xAI has officially launched Grok 4.20, introducing four reasoning modes: Auto, Fast, Expert, and Heavy. Auto is designed to switch dynamically between Fast and Expert, while Fast prioritizes speed and Expert focuses on deeper reasoning. The announcement was shared by founder Elon Musk on X, where he also said major upgrades to Grok will continue on a weekly basis, with user feedback playing a role in future iterations.
Reasoning options aimed at different workloads
The new mode structure suggests xAI is trying to give users more control over the trade-off between latency and analytical depth. Auto offers adaptive selection, Fast is geared toward quicker responses, and Expert is intended for more demanding reasoning tasks. Heavy was also introduced as part of the release, signaling a broader effort to diversify how Grok handles different categories of prompts and workloads.
Out of beta, but still trailing top rivals
Grok 4.20 officially exited beta testing on the day of the launch. In testing cited from Artificial Analysis, the model posted a 78% non-hallucination rate, pointing to progress in factual reliability. However, it scored 48 on the Composite Intelligence Index, placing it behind competing models including Gemini 3.1 Pro and GPT-5.4. That suggests the new release has made gains in some areas, but still faces a performance gap in broader benchmark comparisons.
API pricing and long-context support
xAI also disclosed API pricing for the model. Input is priced at $2 per million tokens, while output costs $6 per million tokens. Grok 4.20 supports up to 2 million tokens of long context, a specification that could make it more attractive for use cases involving long documents, large-scale analysis, and extended multi-step interactions.
Overall, the launch of Grok 4.20 highlights xAI’s push to accelerate product iteration in the generative AI market. For developers and end users, the key question now is whether the promised weekly upgrades will translate into measurable improvements in capability and competitiveness.