Grok 4.3 and the Grok Build Surge: What xAI Just Shipped

Five days, six product surfaces, one direction of travel. Between June 16 and June 21, 2026, xAI pushed an image model upgrade, renamed a coding tool, teased web app hosting, and shipped two Grok Build CLI updates — all while Grok 4.3 quietly continued accruing third-party benchmark coverage. None of it arrived as a keynote. It arrived as changelogs, X posts, and a quiet OpenRouter listing.

TLDR Grok 4.3 is xAI's current flagship reasoning model, released April 30, 2026, with a 1M token context window and tiered pricing at $1.25/$2.50 per 1M tokens (source). The bigger story in mid-June 2026 is the cadence around it: Grok Imagine 1.5 Fast cut 720p video generation from 40+ seconds to ~25 seconds (source), Cursor Composer 2.5 was folded into Grok Composer inside Grok Build, and built-in free web app hosting on xAI hardware is reportedly imminent. xAI has not published first-party benchmarks for 4.3; the public scoring comes from Artificial Analysis, which places Grok 4.3 (high) at 38 on its Intelligence Index.

Key Takeaways

• Grok 4.3 lists at $1.25 input / $2.50 output per 1M tokens with a 1M token context window and an 84% cache-hit discount to $0.20 per 1M.
• Grok Imagine 1.5 Fast advertises ~25-second 720p generation, down from 40+ seconds in the previous model.
• Cursor Composer 2.5 has been renamed Grok Composer inside Grok Build, signaling a product-line merge under xAI branding.
• Grok Build shipped versions 0.2.57 and 0.2.60 within four days, focused on terminal reliability, MCP behavior, and session resume.
• Free hosted deployment of Grok-generated web apps on xAI hardware is being previewed, with no firm launch date.
• No first-party Grok 4.3 benchmarks have been published; third-party scores from Artificial Analysis rank it #37 of 154 on intelligence.

What xAI Actually Shipped This Week

The verifiable surface area, drawn directly from the signal set:

• Grok 4.3 API listing. Live on OpenRouter as of April 30, 2026, priced at $1.25 per 1M input tokens and $2.50 per 1M output tokens, with a 1M token context window and tiered billing for requests over 200K total tokens (OpenRouter listing).
• Grok Imagine 1.5 Fast. Rolled out June 17, 2026. The advertised change is generation speed: 720p videos render in about 25 seconds, down from 40+ in the prior model (TestingCatalog post).
• Grok Composer rename. On June 17, Mark Kretschmann reported that Cursor Composer 2.5 was renamed Grok Composer inside Grok Build, with the two model lines being merged into one.
• Grok Build 0.2.57. Released June 18. Adds network-blip resume for long responses, marketplace-based plugin installs via grok plugin install, fixes for indefinite compaction hangs, and HTML entity rendering corrections in tool output (changelog summary).
• Imagine search. On June 20, TestingCatalog noted that users can now search their Imagine creations — image and video — through a scoped search interface.
• Web app hosting reveal. On the same day, Nima Owji previewed that Grok will soon let users host generated web apps directly on xAI hardware.
• Grok Build 0.2.60. Released June 21. Surfaces current-repo sessions first in /resume, makes slash command completion remember recent commands, fixes compaction hangs when summarizer streams stall, and truncates large MCP tool results inline.

That is what is observable. Everything beyond it is inference.

The Cadence Story: Why Two CLI Updates in Four Days Matter

The single most legible signal in this batch is not Grok 4.3 itself — it is the rate of change in Grok Build. Two minor releases in four days, both touching long-session reliability (compaction hangs, network blip resume, MCP truncation), points to a product being shaped under live developer load rather than in a closed beta.

That cadence matters because it changes what kind of tool Grok Build is. A CLI that ships patches on a multi-day rhythm and merges in a previously separate product (Cursor Composer) inside a week is behaving like the internal tool of a team that is using it daily on real codebases. The bug fixes are revealing: "Long responses now resume after network blips instead of failing the turn" describes a problem that only surfaces on long-horizon coding tasks. "Compaction no longer hangs indefinitely when the summarizer stream stalls" describes a problem that only matters when conversation history exceeds the model's working context.

These are the failure modes of agentic work, not of single-turn chat. The Grok Build Cadence — call it that — is the operational tell that xAI is building toward long-running developer agents, not just a chat surface.

Grok Composer and the Cursor Question

The June 17 rename of Cursor Composer 2.5 to Grok Composer is the most consequential business signal in the bundle. Mark Kretschmann's framing is direct: "The two lines of models are being merged into one." A separate post from @op7418 on June 16 claims SpaceX acquired Cursor for $60B via a stock swap. That claim is unverified — it appears in a single Chinese-language tweet with no corroborating primary source in the signal set — and should not be treated as confirmed.

What is verifiable is the rename itself. Grok Composer now lives inside Grok Build. Whatever the underlying corporate structure, the product line consolidation is real and shipped. For developers, the practical question is whether the Composer coding model that powered Cursor's autonomous code edits is now a Grok-branded capability available through xAI's API. The signal set does not yet confirm that. The Grok 4.3 OpenRouter listing makes no mention of a Composer-derived variant.

Grok 4.3: The Specs That Are Actually Public

For the model itself, the publicly verifiable specifications are narrower than the surrounding noise suggests:

Dimension	Value	Source
Release date	April 30, 2026	OpenRouter
Context window	1M tokens	OpenRouter
Input price	$1.25 per 1M tokens	OpenRouter
Output price	$2.50 per 1M tokens	OpenRouter
Cache hit price	$0.20 per 1M tokens (-84%)	Artificial Analysis
Reasoning levels	none / low / medium / high (default: low)	OpenRouter
Input modalities	text, image	Artificial Analysis
Output modality	text	Artificial Analysis
Tiered billing threshold	200K total tokens	OpenRouter

The Artificial Analysis Intelligence Index places Grok 4.3 (high) at 38, which ranks it #37 of 154 evaluated models, with output speed of 151.8 tokens per second (#19 of 154). Their cost to evaluate the full Intelligence Index on Grok 4.3 (high) was $318.53, with 83M output tokens consumed — flagged as "very concise" against an average of 100M.

What is not public from xAI in this signal set: parameter count, training corpus size, post-training methodology, any first-party benchmark report, or a system card. The earlier community estimate for Grok 4 of ~2.4T parameters circulated by Smol AI (per Interconnects) is not confirmed for Grok 4.3 and should not be propagated forward.

The Configurable Reasoning Lever

The most interesting architectural exposure in the OpenRouter listing is the reasoning effort control: none, low, medium, or high, defaulting to low. This is configurable test-time compute — the same lever GPT-5 series and Claude Opus 4.x reasoning variants expose under different names. Grok 4.3 making it a first-class API parameter, with the default at low rather than medium, is a pricing-aware design. It tells callers that the cheap path is the expected path, and high reasoning is the explicit opt-in.

The Artificial Analysis score of 38 specifically applies to Grok 4.3 (high). The score for default-low has not been published in the signal set. Builders running cost-sensitive workloads should not assume the Intelligence Index figure transfers to the cheaper reasoning tier.

Grok Imagine 1.5 Fast and the Video Latency Race

Grok Imagine 1.5 Fast is the standalone video update of the week. The advertised improvement: 720p generation in ~25 seconds, down from 40+ seconds. That is roughly a 1.6× speedup at the same resolution, and TestingCatalog's framing emphasizes both speed and quality ("better quality and faster generation time"). Mark Kretschmann's hands-on take from June 17 reports cleaner motion, better consistency, and surprisingly cinematic sound.

These are community impressions, not measured comparisons. No FVD, no human preference numbers, no third-party video quality benchmark accompanies the rollout in this signal set. The 25-second figure is the only quantitative claim. For builders comparing Grok Imagine 1.5 Fast against alternatives, that single number is the only one currently grounded in a vendor statement; everything else is vibes-tier evidence.

Web App Hosting: The Loop That Closes the Deployment Gap

The signal with the largest potential downstream effect is the smallest in actual shipped surface: built-in web app hosting. Nima Owji's June 20 post and Mark Kretschmann's same-day analysis both describe a flow where apps generated inside Grok can be deployed on xAI hardware for free, with usage limits, without external infrastructure setup.

The framing matters. Kretschmann: "you can go from idea to generated app to live URL without messing around with external hosting, deployment setup, or infrastructure. The smoother that loop gets, the more people will actually build." That is a vendor lock-in design dressed as a developer experience win — and it is also, accurately, a real reduction in friction for the audience xAI is courting.

What is unverified: the hosting tier, the limits, the bandwidth caps, the data persistence guarantees, the custom domain story, and the launch date. Both posts use "soon," and "soon" in this signal set has no T-zero attached to it.

Grok 4.3 vs Claude Opus 4.x: What the Signal Says

Claude is the only competitor model family named alongside Grok in the link snapshot bundle (via Hackceleration's positioning of Grok as "a serious alternative to ChatGPT and Claude" and the GlobalGPT model list, which surfaces both Claude Opus 4.7 and Grok 4.3 in the same tool catalog). A direct head-to-head comparison from the signal set:

• Context window. Grok 4.3: 1M tokens (OpenRouter). Claude Opus 4.7: unverified — no public number in this signal set.
• Input pricing. Grok 4.3: $1.25 per 1M tokens. Claude Opus 4.7: unverified — no published price in this signal set.
• Output pricing. Grok 4.3: $2.50 per 1M tokens. Claude Opus 4.7: unverified — no published price in this signal set.
• Reasoning controls. Grok 4.3: explicit none/low/medium/high parameter. Claude Opus 4.7: unverified in this signal set.
• Third-party benchmark. Grok 4.3: Artificial Analysis Intelligence Index 38 (#37 of 154). Claude Opus 4.7: not scored in this signal set.

The honest read: this signal set does not support a numerical Grok 4.3 vs Claude Opus 4.7 comparison. Hackceleration's editorial framing positions Grok against ChatGPT and Claude on the basis of "less restrictive content policies" and "real-time X data" rather than on capability deltas. That is positioning, not measurement. Treat any blog post that claims a head-to-head benchmark between these models — without citing a first-party report from either lab — with skepticism.

What We Know vs. What We Don't

What we know (from primary listings and vendor-adjacent posts):

• Grok 4.3 supports a 1 million token context window with no published output token limit, according to its OpenRouter listing.
• Grok 4.3 is priced at $1.25 per 1M input tokens and $2.50 per 1M output tokens on its listed API endpoint, with cache hits billed at $0.20 per 1M tokens — an 84% discount versus list.
• Grok 4.3 was released by xAI in April 2026, per the OpenRouter and Artificial Analysis listings dated April 30, 2026.
• Grok Imagine 1.5 Fast renders 720p videos in about 25 seconds, down from 40+ seconds in the prior model, according to xAI's rollout note surfaced by TestingCatalog.
• Grok Composer is the renamed version of Cursor Composer 2.5 inside Grok Build, signaling that xAI is merging the two product lines into one, as reported by Mark Kretschmann.
• xAI appears to be adding free, limit-bounded hosting for Grok-generated web apps directly on xAI hardware, based on a feature reveal by Nima Owji and independent commentary from Mark Kretschmann.
• Grok Build 0.2.60 reordered /resume to surface current-repo sessions first, made slash command completion remember recent commands, stopped compaction from hanging when summarizer streams stall, and truncated large MCP tool results inline.
• Grok Build 0.2.57 added network-blip resume for long responses, enabled marketplace-based plugin installs via grok plugin install, fixed indefinite compaction hangs, and corrected HTML entity rendering in tool output.

What we don't know (open questions):

• No. The signal set contains no official parameter count for Grok 4.3 — earlier community estimates for Grok 4 (around 2.4T parameters) are not confirmed for the 4.3 release.
• Not in this signal set. Artificial Analysis publishes a third-party Intelligence Index score of 38 for Grok 4.3 (high), but xAI has not released a first-party benchmark report alongside the model.
• No. A single Chinese-language tweet from @op7418 claims a $60B SpaceX-Cursor acquisition via stock swap, but no primary source corroborates it in this signal set — treat as rumor.
• Unknown. Both Nima Owji and Mark Kretschmann describe the feature as imminent ("will soon"), but no firm date, pricing tier, or usage limit has been published by xAI.

What Builders Should Do Today

Three concrete actions, in priority order:

First, treat Grok 4.3 as a price-sensitive reasoning model rather than a frontier benchmark champion. The Artificial Analysis Intelligence Index of 38 is competitive against models of similar price, but it is not best-in-class. At $1.25/$2.50 per 1M tokens with an 84% cache discount, Grok 4.3 is well-positioned for high-volume agentic workloads where the per-token cost dominates the per-task quality. Run your own coding eval at the default low reasoning tier before assuming the high-tier benchmark transfers.

Second, watch Grok Build's release cadence as the leading indicator. Two minor releases in four days, both targeting long-session reliability, is the operational signature of a tool being prepared for serious agentic work. If you build coding agents, fork Grok Build's changelog into your weekly review. The patches reveal the workloads xAI's internal users are running.

Third, do not propagate unverified claims. The SpaceX-Cursor acquisition rumor, the parameter count carry-over from Grok 4, and the "Opus-class" community impressions are not benchmarks. They are noise. The Grok Imagine 1.5 Fast 25-second figure is the only vendor-stated number in the imaging batch — anchor public claims to it, not to vibe reports.

The Week Ahead

Three things to track. Watch xAI's official accounts for a Grok 4.3 system card or benchmark report — its continued absence is itself a signal about how xAI wants 4.3 evaluated (cost and speed, not raw intelligence). Run the Grok Composer rename to its conclusion: if xAI publishes an API surface for Grok Composer separate from Grok 4.3, that confirms the Cursor coding stack is now a first-class xAI product. Pin a date on the web app hosting reveal — "soon" from June 20 should become a concrete launch announcement within two to four weeks if the feature is real and near-ready, and any longer silence suggests the rollout slipped.

Want to call Grok 4.3 via API? kie.ai has it.