Grok 4.3 Is Live in Kilo Code
xAI’s latest model has always-on reasoning and is priced for everyday use
Grok 4.3 is now available in Kilo Code. This new release from xAI is designed to "think" before it speaks for every query—whether you’re planning or coding or just asking questions—and, impressively, it now sits just above Muse Spark and Claude Sonnet 4.6 on the Intelligence Index.
Wait, but didn’t xAI just release 4.20? Yes. 4.20 was released in a couple variations through the month of March (not April as many were predicting). They’re moving as fast as ever, and Grok 4.3 offers always-active reasoning without the high costs you might expect for such strong capabilities.
If you’ve been paying attention to model usage patterns here, you already know where this is going. Grok models have quietly become some of the most-reached-for models in Kilo for coding work — particularly for the long, agentic sessions where context depth and instruction-following actually matter.
Grok 4.3 is the next step in that line, proving that the Cursor deal was not a distraction for Elon’s empire.
What’s New in 4.3
xAI dropped Grok 4.3 quietly — no press release, just a new entry in the model selector and early testers immediately pushing it through its paces.
The new model carries over the architecture from Grok 4.20, including the 16-agent Heavy system and a large context window, while adding native video understanding, sharper reasoning from extended training runs, and the ability to generate structured documents — slides, PDFs, spreadsheets — directly in conversation.
For coding workflows, what matters most is the reasoning depth. xAI attributes the improvement to longer training runs, and early results are consistent with that: the model holds context better across multi-step agentic tasks, follows complex instructions more reliably, and is less likely to lose the thread mid-session.
It’s more unique than one might expect from xAI’s understated release notes about the beta:
“Grok 4.3 is a new pre-trained model matching the scale of Grok 4.20 with an improved architecture and a December 2025 knowledge cutoff.”
The model is clearly designed for long coding sessions and agentic work. What’s new?
Always-on reasoning, no knobs. Unlike most reasoning models, you can’t dial the effort up or down — reasoning is always active. That’s a deliberate bet that the overhead is worth it for every request.
1M token context, no output limit. That’s the full spec — genuinely uncapped output, which is rare and relevant for long agentic sessions.
Price vs. capability gap is notable. $1.25/$2.50 per million tokens for an always-on reasoning model with a million token context is unusually low for this class. The closest comparisons run 5–10x higher — although we should note that, as with a number of new models charging more for higher context, requests exceeding 200k total tokens are billed at a higher rate.
That pricing is significant. Combining benchmark data, Artificial Analysis estimates that Grok 4.3 reduces overall running costs by approximately 20% compared to Grok 4.20. It’s meaningul that xAI chose not to raise the price versus Grok 4.20, which seems like a play to become the new daily driver in agents like Kilo, as Grok Code Fast used to be (even after it stopped being free).
That’s not a cost you need to justify. For most coding tasks in Kilo — file edits, multi-turn refactors, agent loops — you’re looking at fractions of a cent per run. A model with a million-token context window and always-on reasoning at that price used to be a contradiction. Now it’s just a typical day using Kilo Code.
This has always been part of why Grok models have done well in Kilo. Developers run them harder and longer than they’d run a $15/M model, which means they get more out of each session. Grok 4.3 extends that logic with meaningfully better performance.
The Best Way to Run It
Grok 4.3 is already live in the Kilo model switcher, which means that anybody can give it a run in the IDE, CLI, cloud agents, KiloClaw and beyond. No waitlist, no tier gating.
Congrats to xAI on this new release. It’s one of the most credit-efficient high-intelligence models you can run in Kilo, which means it stacks especially well with Kilo Pass if you want to get more bang for the buck.
Give it a spin and let us know how it feels!
Curious about AI coding models? Join us for KiloWorks, a practical AI workshop featuring talks from AI researchers at leading labs.