Grok 4.3 Surfaces: What the xAI Codename Shuffle, Cursor Rumor, and Imagine 1.5 Rollout Actually Tell Us

Grok 4.3 did not arrive with a livestream. It arrived as an OpenRouter listing, a Cursor-shaped rumor, a video model bump, and a one-line product rename — four small signals in 72 hours that together describe what xAI's release cadence now looks like.

TLDR Grok 4.3 is publicly listed at $1.25/$2.50 per 1M tokens with a 1M-token context and configurable reasoning effort, per the OpenRouter model page. Artificial Analysis measured the high-reasoning variant at 38 on the AA Intelligence Index v4.1 and roughly 152 output tokens per second. Around it, Grok Imagine 1.5 Fast shipped, Cursor Composer 2.5 was renamed to "Grok Composer", and Grok Build 0.2.57 dropped with a terminal-reliability changelog. No first-party xAI post for the 4.3 model is visible in the signals — making this a Quiet Catalog Release.

Key Takeaways

• Grok 4.3 is a reasoning model with four effort levels (none/low/medium/high, default low) and a 1M-token context window per OpenRouter.
• Public pricing lands at $1.25 input / $2.50 output per 1M tokens, with cached input at $0.20 — an 84% cache discount.
• Artificial Analysis ranks Grok 4.3 (high) at 38 on AA Intelligence Index v4.1 (rank 37/154) with 83M tokens generated to complete the index.
• xAI appears to be merging the Cursor Composer line into Grok Composer inside Grok Build, per community observation.
• Grok Imagine 1.5 Fast cut 720p render time from 40+ seconds to roughly 25 seconds, according to TestingCatalog.
• No first-party xAI announcement post for the 4.3 model has surfaced in this signal set — most fact-anchoring runs through third-party catalogs.

What Actually Surfaced This Week

Four discrete events. None of them is a launch event. All of them are real.

First, Grok 4.3 has a live, fact-anchored listing on OpenRouter, dated April 30, 2026, single-provider, with the spec sheet laid out: text and image input, text output, 1M context, configurable reasoning effort (none/low/medium/high, default low), and tiered pricing where requests above 200K total tokens get billed at a higher rate. Artificial Analysis independently measured the high-reasoning variant at 38 on the AA Intelligence Index v4.1, 151.8 output tokens per second, $318.53 total to run the index — and noted the model generated 83M tokens, well below the 100M cross-model average. Verbosity rank: 7 of 154.

Second, Grok Imagine 1.5 Fast rolled out, with TestingCatalog noting that 720p videos now render in about 25 seconds, down from 40+ in the prior generation. Mark Kretschmann's hands-on take reports cleaner motion, better consistency, and more cinematic audio — a community impression, not a measured eval.

Third, Cursor Composer 2.5 was renamed to "Grok Composer" inside Grok Build, which Kretschmann reads as the two product lines being merged into one. Around the same time, a Chinese-language post by @op7418 circulated an unverified claim that SpaceX acquired Cursor for $60 billion in a stock-based deal. Treat the acquisition number as community rumor, not a confirmed corporate event.

Fourth, Grok Build 0.2.57 shipped with a changelog focused on terminal reliability: long responses now resume after network blips instead of failing the turn, grok plugin install <name> can pull from registered marketplaces (not just local paths), long-running conversation compaction no longer hangs, notification hooks fire only on real user-attention events, and HTML entities like < and > finally render correctly in tool output.

The Quiet Catalog Release Pattern

Coined term, used deliberately: a Quiet Catalog Release is what happens when a model's existence is established by third-party aggregator pages, OpenAI-compatible router listings, and resold inference endpoints before any first-party launch post lands — or instead of one. Grok 4.3 fits this shape exactly. The OpenRouter page is dated April 30, 2026. The Artificial Analysis page calls it "Released April 2026." GlobalGPT lists Grok 4.3 in its catalog alongside Claude Opus 4.8, GPT-5.5, and Gemini 3.5 Flash. An @aditiitwt post on June 18 framed Grok 4.3 reaching GlobalGPT as "BREAKING NEWS," which is itself a tell — it's news because the broader market is still catching up to a model that has, mechanically, already shipped.

This matters for builders because the model can be benchmarked, priced, and integrated before xAI has produced a model card to argue about. The community is now the first-pass evaluator.

Pricing and Spec Sheet

Dimension	Grok 4.3	Source
Input price	$1.25 / 1M tokens	OpenRouter
Output price	$2.50 / 1M tokens	OpenRouter
Cache write	$1.25 / 1M tokens	Artificial Analysis
Cache hit	$0.20 / 1M tokens (-84%)	Artificial Analysis
Context window	1,000,000 tokens	OpenRouter
Tiered billing	Above 200K total tokens	OpenRouter
Reasoning effort	none / low / medium / high (default low)	OpenRouter
AA Intelligence Index v4.1	38 (rank 37/154)	Artificial Analysis
Output speed	151.8 tok/s (rank 19/154)	Artificial Analysis
Verbosity (tokens to complete index)	83M (rank 7/154, concise)	Artificial Analysis
Release date listed	April 30, 2026	OpenRouter
Provider count on OpenRouter	1	OpenRouter

Two numbers anchor the value argument. The 84% cache discount on input tokens is aggressive — most reasoning models in this tier charge closer to the full input rate even on cache hits. And at $2.50 output, Grok 4.3 sits well under the $8.05 average output price Artificial Analysis cites for comparable models, while still placing in the top quartile on the Intelligence Index.

Grok 4.3 vs Grok 4: What the Signal Says

The only first-party Grok comparison available in this signal set is the Interconnects deep-dive on Grok 4, which framed the earlier model as o3-class, rumored at roughly 2.4T parameters, with Grok 4 Heavy gated behind a $300/month tier. That gives a baseline.

Dimension	Grok 4.3	Grok 4
Context window	1M tokens (OpenRouter)	"Top overall by ArtificialAnalysis, dethroning Gemini 2.5 Pro on long context" (Interconnects)
Input/output price	$1.25 / $2.50 per 1M tokens (OpenRouter)	unverified — no public per-token number from either lab in this signal set
Parameter count	unverified — no public number from this signal set	"Rumored to be 2.4T params" per Interconnects, citing Swyx
Reasoning effort control