Grok 4.3: xAI's Cheap Frontier Model (May 2026 Guide)

John -
Quick answer. Grok 4.3 is xAI's reasoning-first flagship released April 30, 2026 (general API). It ships a 1M token context window, native video input (an xAI first), and aggressive pricing: $1.25 per 1M input tokens and $2.50 per 1M output tokens. It scores 53 on the Artificial Analysis Intelligence Index, sitting below GPT-5.5 (60) and Gemini 3.1 Pro Preview (57), but at a fraction of their price.

What is Grok 4.3?

Grok 4.3 is xAI's latest reasoning-centric large language model, succeeding Grok 4.20. It went into beta on grok.com and the SuperGrok mobile apps on April 17, 2026, and opened up on the public API on April 30, 2026, with the broader launch wave landing the week of May 4, 2026. xAI positions it as a cost-efficient frontier model: chain-of-thought reasoning is on by default for every request, and the model is multimodal across text, images, and now video.

Unlike previous Grok versions where reasoning was a separate toggle, 4.3 folds always-on reasoning into the base behavior. That means slightly higher time-to-first-token, but consistently stronger results on multi-step tasks without the developer having to pick the right "mode."

What's new in Grok 4.3?

Six concrete additions over Grok 4.20:

  • Native video input. The vision encoder now ingests video directly (mp4/mov/webm, up to 5 minutes, up to 1080p) instead of requiring frame extraction or transcription upstream. xAI says this covers speech transcription, speaker segmentation, object tracking, and motion causality in a single pass.
  • 1M token context window. Up from previous Grok generations; enough to fit roughly 1,500 pages of text, a full codebase, or hours of meeting transcripts in one request.
  • Native artifact generation. The model can directly produce PDFs, PowerPoint-style slide decks, and spreadsheets as outputs, not just describe how to build them.
  • Stronger agentic tool calling. xAI reports an Elo gain of over 300 points on GDPval-AA (a tool-use and agentic benchmark) versus Grok 4.20.
  • Higher factual accuracy. Artificial Analysis reports an 8-point gain in AA-Omniscience accuracy versus Grok 4.20, though the non-hallucination rate slipped by 8 points in the same suite — Grok 4.20 still leads xAI's lineup on that specific metric.
  • Aggressive price cut. Per Artificial Analysis, input is 37.5% cheaper and output 58.3% cheaper than Grok 4.20, with prompt caching at $0.20 per 1M tokens for repeated context.

Output throughput sits at around 156 tokens per second on the Artificial Analysis benchmark — fast for a reasoning model, though time-to-first-token (~19.7s) is on the slow side because of the always-on chain-of-thought.

How does Grok 4.3 benchmark?

On the Artificial Analysis Intelligence Index, Grok 4.3 (high) scores 53. For context, the median across tracked reasoning models is around 36, so it sits comfortably above average. The standout single-benchmark gain is on GDPval-AA, where Grok 4.3 reached an Elo of 1500 — a 321-point jump from Grok 4.20's 1179.

The reasoning gains are real but uneven. Grok 4.3 hits roughly 90% on GPQA Diamond and is strong on math and analytical tasks, but on agentic coding benchmarks like SWE-bench Pro, it trails Anthropic's Claude Opus 4.7 by double-digit percentage points. xAI has not published its own SWE-bench Verified number, which is itself a tell.

The honest summary: 4.3 is no longer chasing the top of every leaderboard. It is chasing the best price-per-intelligence ratio in the frontier tier, and it wins that fight cleanly.

How does Grok 4.3 compare to Claude Opus 4.7?

Different tools for different jobs.

  • Coding. Claude Opus 4.7 still holds the industry gold standard on SWE-bench Verified (~87%) and SWE-bench Pro (~64%). On hard cross-file refactors, complex repo bugs, and mission-critical code, Opus 4.7 is meaningfully ahead. For function-level generation, unit tests, simple bug fixes, and long-chain agents, third-party comparisons put the gap at under 5 percentage points.
  • Price. Opus 4.7 lists at $5 / $25 per 1M input / output. Grok 4.3 lists at $1.25 / $2.50. That is roughly a 5x gap on input and 10x on output. For high-volume agentic workflows, the difference is the difference between "affordable experiment" and "line-item in the P&L."
  • Context. Both offer 1M-token context windows.
  • Multimodal. Grok 4.3 has native video, which Opus 4.7 does not match natively.

Rule of thumb: if the work is hard coding or sensitive long-horizon agent loops, pay for Opus 4.7. If the work is high-volume reasoning, video understanding, document Q&A, or anything where cost-per-call matters, 4.3 is the rational pick. For a deeper look at the Anthropic side, see our Claude Opus 4.7 complete guide.

How does Grok 4.3 compare to GPT-5.5 and Gemini 3.1 Pro?

Both OpenAI and Google currently sit above Grok 4.3 on aggregate intelligence indices. GPT-5.5 (xhigh) leads at 60 on the Artificial Analysis Intelligence Index; Gemini 3.1 Pro Preview scores 57; Grok 4.3 (high) scores 53.

But Grok 4.3 wins the price-per-token race against both:

  • Gemini 3.1 Pro Preview is roughly $2.00 input / $12.00 output per 1M tokens — 60% more on input, almost 5x more on output.
  • GPT-5.5 sits in a similar premium tier.

If you are building a product where you can tolerate "top-5 frontier intelligence" instead of "absolute best," Grok 4.3 currently delivers it for half to a fifth of the cost. If you need the actual ceiling — best-in-class on the hardest reasoning, coding, or multimodal benchmarks — pay for GPT-5.5, Gemini 3.1 Pro, or Claude Opus 4.7. Our GPT-5.5 complete guide covers the OpenAI side in detail.

How much does Grok 4.3 cost?

Public API pricing per the xAI model docs, current as of May 2026:

  • Input: $1.25 per 1M tokens (text or image)
  • Output: $2.50 per 1M tokens
  • Cached input: $0.20 per 1M tokens (84% discount on repeated context)
  • Long-context tier: requests over 200,000 input tokens are billed at double the standard rate

On the consumer side, the chat experience at grok.com runs free with rate limits, while SuperGrok Heavy at $300/month unlocks the highest reasoning effort, longer sessions, and the new artifact-generation features.

For a typical agentic workload — long context in, short structured response out — Grok 4.3 ends up roughly 5–10x cheaper than Opus 4.7 and 4–5x cheaper than Gemini 3.1 Pro Preview at the same call volume.

How do I access Grok 4.3?

Three paths, depending on what you are building:

  1. Consumer chat: grok.com (or the iOS/Android Grok apps). Free with limits; SuperGrok subscriptions unlock heavier use.
  2. xAI API: sign up at console.x.ai, generate an API key, and call the OpenAI-compatible endpoint at https://api.x.ai/v1. The model name is grok-4.3. New accounts receive $25 in promotional credits.
  3. Third-party gateways: OpenRouter, Requesty, and similar routers expose grok-4.3 alongside the rest of the frontier catalog if you prefer a single billing relationship across providers.

Because xAI's API speaks the OpenAI SDK dialect, switching an existing GPT-based client to Grok 4.3 is usually a base URL plus model name change — minutes of work, not days. xAI also retired eight legacy Grok models on May 15, 2026, so any apps still pointing at older Grok versions need to migrate (requests to the deprecated slugs now silently route to grok-4.3 at the new pricing).

When should you use Grok 4.3?

Strong fits:

  • High-volume agentic workflows where every cent per call shows up at the bottom of the month
  • Video understanding and analysis — Grok 4.3 is currently the only API model in this tier with native video input
  • Long-document Q&A and codebase exploration where you want the full 1M context window without paying frontier prices
  • Latency-tolerant pipelines that can absorb the ~20s time-to-first-token in exchange for cheap reasoning

Weaker fits:

  • Hard agentic coding (cross-file refactors, complex repo bugs) — pay for Opus 4.7
  • Sub-second response latency requirements — chain-of-thought is always on, so first tokens take a while
  • The absolute top of reasoning, math, or multimodal leaderboards — GPT-5.5 and Gemini 3.1 Pro still lead

If you are scoping a new feature, the sane move is to A/B Grok 4.3 against your incumbent on the actual task you care about. The pricing makes it cheap to run that bake-off honestly. For a wider view of what landed this month, see our AI model releases May 2026 roundup.

FAQ

When was Grok 4.3 released?

Grok 4.3 entered beta on grok.com and the SuperGrok apps on April 17, 2026, opened to the public API on April 30, 2026, and reached general availability the week of May 4, 2026.

How much does the Grok 4.3 API cost?

$1.25 per 1M input tokens and $2.50 per 1M output tokens. Cached input is $0.20 per 1M tokens. Requests over 200,000 input tokens are billed at double the standard rate.

What is the Grok 4.3 context window?

One million tokens. There is no documented hard cap on output tokens, though latency scales with output length.

Is Grok 4.3 better than Claude Opus 4.7?

No, not on the hardest agentic coding benchmarks — Opus 4.7 still leads SWE-bench Verified and SWE-bench Pro by double digits. Grok 4.3 closes the gap on most general reasoning tasks and is roughly 5–10x cheaper, which makes it the better default for cost-sensitive workloads.

Does Grok 4.3 support video?

Yes. Grok 4.3 is the first xAI model with native video input via the vision encoder. It accepts mp4/mov/webm up to 5 minutes at 1080p, and handles speech transcription, speaker segmentation, object tracking, and motion reasoning in a single pass.

How do I get a Grok API key?

Sign up at console.x.ai, create an account, and generate an API key from the API Keys page. The API is OpenAI-compatible — point your existing client at https://api.x.ai/v1 and pass grok-4.3 as the model name.

What is the difference between SuperGrok and the Grok 4.3 API?

SuperGrok is the consumer chat subscription on grok.com (and the mobile apps). The API is the developer surface at api.x.ai. Same underlying models, separate billing, separate rate limits, separate accounts.