and grok-build-plugin-cc open-source the harness and Claude Code plugin; the model still goes through the xAI API. For agent builders, Build's value is an auditable, forkable harness — not a self-hostable model.
Competitive Position
Grok in the 2026 landscape:
- vs. GPT-5.6: GPT still leads on hard agentic boards (SWE-bench Verified / Terminal-Bench / BrowseComp); Grok chases with 2M/500K and lower pricing
- vs. Claude Opus 5: Claude leads at 96% SWE-bench Verified and agentic stability; Grok's edge is distribution and terminal-native (Build vs. Claude Code)
- vs. Gemini 3.1 Pro: Gemini leads on GPQA 94.1% / ARC-AGI-2 77.1% for science; Grok has highlights on HLE/ARC-AGI v2 but not a sweeping win
- vs. DeepSeek V4: DeepSeek matches Claude on SWE-bench at 1/28 the price under MIT and is self-hostable; Grok has no commercially self-hostable flagship (only Grok-1) and requires official xAI deployment for private use
- vs. Qwen3.8 / Kimi K3 / GLM-5.3: All three ship 1T+ open weights; Grok has no equivalent open flagship — its differentiation is Imagine and Build breadth
- vs. open small models: No Phi/Gemma/Ornith-equivalent open small line; edge/local is not Grok's lane
What This Means for Agent Builders
- Long reasoning → Grok 4 Fast (2M) or Grok 4.6 (500K) — watch the >200K doubling; 4.3/4.20's 1M + batch discount suits off-peak batches
- Terminal coding → Grok Build (Apache 2.0 harness) + grok-build-0.1/4.6 — headless for CI, ACP for editors, competing head-on with Claude Code
- Image/video generation → Imagine ($0.02–$0.08) — same account, same API, same distribution as text reasoning
- X / Tesla ecosystem → If your product already lives in the X or Tesla stack, Grok has the lowest integration friction
- High-throughput / cost-sensitive → Grok 4 Fast's 98% delta + 2M is the most aggressive value play today; for well-defined bug fixes DeepSeek V4 Pro (SWE-bench 96.4%, ultra-low cost) remains an option
- Local / private deployment → Only Grok-1 is self-hostable; private use goes through the xAI API / official deployment — factor in vendor lock-in and the >200K doubling
Pragmatic mix-and-match:
| Task | Recommendation | Why |
|---|---|---|
| Long reasoning / research | Grok 4 Fast 2M or Grok 4.6 500K | 2M/500K + four tunable reasoning levels |
| Terminal coding agent | Grok Build + grok-build-0.1 | Auditable, embeddable harness |
| Image/video generation | Imagine image 2.0 / video 1.5 | One xAI account end-to-end |
| Well-defined bug fix | DeepSeek V4 Pro | SWE-bench 96.4%, ultra-low cost |
| Complex agentic orchestration | Claude Opus 5 / GPT-5.6 Sol | Agentic stability still ahead of Grok |
| Local deployment | Qwen3.8-27B / Gemma 3 27B | No current open Grok flagship |
Overall
Grok's story is "catch up with distribution and compute." xAI proved good faith by open-sourcing 314B, then bet everything on Colossus RL scaling and distribution through X / Tesla / Bedrock — the model as a traffic entry point for the Elon system. By 2026 Grok is no longer a single model but a quartet: 4.6 at 500K, 4 Fast at 2M, Imagine for images, Build for the terminal.
The price is openness and the hard-coding gap. No open flagship, Build opens the harness not the model; HLE/ARC-AGI highlights aside, it still trails GPT-5.6 and Claude on SWE-bench / Terminal-Bench engineering boards. That defines Grok's pragmatic fit: the default inside its distribution, the long-context value king, the second choice for terminal-native agents — not a replacement for the strongest coding model.
References
- xAI Official Site
- Announcing Grok — xAI Blog — Nov 3, 2023 debut
- Open Release of Grok-1 — xAI / xai-org/grok-1 — GitHub — 314B MoE, Apache 2.0
- Grok-1 — HuggingFace (xai-org/grok-1) — 314B, Apache 2.0
- Announcing Grok-1.5 — xAI Blog — 128K, NIAH
- Grok-1.5 Vision Preview — xAI Blog — First multimodal
- Grok-2 Beta Release — xAI Blog — Grok-2 / mini, LMSYS
- Grok 3 Beta — The Age of Reasoning Agents — xAI Blog — 1M, Elo 1402, 10× Colossus
- Grok 4 — xAI News — 256K, Heavy, HLE/ARC-AGI
- Grok Code Fast 1 — xAI News — Terminal/editor coding
- Grok 4 Fast — xAI News — 2M, 98% price delta, 40% fewer thinking tokens
- xAI Docs — Models — Current model catalog, contexts, reasoning levels
- xAI Docs — Pricing — Full pricing, >200K doubling, batch
- Grok 4.6 — xAI Docs — 500K, $2/$6
- Grok Build 0.1 — xAI Docs — 256K, $1/$2
- Grok Imagine Image 2.0 — xAI Docs — $0.04
- Grok Build — GitHub (xai-org/grok-build) — 26k stars, Apache 2.0, TUI/ACP
- Grok Build Plugin CC — GitHub — Claude Code plugin
- Grok Build Overview — xAI Docs — Three run modes, ACP
- AI Model Evaluation Sources Guide — This site
- AI Model Landscape Overview — This site
Loading...