Grok 4.3

What is Grok 4.3?#

Grok 4.3 is xAI's fast, low-cost reasoning model, released April 30, 2026 and built for tool calling and instruction following rather than frontier coding. It accepts text and image input, outputs text, and runs on Unifically as xai/grok-4.3. The context window is 1M tokens, and reasoning effort runs from none through low and medium to high, so the same model answers simple requests without thinking and spends more only on the hard ones. When xAI retired Grok 4 and Grok 3 in May 2026, Grok 4.3 was the model it pointed those IDs at.

Key features of Grok 4.3#

Fast decode at a low cost per task

Artificial Analysis measures 115.8 output tokens per second, #36 of 195 models, and $0.15 per Intelligence Index task, #6 of 195. Its Intelligence Index of 38 sits above the class median of 36, so the speed does not come from a weak model.

Tool calling and instruction following

98% on Tau2-Bench Telecom, 81% on IFBench, and a GDPval-AA Elo of 1500 in Artificial Analysis's launch evaluation, a 321-point Elo jump over Grok 4.20 on agentic real-world work.

Low hallucination rate

On AA-Omniscience it hallucinates on 25% of unknown questions at high effort and 16% at medium, among the lowest rates Artificial Analysis tracks for frontier models. It prefers to say it does not know.

1M context with a no-thinking mode

A 1M-token window takes whole codebases or document sets in one request, and reasoning effort none turns it into a plain fast model for extraction, routing, and short replies.

Best for#

Use cases#

Use Grok 4.3 as the default model behind support and operations agents that call tools many times per conversation: it follows the schema, keeps to the instructions, and answers quickly enough for a live chat. Run it at effort none for batch extraction over invoices, tickets, and forms, then switch to medium or high for the cases that need judgment. The 1M context makes it a good first reader for long contracts, log dumps, and research folders, and the low hallucination rate suits knowledge bots where a confident wrong answer costs more than a refusal. Pair it with Grok 4.5 or Grok 4.6 for the coding tasks it hands off.

Limitations#

Grok 4.3 is a mid-intelligence model. Its Artificial Analysis Intelligence Index of 38 trails Grok 4.5 at 56 and Grok 4.6 at 61, and Vals AI's SWE-bench Verified re-run puts it at 71.4%, fifty-seventh of 88 models, so keep it away from hard coding.

Fast decode does not mean fast answers at high effort: Artificial Analysis measures about 17 seconds to first token there. Use none or low effort for interactive work.

It is also somewhat verbose, using 83M output tokens across the Intelligence Index against a 71M median, which eats into the low per-token price on long tasks.

Grok 4.3 vs Grok 4.5#

Grok 4.5 is the smarter model by a wide margin: 56 against 38 on the Intelligence Index and 86.6% against 71.4% on SWE-bench Verified. Grok 4.3 wins on everything around the answer: 115.8 against 52 output tokens per second, $0.15 against $0.43 per index task, a 1M context window against 500k, and an effort none mode Grok 4.5 does not have. Route tool-heavy, high-volume, and latency-sensitive traffic to Grok 4.3; send coding and multi-hour agent runs to Grok 4.5.

When to use Grok 4.3#

Use Grok 4.3 when a request has to be fast and cheap, calls tools, or needs a million tokens of context, and the task is not frontier coding. Skip it when you need the best answer xAI can produce, which today means Grok 4.6, or when the task is a real software engineering job, where Grok 4.5 is the better fit.

API examples#

Call Grok 4.3 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/xai/grok-4.3.

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "xai/grok-4.3",
    "reasoning_effort": "none",
    "messages": [
      { "role": "user", "content": "Extract the customer name, order number, and issue from this ticket as JSON." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate, and raise reasoning_effort to medium or high for requests that need more thought.