xAI · Closed-API Model
Grok 4.3 Review: xAI's Aggressive Price Cut, Tested
xAI launched Grok 4.3 on April 30, 2026 and the headline is the price: $1.25 input / $2.50 output per million tokens — roughly 4.8× cheaper than Grok 4's $3 / $15, and well under GPT-5.5 ($5 / $30). You get a 1M-token context window, always-on reasoning you cannot disable, and native video input. This review covers the pricing, benchmarks, what's actually new, and which open-weight models come closest if you need to self-host.
Note: Grok 4.3 is API-only — it cannot be downloaded or run locally. For self-hostable models with comparable performance, see DeepSeek V4, GLM-5, and Qwen3-Coder-Next.
Key takeaways
- →Price collapse: $1.25 / $2.50 per Mtok — ~4.8× cheaper than Grok 4, and cheaper than GPT-5.5 and Claude Opus 4.7.
- →1M context, no fixed output cap — on par with Gemini 3.1 Pro and GPT-5.5.
- →Always-on reasoning — you cannot toggle it off; the model thinks before every reply (~100 tok/s).
- →Native video input — ingests mp4/mov/webm directly; plus a Custom Voices cloning suite.
- →Coding trails — roughly 14 points behind Claude Opus 4.7 on SWE-Bench Verified.
Quick verdict
Grok 4.3 is the cheapest frontier-tier closed model right now, and that — more than any benchmark — is why it matters. If you run high-volume agentic pipelines, process video, or just want a capable reasoning model without GPT-5.5-level token bills, Grok 4.3 is the value pick. Always-on reasoning and a 1M context make it a solid default for long-horizon tasks.
Where it loses: coding. On SWE-Bench Verified it trails Claude Opus 4.7 by roughly 14 points, so for production code the leader is elsewhere. And the cheap headline price only applies up to 200K tokens — long prompts cost more. For local-first deployment, DeepSeek V4-Pro gets you most of Grok 4.3's general capability at zero per-token cost.
Specs at a glance
| Vendor | xAI |
| Release date | April 30, 2026 (beta Apr 17; Custom Voices May 2) |
| API price | $1.25 input / $2.50 output per 1M tokens (≤200K) |
| Context window | 1,000,000 tokens |
| Max output | No fixed cap published |
| Reasoning | Always-on (cannot be disabled) |
| Modalities | Text · Image · Video (input) · File gen (PDF/PPTX/XLSX) · Voice |
| License | Proprietary (xAI Terms of Service) |
| Local self-hostable? | No |
| API model ID | grok-4.3 |
| Knowledge cutoff | December 2025 |
| Access | xAI API (console.x.ai) · Grok app · X Premium+ · OpenRouter · Vercel AI Gateway |
The price collapse
The story of Grok 4.3 is cost. xAI cut input pricing by ~58% and output by ~83% versus Grok 4, landing at a price that undercuts every other frontier closed model. Here is how it stacks up per million tokens:
| Model | Input | Output |
|---|---|---|
| Grok 4.3 | $1.25 | $2.50 |
| Grok 4 (predecessor) | $3.00 | $15.00 |
| Gemini 3.1 Pro | ~$2.00 | ~$12.00 |
| GPT-5.5 (standard) | $5.00 | $30.00 |
| Claude Opus 4.7 | Higher tier | Higher tier |
The catch: the headline rate applies to requests up to 200K tokens. Prompts above that are billed at a higher rate, so whole-codebase analysis at the full 1M context costs more per token than $1.25 suggests. For sustained high volume, a self-hosted open-weight model still wins on total cost.
Sources: xAI / VentureBeat launch coverage (Apr 30, 2026), OpenRouter and Vercel AI Gateway model pages. Competitor prices are list rates and change frequently — verify before budgeting.
Benchmarks
Grok 4.3 is a strong reasoning and agentic model that gives up some ground on pure coding. Where independent trackers disagree on exact figures, we've flagged it — treat coding numbers as approximate.
| Benchmark | Grok 4.3 | Notes |
|---|---|---|
| GPQA Diamond (science) | ~90% | Frontier-tier reasoning. |
| τ²-Bench (agentic tool use) | ~98% | A standout — strong tool/agent behavior. |
| SWE-Bench Verified (coding) | ~low-70s% | ~14 pts behind Claude Opus 4.7 (~87.6%). Figures vary by harness. |
| Vending-Bench (long-horizon agent) | Leads | Reported ~1.26× over Opus 4.7 on long-sequence sim. |
| Throughput | ~100 tok/s | Among the faster frontier-tier API models. |
Sources: Artificial Analysis, vals.ai, and OpenRouter Grok 4.3 benchmark pages (2026); Anthropic Opus 4.7 announcement for the SWE-Bench comparison. Exact SWE-Bench Verified figures for Grok 4.3 vary across trackers (low-70s%); we report a conservative range.
Native video & voice
The other big change in Grok 4.3 is multimodality. The vision encoder ingests video directly — mp4, mov, and webm (reported up to ~5 minutes and 1080p) — and handles transcription, speaker segmentation, object tracking, and motion in a single pass, instead of chaining a separate transcriber, chunker, and language model. That collapses a common multi-tool video pipeline into one API call.
Video input
Direct mp4/mov/webm ingestion. Testers noted occasional hallucinated frame descriptions and intermittent latency — verify on critical tasks.
File generation
Generates PDF, PPTX, and XLSX outputs, plus tool use, vision input, and web search for agentic workflows.
Custom Voices (voice cloning)
Shipped May 2, 2026 alongside Grok 4.3 — clones a voice from ~1 minute of speech in under two minutes, free on the xAI console, sharing TTS/voice-agent APIs with 80+ preset voices.
When to pick Grok 4.3 vs alternatives
| Workload | Best pick | Why |
|---|---|---|
| Cost-sensitive agentic pipelines | Grok 4.3 | $1.25/$2.50 per Mtok + always-on reasoning + ~98% τ²-Bench. |
| Video understanding | Grok 4.3 | Native video input in one pass — few closed models match it. |
| Production coding | Claude Opus 4.7 | SWE-Bench Verified leader (~87.6%); ~14 pts ahead of Grok 4.3. |
| Math + ChatGPT ecosystem | GPT-5.5 | 95.2% AIME; custom GPTs, plugins, mature function calling. |
| Cheapest long-context analysis | Gemini 3.1 Pro | 1M context; competitive per-token cost at long context. |
| Privacy-required workloads | DeepSeek V4 | Self-hostable, MIT licensed, 1M context. |
Open-weight alternatives you can run locally
Grok 4.3 is API-only — no published weights, nothing to download. Its low price helps, but it's still a per-token bill on someone else's servers, with your data leaving your network. If you need a frontier-class model you can self-host for privacy, predictable cost, or offline use, these come closest:
| Model | License | Best for |
|---|---|---|
| DeepSeek V4-Pro | MIT | Closest general-purpose match; 1M context |
| GLM-5 | MIT | Reasoning on a smaller hardware footprint |
| Qwen3-Coder-Next | Apache 2.0 | Coding-heavy workflows (where Grok 4.3 lags) |
| Kimi K2.6 | Modified MIT | 1T MoE; strong agentic + coding |
A note on Grok 5
To be clear: Grok 5 has not been released. As of June 2026, xAI confirmed in its January 2026 funding announcement that the next model — reportedly ~6 trillion parameters with a Mixture-of-Experts architecture — was still in training, and the originally floated Q1/Q2 2026 launch windows passed without a release. Grok 4.3 is xAI's current flagship. Ignore any "Grok 5 benchmarks" you see floating around; nothing official has shipped.
Frequently asked questions
Can I run Grok 4.3 locally?
How much does the Grok 4.3 API cost?
What changed from Grok 4 to Grok 4.3?
Grok 4.3 vs GPT-5.5 vs Claude Opus 4.7: which is best?
Is Grok 4.3 good at coding?
How big is Grok 4.3's context window?
Does Grok 4.3 support video?
Is Grok 5 out yet?
When should I use an open-weight model instead of Grok 4.3?
Cut your AI API bill — run open weights locally
Grok 4.3 is cheap, but it's still a metered API on someone else's servers. The Local AI Master deployment course shows you how to run open-weight alternatives like DeepSeek V4 and Qwen3-Coder-Next on your own hardware — unlimited inference, full data privacy, zero per-token cost.
See the deployment course →Related models
- → GPT-5.5 — ChatGPT default; math leader, pricier at $5/$30
- → Claude Opus 4.7 — current SWE-Bench Verified coding leader
- → Gemini 3.1 Pro — 1M context + thinking tiers
- → DeepSeek V4 — open-weight frontier alternative you can self-host
- → GLM-5 — MIT-licensed reasoning model, smaller footprint
- → Best AI models 2026: complete comparison
Go from reading about AI to building with AI
20 structured courses. Hands-on projects. Runs on your machine. Start free.
Written by the Local AI Master Team
The team behind Local AI Master
We build Local AI Master around practical, testable local AI workflows: model selection, hardware planning, RAG systems, agents, and MLOps. The goal is to turn scattered tutorials into a structured learning path you can follow on your own hardware.