xAI Opens Grok 4.3 to All Developers: Million-Token Context, Video Input, and Voice Cloning at 40% Lower Cost
News Summary
xAI officially opened its Grok 4.3 model to full API access on April 30, 2026 (Pacific Time), completing a staged rollout that began with a SuperGrok Heavy subscriber beta on April 17, 2026 (Pacific Time). The release bundles five substantive upgrades — a 1 million-token context window, always-on chain-of-thought reasoning, native video input, a roughly 40% price reduction, and a new voice-cloning suite — representing one of the most feature-dense updates in the Grok model lineage to date.
What Is Grok 4.3 and Where Does It Fit
Grok 4.3 is xAI's latest large language model, succeeding Grok 4.20. The model targets developers building agentic and enterprise applications where long context, low latency, and predictable reasoning are critical. It is accessible through the xAI API directly, via Vercel AI Gateway, and through OpenRouter. For developers already working within the OpenAI SDK ecosystem, Grok 4.3 uses an OpenAI-compatible base URL — no code changes are needed beyond swapping the model field to grok-4.3. This compatibility lowers migration costs substantially and is widely seen as a deliberate effort to attract developers from competing platforms.
The 1 Million-Token Context Window
One of the headline capabilities is a context window of 1,000,000 tokens, a significant step up from earlier Grok generations. At roughly 750 words per token, a million tokens can accommodate tens of thousands of pages of text — the equivalent of a large software codebase, multiple lengthy research reports, or an extended customer interaction history. Costs do scale at the 200,000-token input boundary, where per-token prices double, so developers processing very large payloads will need to account for tiered billing in their architecture.
Always-On Reasoning
Grok 4.3 implements a persistent, non-toggleable chain-of-thought (CoT) mechanism. Unlike models where extended reasoning is an optional parameter, every Grok 4.3 API call automatically performs internal reasoning before producing a response. xAI describes this as "always-on" reasoning. On the τ²-Bench Telecom benchmark — a standard for agentic customer support accuracy — Grok 4.3 achieved 98%, a 5-point improvement over Grok 4.20.
On the GDPval-AA agentic benchmark, Grok 4.3 recorded an ELO score of 1,500, up 321 points from Grok 4.20's score of 1,179, surpassing several competitive models in the same evaluation window. Independent analysis from Artificial Analysis ranked the model just above Claude Sonnet 4.6 and Muse Spark on their composite Intelligence Index while simultaneously reducing cost per benchmark run. The persistent reasoning does introduce some additional latency per call, which developers should evaluate against their application's response-time requirements.
Native Video Input
Grok 4.3 is the first model in the xAI API to support native video input. Developers can pass a video URL directly to the model without pre-processing steps such as frame extraction or transcoding. Use cases include automated content analysis, video-based question answering, and multimodal agentic workflows that require temporal understanding of visual information.
Pricing and Output Speed
The new pricing sets input tokens at $1.25 per million (down from approximately $2.00 with Grok 4.20) and output tokens at $2.50 per million (down from approximately $6.00) — a reduction of 37.5% on inputs and 58.3% on outputs, averaging roughly 40% overall. At a reported throughput of 159 tokens per second, Grok 4.3 also delivers faster output than its predecessor. Together, the lower price per token and higher throughput substantially improve the economics of high-volume or real-time deployments.
Custom Voices: The Voice Cloning Suite
Alongside the language model update, xAI released Custom Voices — a voice-cloning API and web-based creation suite. Developers can clone a voice from a reference audio clip as short as 120 seconds. The resulting voice ID can then be used across xAI's Text-to-Speech (TTS) and Voice Agent APIs, enabling consistent synthetic voices in conversational and customer-facing applications. This positions xAI as a more complete platform for voice-first agentic products, expanding the company's addressable market beyond text inference.
Access and Availability
The staged rollout began April 17, 2026 (Pacific Time), giving SuperGrok Heavy subscribers ($30/month) and X Premium+ subscribers (up to $40/month) early access via iOS, Android, and web. Full public API access opened on April 30, 2026 (Pacific Time). The model is currently listed in Oracle Cloud Infrastructure's Generative AI service alongside xAI's API and OpenRouter, broadening enterprise deployment options for teams already operating within those cloud environments.