Grok 4.3 Beta went live on April 17, 2026. No press release, no blog post — just a new "Early Access" entry in the model selector on grok.com, iOS, and Android. If you are on SuperGrok Heavy at $300/month, you can use it now. If you are on the standard SuperGrok plan at $30/month, you can see it in the dropdown but cannot activate it. Full rollout to lower tiers is estimated for mid-to-late May 2026.
Here is everything confirmed from early testers, xAI's release notes, and independent benchmarks — no speculation, no padding.
What Is Grok 4.3 Beta?
Grok 4.3 Beta is xAI's latest iteration of its flagship Grok 4 series, building on the four-agent architecture (Grok, Harper, Benjamin, and Lucas) introduced in Grok 4.20. The 16-agent Heavy system and the 2 million token context window from 4.20 are both retained — the largest context window among Western closed frontier models. Early testers on X report Grok 4.3 may be running at roughly double the parameter count of 4.20, though xAI has not published a model card or confirmed this figure. Enhanced reasoning depth is attributed to longer training runs.
The biggest additions in 4.3 are not architecture tweaks. They are output types and API expansions that make Grok substantially more useful as a production tool.
Grok 4.3 Beta: New Features Confirmed
Native Video Input
Grok 4.20 handled images. Grok 4.3 processes and understands video content conversationally. You can share a video clip and ask Grok to reason about what is happening in it — summarize content, extract quotes, or analyze visual sequences. This is a meaningful multimodal step beyond image-only understanding.
Document Generation: PDFs, Spreadsheets, and Slides
Grok 4.3 can now generate downloadable PDFs, fully populated spreadsheets, and PowerPoint-style presentation decks directly from conversation. Early testers are reporting formatted outputs you could actually hand to someone — not rough-draft HTML exports. This is the feature that has generated the most traction on X since launch. For a dedicated comparison of AI tools that generate slide decks, see our guide to best AI tools for PowerPoint presentations in 2026.
Expanded Batch API: Image and Video Generation
xAI expanded the Batch API to support image generation, image editing, and video generation in addition to chat completions. Both server-side tools and client-side function tools are now supported in batch requests. This makes Grok significantly more useful for creative production pipelines that need to process large volumes of assets asynchronously.
Grok Computer Integration
Launching in parallel with Grok 4.3, Grok Computer — xAI's autonomous desktop agent — entered wider beta access for select SuperGrok accounts. Grok Computer can open applications, navigate UIs, fill forms, and chain multi-step desktop tasks without requiring API access. It works by reading pixels, meaning it can operate any software including legacy programs. The architecture is deliberate: Grok 4.3 is the reasoning engine, Grok Computer is the execution layer. Together they are xAI's version of an AI that does not just answer questions — it does work. For context on how Claude handles similar desktop automation, see our review of Claude computer use on Windows.
Standalone Speech-to-Text and Text-to-Speech APIs
Launched the same day as Grok 4.3, xAI's Speech-to-Text (STT) and Text-to-Speech (TTS) APIs are now generally available — the same stack powering Grok Voice, Tesla vehicle infotainment, and Starlink customer support.
The STT API supports 25+ languages and 12 audio formats. Key capabilities include word-level timestamps, multichannel audio support, Inverse Text Normalization, speaker diarization (word-level speaker identification for call center and interview transcription), and two operating modes: real-time WebSocket streaming and batch REST processing. xAI benchmarks Grok STT at a 5.0% entity recognition error rate on phone call audio, versus ElevenLabs at 12%, Deepgram at 13.5%, and AssemblyAI at 21.3%.
The TTS API offers five voices — Ara, Eve, Leo, Rex, and Sal — across 20+ auto-detected languages. It supports expressive speech tags: inline markers like [laugh], [sigh], and [whisper] for controlling vocal delivery without a separate fine-tuned model. Pricing is $4.20 per 1 million characters — compared to approximately $30/million for OpenAI and $50/million for ElevenLabs. That is an 86–92% cost reduction. To learn more about Grok's voice capabilities, see our guide on how to enable Grok voice mode and the detailed breakdown of the Grok Imagine API access and pricing.
Grok 4.3 Beta vs Grok 4.20: What Changed
| Capability | Grok 4.20 | Grok 4.3 Beta |
|---|
| Video input | No | Yes — native video understanding |
| Document generation | No | PDF, spreadsheet, slide deck output |
| Context window | 2M tokens | 2M tokens (retained) |
| Agent architecture | 16-agent Heavy system | 16-agent Heavy system (expanded) |
| Batch API | Chat completions only | + image, image editing, video gen |
| Desktop agent | Not available | Grok Computer (parallel beta) |
| STT / TTS APIs | Not available | GA — $4.20/million chars TTS |
| Persistent memory | None | None (still absent) |
| Access tier | SuperGrok Heavy ($300/mo) | SuperGrok Heavy ($300/mo) |
Access and Pricing
As of April 17, 2026, access to Grok 4.3 Beta is structured as follows:
- Free tier: No access. Grok 4.3 is not visible or selectable.
- SuperGrok ($30/month): Grok 4.3 appears in the model selector with an "Early Access" label but cannot be activated.
- SuperGrok Heavy ($300/month): Full access to Grok 4.3 Beta now. This is the only way to use it during the beta period.
- X Premium+ ($40/month): No Grok 4.3 access — X platform plans do not include the Heavy tier.
xAI has not announced a specific date for broader rollout. Based on the Grok 4.20 beta pattern — narrow launch, daily iteration, expand over weeks — the mid-to-late May estimate for standard SuperGrok access is reasonable but not confirmed.
At $300/month, xAI is positioned above ChatGPT Pro ($200/month) and Claude Max ($200/month). For a direct feature comparison across the major AI platforms, see our ChatGPT vs Gemini 2026 comparison and the full April 2026 AI tools roundup covering every major model release this month.
The One Major Gap: Still No Persistent Memory
Grok 4.3 still has no persistent memory between sessions. At $300/month, this is the most legitimate criticism of the Heavy tier. ChatGPT and Claude have both offered persistent memory for over a year. If you expect Grok to remember your preferences, ongoing projects, or context from previous sessions, it will not. Every conversation starts cold. This is arguably the more important story from the April 17 launch, and it did not get enough coverage relative to the new features.
Who Should Upgrade to Grok 4.3 Beta?
Upgrade now if you are already on SuperGrok Heavy and run workflows that would directly benefit from native video analysis or automated document generation — formatted PDFs, spreadsheets, or slide decks you can hand to clients.
Wait if you are considering upgrading from standard SuperGrok ($30) to Heavy ($300) specifically for 4.3. The lack of persistent memory at that price is a real friction point. Hold until the full rollout in May to see what else ships before committing to a 10x price jump.
Evaluate the API instead if you are building batch creative or transcription workflows. The STT API at 5.0% entity error rate and the TTS API at $4.20/million characters are worth a pilot run regardless of which subscription tier you are on. xAI also reduced agent tool API pricing by up to 50% in April 2026, capping costs at $5 per 1,000 successful calls.
Stick with alternatives if persistent memory is essential to your workflow. Claude (via Projects) and ChatGPT both handle cross-session memory today. Grok 4.3 does not. For how Claude manages long-running tasks, see our breakdown of Claude Managed Agents in public beta.
Grok 4.3 in Context: April 2026 Model Race
Grok 4.3 Beta launched the day after Anthropic released Claude Opus 4.7 (April 16, 2026). The timing was not accidental. In April 2026 alone, 19 major AI models or significant updates shipped — including Llama 4 Scout and Maverick from Meta, Gemma 4 from Google, and GLM-5.1 from Zhipu AI. On coding benchmarks, Claude Opus 4.7 leads at 87.6% on SWE-bench Verified. For computer use tasks, GPT-5.4 has a head start from its March 2026 launch. Grok 4.3's edge is the 2 million token context window and now native document creation — if those are your bottlenecks, the $300 price becomes more defensible. For more detail on the broader competitive landscape, see the complete Grok usage guide and the latest Grok app updates.
Frequently Asked Questions
When will Grok 4.3 be available to regular SuperGrok subscribers?
xAI has not given a specific date. Based on the rollout pattern used for Grok 4.20, the mid-to-late May 2026 estimate is reasonable. Standard SuperGrok subscribers ($30/month) can currently see Grok 4.3 in the model dropdown but cannot activate it. The beta label will likely be removed when xAI widens access.
What is the difference between Grok 4.3 and Grok 4.20?
The core architecture is the same — 16-agent Heavy system, 2 million token context. What Grok 4.3 adds on top of 4.20: native video input (4.20 handled images only), document generation (downloadable PDFs, spreadsheets, and slide decks from conversation), Batch API expansion to include image and video generation, tighter Grok Computer integration, and the new standalone STT and TTS APIs. There are no published benchmark differences yet — xAI has not released a model card for 4.3.
Is Grok 4.3 worth the $300/month SuperGrok Heavy subscription?
If you are already on Heavy and use it for deep research, multi-agent analysis, or batch creative workflows, the document generation and video input features are genuine additions. If you are considering upgrading from standard SuperGrok ($30/month) specifically for 4.3, wait until the full May rollout to see what else ships first. The absence of persistent memory at $300/month is a legitimate gap that ChatGPT Pro and Claude Max both cover at a lower price.
Does Grok 4.3 have persistent memory?
No. Grok 4.3 Beta still does not support persistent memory between sessions. Every conversation starts fresh. This applies at all tiers, including SuperGrok Heavy at $300/month. ChatGPT and Claude have both had persistent memory for over a year. xAI has not announced a timeline for adding this feature to Grok.
How do I access Grok 4.3 Beta right now?
You need an active SuperGrok Heavy subscription at $300/month (or $3,000/year). Go to grok.com, open a new conversation, click the model selector, and look for "Grok 4.3 (beta)" labeled as Early Access. It is also available in the iOS and Android Grok apps under the same model picker. If you are on standard SuperGrok, you will see it listed but it will be locked — there is no workaround available during the beta period.
How does the new Grok TTS API compare to ElevenLabs and OpenAI?
xAI priced the TTS API at $4.20 per 1 million characters. OpenAI TTS runs at approximately $30 per million characters, and ElevenLabs at approximately $50 per million characters. xAI is undercutting both by 86–92%. The Grok TTS API supports five voices (Ara, Eve, Leo, Rex, Sal), 20+ languages with auto-detection, and expressive speech tags like [laugh], [sigh], and [whisper] for controlling vocal delivery. For a full walkthrough of Grok's voice stack, see our guide on enabling Grok voice mode on iPhone, Android, and desktop.
What is Grok Computer and how does it relate to Grok 4.3?
Grok Computer is xAI's autonomous desktop agent — separate from Grok 4.3 but launched the same week. It can open applications, navigate UIs, fill forms, and chain multi-step desktop workflows without requiring API access. It operates by reading pixels, so it works with any software including legacy programs. The private beta started April 13, 2026; wider beta access for select SuperGrok accounts began April 17. Grok 4.3 is the reasoning engine. Grok Computer is the execution layer. Together they represent xAI's answer to autonomous AI that does work rather than just answering questions.