Grok 4.20

“A capable, supported legacy flagship — but if my workload fits 1M tokens, I migrate to the cheaper, newer Grok 4.3.”

Strategically, Grok 4.20 is now a continuity choice rather than a new bet. It retains real value where a runtime reasoning toggle, long context, and best-in-class non-hallucination matter together — but xAI itself recommends 4.3 as the default, signaling where investment and roadmap attention go. Vendor risk is the same as the family: thin disclosure, no certs, no published safety framework. Lock-in is low (SDK compatibility). The decision is simple: keep 4.20 only for the specific niches; otherwise plan migration. Roadmap confidence in 4.20 specifically is declining as 4.3 absorbs its use cases.

“Its one durable edge is the lowest hallucination rate at launch — but the market has moved to 4.3, and that's where the moat now lives.”

In positioning terms, Grok 4.20 carries the same structural moat as the family — live X data — but the differentiation has migrated to Grok 4.3, which is cheaper, fresher, and adds video. 4.20's distinct selling point is reliability: the best-in-class non-hallucination rate gives it a credible pitch to error-sensitive verticals. But market timing works against it: launched March, superseded by April's 4.3, repriced to match it. As a standalone competitive play it has limited runway; its role is to retain users until they migrate. Differentiation versus non-xAI rivals rests entirely on the X-data and reliability angles.

“Now repriced to match 4.3 — so there's no cost reason to choose 4.20 over the newer model unless you specifically need its niche.”

At launch, 4.20's $2 / $6 made it a clear loser to 4.3's $1.25 / $2.50. xAI has since repriced 4.20 down to $1.25 / $2.50 / $0.20 cached on the docs card, erasing the cost penalty — but that also erases any financial reason to prefer it, since 4.3 is the same price with newer training and a higher AA Index. The lingering finance risk is the source conflict: AA still bills it mentally at $2 / $6 with $1.10 cached, so anyone modeling from aggregators will misprice. Rule for finance: confirm the live docs.x.ai rate, then ask whether 4.3 wouldn't simply be the better-value choice at identical pricing.

“The reasoning toggle is the real win — ship one integration, flip a boolean between fast and deep, and parse cleaner output thanks to strict adherence.”

For builders, 4.20's standout is the runtime reasoning toggle: one integration covers both a fast conversational mode and a deep reasoning mode, switched by a parameter or slug rather than a model swap. Strict prompt adherence means less defensive parsing code — when you say "JSON only," you get JSON. Tool calling, structured outputs, and the live-X search tool all work cleanly. The 1M context simplifies long-context plumbing. Friction: SDK/docs polish trails OpenAI and Anthropic, reasoning visibility is summary-only, and coding is better done elsewhere. Reliability is decent; rate limits are spend-tiered.

“It was the standard Grok through April — looser, opinionated, live X data — but consumer surfaces have since moved everyone to 4.3.”

For everyday users on grok.com or X Premium, Grok 4.20 was the default Grok experience through April 2026: the looser personality, ready opinions, and live-X integration that define the brand. Its 2M/1M context is invisible to typical chat use. Once 4.3 became the consumer default, users migrated automatically, so most direct 4.20 use today is via teams pinned to a specific API slug for stability. As a daily driver it was fine; it is simply no longer the one most people touch.

“The 2M-context headline is already contradicted by xAI's own current docs (1M), and the pricing it launched at quietly evaporated — read the live card, not the launch post.”

Adversarially, Grok 4.20 is a case study in why xAI's thin disclosure matters. Its marquee launch claim — 2M context — now conflicts with xAI's own docs card showing 1M for the reasoning/non-reasoning slugs, while AA and OpenRouter still show 2M; nobody should cite a Grok context number without checking the live source. The launch pricing ($2 / $6) silently dropped to $1.25 / $2.50, so any cost analysis older than a few weeks is wrong. Architecture is undisclosed, no SWE-bench, no safety framework. The one claim that holds up well is the non-hallucination leadership (AA-Omniscience 78%) — that is independently sourced and genuinely strong.