New Grok 4.3 Update: Elon Musk's BEST Model?
Grok 4.3 was quietly released by xAI on April 17, 2026, with no formal announcement, bringing major upgrades including native video input, document generation, and significantly improved agentic task performance. The model scores 53 on the Artificial Analysis intelligence index and made a massive 321-point ELO leap on agentic benchmarks, while also becoming cheaper to run than its predecessor. However, it still lacks persistent memory, a feature competitors like ChatGPT and Claude have offered for over a year.
Summary
Grok 4.3 was released by xAI on April 17, 2026, without a press release or launch event — it simply appeared in the model picker on grok.com marked as early access. Access is currently restricted to Super Grok Heavy tier subscribers, with a wider rollout expected in mid-to-late May 2026. xAI has confirmed plans to ship daily improvements during the beta period, making this an actively evolving release rather than a polished product.
The most significant new capabilities include native video input, allowing users to drop in video clips and have contextual conversations about their content without needing transcripts. Grok 4.3 can also generate full PDFs, populated spreadsheets, and PowerPoint presentations directly from chat. It also features tighter integration with Grok Computer, xAI's autonomous desktop agent, enabling more reliable multi-step task execution across applications.
On benchmarks, Grok 4.3 scores 53 on Artificial Analysis's intelligence index, placing it above Claude Sonic 4.6 and four points ahead of the previous Grok 4.20. Its most dramatic improvement came on the GDPV benchmark for agentic tasks, where it scored an ELO of 1500 — a 321-point jump — beating Gemini 3.1 Pro Preview, GPT 5.4 mini, and Kimi K2.5, with only GPT 5.5 still ahead. It also scored 81% on instruction following and 98% on the TA2 Bench telecom customer support benchmark.
A notable trade-off: while Grok 4.3 gained 8 points in accuracy in one test, it also lost 8 points on non-hallucination rate in the same test, meaning it knows more but fabricates information slightly more often than its predecessor. Despite improved capabilities, the model is actually cheaper to run — input pricing dropped ~37%, output pricing dropped ~58%, and overall efficiency improved by ~20%.
xAI also quietly launched standalone speech-to-text and text-to-speech APIs supporting 25+ languages, 12 audio formats, word-level timestamps, speaker detection, and real-time streaming. The text-to-speech API supports expressive inline tags like sighs and laughs. Additionally, xAI launched XChat, an encrypted iOS messaging app built in Rust with end-to-end encryption, disappearing messages, voice/video calls, and Grok integration, requiring iOS 26 or higher.
The one major gap highlighted is the continued absence of persistent memory. Unlike ChatGPT and Claude, which have offered memory for over a year, Grok 4.3 still starts every conversation fresh with no recollection of past interactions, preferences, or projects. The video concludes by contextualizing Grok 4.3 within a highly competitive April 2026 AI landscape, noting that the key skill is knowing which tool to use for which task rather than relying on a single model.
Key Insights
- Grok 4.3 made a 321-point ELO leap on the GDPV agentic benchmark, scoring 1500 and beating Gemini 3.1 Pro Preview, GPT 5.4 mini, and Kimi K2.5 — with only OpenAI's GPT 5.5 still ahead of it on this specific test.
- Despite being a more capable model, Grok 4.3 is actually cheaper to run than its predecessor — input pricing dropped ~37%, output pricing dropped ~58%, and overall benchmark efficiency improved by ~20%, which the speaker notes is almost unheard of in the AI industry.
- Grok 4.3 gained 8 points in accuracy on one test but simultaneously lost 8 points on the non-hallucination rate in the same test, meaning it is both more knowledgeable and more prone to making things up than the previous version.
- xAI quietly launched standalone speech-to-text and text-to-speech APIs supporting 25+ languages, 12 audio formats, real-time streaming, and expressive inline voice tags — the same voice stack that powers Grok Voice and Tesla's in-car voice system.
- Grok 4.3 still lacks persistent memory, meaning every conversation starts completely fresh with no recall of past sessions, preferences, or projects — a feature that both ChatGPT and Claude have offered for over a year, which the speaker identifies as the model's single biggest missing capability.
Topics
Full transcript available for MurmurCast members
Sign Up to Access