Grok 4.20 drops the beta tag and dares users to feed it two-million-token tomes
xAI just yanked the safety net off Grok 4.20, pushing the chatbot onto the public stage with a two-million-token context window that swallows entire codebases, legal briefs, or season-long TV scripts in a single gulp.
Musk promises weekly brain transplants
The serial upgrader tweeted from his X account that the model will receive “major updates every week,” a cadence that turns the usual quarterly AI refresh cycle into roadkill. Early testers report that the multi-agent architecture spins up sub-models on demand, letting one instance handle math proofs while another drafts Python, all under the same conversational thread.
Speed gains come from a rewritten inference stack that trims first-token latency to 240 milliseconds on Nvidia H100 clusters, according to benchmarks the company quietly slipped onto GitHub last night. That is 38 % faster than the beta build leaked two weeks ago on Spanish Discord servers.
Four personalities, zero hallucinations—almost
Users can toggle between Auto, Fast, Expert and Advanced modes. The last one chains Monte-Carlo rollouts until it hits a self-imposed confidence ceiling, a trick that xAI claims pushes factual error rates below 0.9 % on the MMLU suite. Even if marketing gloss inflates that figure, the raw logs circulating on Hacker News show fewer phantom citations than Claude 3 Opus in head-to-head trials.
The catch: you need an X Premium+ subscription, now priced at $50 a month after last week’s hike. Musk is betting that power users will trade the price of two lattes for the ability to park an army of agents inside one chat window.
Enterprise customers get an on-prem sandbox later this quarter, letting banks and biotech labs keep their million-token prompts off the social graph. The move positions xAI to raid OpenAI’s lucrative compliance-driven client roster before GPT-5 hits the circuit.
Two million tokens is not a party trick; it is a declaration of war on context-hungry rivals. If Grok 4.20 keeps its weekly promise, the next six months will feel like dog years for the rest of the field.