TechSifted has no affiliate relationship with xAI.
xAI didn't send a press release. No blog post, no fanfare. Grok 4.3 Beta just appeared in the model selector on grok.com on April 17 — flagged as "Early Access" — and early testers started posting about it while the broader AI internet was distracted by other things.
That's a pattern xAI has leaned into. Quiet drop, let users find it, let Elon tweet about it later.
So here's what's actually in this release, what's new, and who actually needs to care.
What Is Grok 4.3 Beta — and Who Can Use It
Grok 4.3 Beta is locked behind the SuperGrok Heavy tier at $300/month. If you're on the regular SuperGrok plan ($30/month), you can see it in the dropdown. You just can't use it. The full rollout is estimated for mid-to-late May 2026.
At $300/month, xAI is going head-to-head with ChatGPT Pro ($200/month) and Claude Max ($200/month). That's a steep premium — and it's worth being honest about whether the upgrade justifies it.
What's New vs. Grok 4.20
The biggest additions in 4.3 aren't reasoning tweaks. They're output types.
Grok 4.3 can now generate downloadable PDFs, fully populated spreadsheets, and PowerPoint decks directly from conversation. These aren't rough drafts — early testers are reporting formatted outputs you could actually hand to someone. That's a meaningful step beyond what 4.20 could do.
The other big one: native video input. Grok 4.20 handled images. Grok 4.3 processes and understands video content conversationally. You can share a video clip and actually reason about what's in it.
There's also tighter integration with Grok Computer, xAI's autonomous desktop automation agent. Think: Grok planning and Grok executing, running in parallel on your machine.
Architecture-wise, the 16-agent Heavy system and the 2 million token context window from 4.20 are both retained — that context window is still the largest among Western closed models. Early reports suggest 4.3 launched at roughly 0.5T parameters, with a 1T checkpoint completing training. Enhanced reasoning depth is attributed to longer training runs, though xAI hasn't published a model card.
One notable gap: still no persistent memory between sessions. ChatGPT and Claude have had this for over a year. At $300/month, its absence is genuinely hard to defend. If you're expecting Grok to remember your preferences or ongoing projects, it won't.
The New Audio APIs: Speech-to-Text and Text-to-Speech
This is arguably the bigger developer story from April 17, and it didn't get nearly enough coverage.
xAI quietly launched standalone Speech-to-Text (STT) and Text-to-Speech (TTS) APIs — the same stack that powers Grok Voice, Tesla vehicle infotainment, and Starlink customer support.
Speech-to-Text
The STT API covers 25+ languages and 12 audio formats. The headline features are word-level timestamps, multichannel support, Inverse Text Normalization, and — most useful for anyone building call center or interview tools — speaker diarization. It runs in two modes: real-time WebSocket streaming and batch processing via REST.
Pricing:
- Batch: $0.10/hour
- Streaming: $0.20/hour
On phone call entity recognition error rate, xAI benchmarks Grok STT at 5.0% vs. ElevenLabs at 12%, Deepgram at 13.5%, and AssemblyAI at 21.3%. That's a notable gap — though "phone call entity recognition" is a specific enough test that I'd want to see broader benchmark coverage before treating it as gospel.
That said, for medical, legal, and financial transcription use cases, the early signals look strong.
Text-to-Speech
Five voices: Ara, Eve, Leo, Rex, and Sal. Twenty-plus languages, auto-detected. The TTS API supports expressive speech tags — inline markers like [laugh], [sigh], and [whisper] that let you control vocal delivery without needing a separate fine-tuned model.
Pricing: $4.20 per 1 million characters.
Compare that to OpenAI at ~$30/1M chars and ElevenLabs, which has been the go-to for expressive AI voice, at ~$50/1M chars. xAI is undercutting both by roughly 86–92%.
That's not a small pricing difference. That's a different category of cost.
Batch API Expansion: Images and Video
Less flashy, but practically important for developers.
xAI's Batch API now supports image generation, image editing, and video generation in addition to existing chat completions. You submit a JSONL file with mixed request types via the Files API. Everything runs asynchronously, completing within 24 hours.
One thing to know: generated image and video URLs expire after 1 hour. Download immediately.
The pricing with the 50% batch discount:
| Type | Batch Rate |
|---|---|
| Image Generation (Standard) | $0.01/image |
| Image Generation (Pro) | $0.035/image |
| Video Generation | $0.025/second |
At $10 for 1,000 standard-quality product images, xAI is competitive — actually undercutting most alternatives on per-unit cost. For any team running batch creative workflows at scale, this is worth a serious look.
XChat: The Standalone App
XChat launched on the Apple App Store on April 17, right alongside the model update.
It's xAI's encrypted messaging app — think Signal crossed with a WeChat-style ambition. Built in Rust. End-to-end encryption, no ads, voice and video calls, disappearing messages, screenshot blocking. Group chats up to 481 people, 45 languages supported.
Grok is baked into the experience. Long-press any message, tap "Ask Grok," and you get real-time AI analysis. The catch: that query goes to Grok unencrypted. The surrounding chat stays encrypted. Make of that what you will.
It's iOS-only at launch, and it requires iOS 26 or higher — which locks out a meaningful chunk of iPhone users. No Android timeline announced.
XChat is clearly Elon's "super app" play — a Western WeChat. Whether it gets there from here is a much longer conversation. For now, it's an early-access messaging app with a capable AI assistant built in. That's worth something, but it's a long way from platform.
Who Should Actually Care
If you're a developer building voice products: the STT and TTS APIs are the most immediately actionable thing here. The pricing alone makes them worth evaluating — especially if you've been paying ElevenLabs or OpenAI rates. If you're building on Cursor too, the xAI/Cursor partnership we covered last week means xAI's ecosystem is becoming more central to the developer stack either way.
If you're a power user already on SuperGrok Heavy: native video input and the document generation features are genuine additions. Not transformative, but real.
If you're considering upgrading from regular SuperGrok to Heavy: $300/month is a meaningful spend, and the lack of persistent memory at that price is a real friction point. Wait for the full rollout in May to see if anything else ships before committing.
If you're evaluating LLMs for batch creative workflows: the Batch API expansion with image and video support, at those per-unit rates, is worth running a small pilot.
And if you want to compare where the broader LLM field is right now — including Anthropic's recent Claude Opus 4.7 release — there's a lot moving fast this month. Grok 4.3 is a real update. But it's one of several.
xAI released Grok 4.3 Beta on April 17, 2026. All pricing, features, and access information reflects what was publicly available at time of writing. TechSifted has no affiliate relationship with xAI.
Top comments (0)