Name: Grok 4.3
Author: xAI

Grok 4.3 is the latest update to xAI's flagship Grok 4 family, launching in beta on April 17, 2026, without a press release or announcement. The model appeared silently in the model selector on grok.com and in the iOS and Android apps for SuperGrok Heavy subscribers. The two additions over Grok 4.20 are native video understanding and structured document output - capabilities the previous version lacked - while carrying forward the 2-million-token context window and 16-agent Heavy mode.

TL;DR

First Grok model with native video input; also generates downloadable PDFs, PowerPoint slides, and spreadsheets
2M token context window retained from Grok 4.20 - still the largest among Western closed models
Beta-only; locked to SuperGrok Heavy at $300/month with no API access as of April 23, 2026

Elon Musk confirmed on social media that the currently live checkpoint is roughly 0.5 trillion parameters. A larger 1T checkpoint was described as "five days away" from completing training at the time of the April 17 announcement. The SpaceXAI model factory - the joint compute infrastructure announced with Grok 4.20 - is now operational and running training cycles on a planned two-week release cadence.

The lack of a formal announcement and the immediate paywall have slowed independent evaluation. As of April 23, 2026, xAI hasn't published an official model card, benchmark report, or pricing page for Grok 4.3.

Key Specifications

Specification	Details
Provider	xAI
Model Family	Grok
Parameters	~0.5T (confirmed by Musk); 1T checkpoint in training
Context Window	2,000,000 tokens
Input Price	Not yet disclosed
Output Price	Not yet disclosed
Release Date	April 17, 2026 (beta)
Knowledge Cutoff	Not disclosed
License	Proprietary
Modalities	Text + Image + Video input, Text + Document output
Open Source	No

Benchmark Performance

xAI hasn't published benchmark results for Grok 4.3. The table below uses independently verified scores for Grok 4.20 as the nearest available baseline.

Benchmark	Grok 4.20	Claude Opus 4.7	GPT-5.4	Gemini 3.1 Pro
Chatbot Arena ELO	~1493	Not confirmed	Not confirmed	~1485
GPQA Diamond	Not published	Not confirmed	Not confirmed	94.3%
Output speed (tokens/sec)	234.9	~80	~100	~120
AI Index (Artificial Analysis)	48/100	Not listed	Not listed	Not listed

Early community tests with SuperGrok Heavy access report improvements in video question-answering and long-document retrieval, but reproducible results aren't available yet. xAI's pattern on Grok 4 and Grok 4.20 was to publish benchmark data weeks after launch. If that holds, expect official numbers in May 2026.

The reasoning benchmarks leaderboard and multimodal benchmarks leaderboard are the best places to track third-party evaluations as they appear. The agentic AI benchmarks leaderboard will pick up multi-agent performance data once API access opens.

Close-up photograph of a green and black eye with detailed iris texture Grok 4.3 is the first Grok model to accept video as direct input - prior versions were limited to still images with text. Source: unsplash.com

Key Capabilities

Native Video Understanding

Grok 4.20 accepted images alongside text. Grok 4.3 extends that to full video clips. Early users report the model can identify objects, describe event sequences, and answer questions about specific timestamps within a single conversational turn. No maximum video length has been officially specified. Video processing runs through xAI's server-side infrastructure, so no client-side preprocessing is required.

This matters for how the model stacks up against competitors. Gemini 3.1 Pro has had video understanding since launch. GPT-5.4 handles video through the API. Grok 4.3 closes that gap in the consumer product, though API availability for video input isn't confirmed yet.

Structured Document Generation

The other headline addition is native document output. Grok 4.3 generates downloadable PDFs, populated spreadsheets, and formatted PowerPoint slide decks directly from a conversation - no plugins or external integrations needed. A senior xAI engineer demonstrated the model converting a dense neuroscience paper into a polished nine-slide presentation. Early access users describe the outputs as presentation-ready rather than rough drafts.

Claude and GPT-5.4 can produce structured markdown and code, but neither creates downloadable office-format documents natively at this point. Whether this is a lasting advantage or a temporary gap depends on how quickly the other labs respond.

Grok Computer Integration

Grok 4.3 shipped alongside a broader beta rollout of Grok Computer, xAI's autonomous desktop agent. The integration allows Grok 4.3 to execute multi-step task chains across desktop applications - opening files, running operations, summarizing outputs. It's early-stage and behind the same SuperGrok Heavy paywall, but it's the first time the language model and the desktop agent have shipped together as a combined product.

Person holding a professional video camera while filming outdoors Native video input means Grok 4.3 can reason about footage directly - a capability Grok 4.20 didn't have. Source: unsplash.com

Architecture Continuity

The 2M-token context window carries forward unchanged from Grok 4.20, keeping Grok's position on the long-context benchmarks leaderboard. So does the 16-agent Heavy multi-agent mode, where a leader agent spawns up to 16 sub-agents that research and cross-check in parallel before synthesizing a single response.

One new addition separate from the core model: xAI launched standalone Speech-to-Text and Text-to-Speech API endpoints with the 4.3 beta. These aren't part of the language model itself - they're the same stack powering Grok Voice, Tesla in-vehicle AI, and Starlink customer support. The STT endpoint covers 25+ languages and 12 audio formats.

Pricing and Availability

Grok 4.3 is currently exclusive to the SuperGrok Heavy tier at $300/month. That puts xAI $100 above ChatGPT Pro ($200/month) and Claude Max ($200/month). Standard SuperGrok subscribers ($30/month) can see the model in the dropdown but can't use it. Full rollout to all tiers is estimated for mid-to-late May 2026.

Tier	Price
Grok 4.3 beta (SuperGrok Heavy)	$300/month
ChatGPT Pro (GPT-5.4)	$200/month
Claude Max (Claude Opus 4.7)	$200/month
Grok 4.20 API (predecessor)	$2.00/M input, $6.00/M output

API access for Grok 4.3 hasn't been announced. The predecessor Grok 4.20 is available at api.x.ai and on OpenRouter at $2.00/M input and $6.00/M output - a reasonable baseline expectation if xAI maintains pricing continuity across minor versions.

The $300/month price is difficult to defend without persistent memory between sessions. ChatGPT and Claude have offered cross-session memory for over a year. At this subscription level, users expecting session continuity will be disappointed.

Strengths

Native video input - first time in the Grok family, closing a gap vs. Gemini and GPT
Document generation (PDF, PPTX, spreadsheets) without plugins
2M-token context window, still the largest among Western closed models
16-agent Heavy multi-agent mode retained
0.5T live checkpoint with 1T version completing training
Integrated Grok Computer desktop automation expanded to wider beta
New STT/TTS APIs covering 25+ languages

Weaknesses

No persistent memory between sessions - a real gap at $300/month
Beta-only; no API access announced as of April 23, 2026
No official benchmarks published; can't verify performance claims
API pricing unknown
Knowledge cutoff not disclosed
Full rollout to lower tiers not until mid-to-late May 2026

Predecessor models: Grok 4.20 | Grok 4
News: Grok 4.20 Multi-Agent Launch | Grok 4.20 Search Arena Lead
Reviews: Grok 4.20 Review
Leaderboards: Reasoning Benchmarks | Agentic AI Benchmarks | Multimodal Benchmarks | Long Context
Competitors: Claude Opus 4.7 | GPT-5.4 | Gemini 3.1 Pro

FAQ

What's new in Grok 4.3 compared to Grok 4.20?

Native video input and structured document generation (PDFs, spreadsheets, PowerPoint) are the headline additions. The 2M context window and 16-agent multi-agent mode carry over. Elon Musk confirmed the live checkpoint is 0.5T parameters with a 1T version in training.

Is Grok 4.3 available via API?

No. As of April 23, 2026, Grok 4.3 is beta-only and limited to SuperGrok Heavy subscribers at $300/month. API access hasn't been announced.

Does Grok 4.3 have persistent memory?

No. Session memory still resets between conversations, the same limitation as Grok 4.20. ChatGPT and Claude have offered persistent memory for over a year.

How does Grok 4.3 compare to Claude Opus 4.7?

Grok 4.3 has a larger context window (2M vs 200K tokens) and native document generation. Claude Opus 4.7 has persistent cross-session memory; Grok 4.3 doesn't. Neither has published head-to-head benchmark results for this specific version pair.

When will Grok 4.3 be available to standard subscribers?

xAI estimates mid-to-late May 2026 for broader rollout. Currently limited to the $300/month SuperGrok Heavy tier.

Will there be a 1T parameter version?

Yes. Elon Musk stated on April 17, 2026 that a 1T checkpoint was approximately five days from completing training. No release date or separate model designation has been announced for it.

Sources: