A 1M-context workhorse with reasoning you can switch off
Positioned below the Grok 4.5 flagship, trading peak reasoning for a larger window and lower cost.
Overview
Grok 4.3 is xAI's long-context workhorse in the Grok 4 line, released on April 17, 2026. It pairs a very large 1M-token window with an adjustable reasoning control that includes a none setting, so the same model can run as a fast, non-reasoning responder or as a deliberate step-by-step reasoner depending on the request.
Positioned below the newer Grok 4.5 flagship, 4.3 trades a little peak reasoning for a much larger context window and lower cost, which makes it a practical default for whole-repository work, long documents, and high-volume tool loops. Like the rest of the family, it is reachable through an OpenAI-compatible API with native function calling, vision, and Live Search.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 1,000,000 tokens |
| Max output | 1,000,000 tokens |
| Input modalities | Text, image |
| Output modalities | Text |
| Reasoning | Adjustable effort: none / low / medium / high |
| Tools | function calling, structured outputs, streaming, vision, Live Search |
Inputs above 200K tokens enter a higher long-context tier (2x pricing). Cached input is billed well below fresh input. See live pricing in the model catalog.
Long context and reasoning modes
One knob covers cheap throughput and careful reasoning
Use none for latency-sensitive volume; turn up to high when a task needs verified multi-step work.
The 1M-token window is the headline: entire codebases, long contract sets, research corpora, and multi-file histories fit in a single request, and the 1M max-output ceiling means large generated artifacts are not truncated by a separate output cap.
The none effort setting is what makes 4.3 versatile. Turn reasoning off for latency-sensitive, high-volume calls where a direct answer is enough; turn it up to high when a task needs verified multi-step work. That one knob lets a single deployment cover both cheap throughput and careful reasoning.
Benchmarks
Reasoning-first: math, science, and agentic tool use
xAI's published benchmarks; competition math is near-perfect, and τ²-Bench reflects stable multi-step tool use.
Grok 4.3 is a reasoning-first model, and xAI's published numbers lean that way: 95 on AIME (near-perfect competition math), 90.1% on GPQA Diamond (graduate-level science), and 97.7% on τ²-Bench, which measures stable multi-step tool use across agent loops. On engineering it posts 75 on SWE-Bench and 47.3% on SciCode. Where Grok 4.5's launch leads with agentic coding and token efficiency, 4.3's strength is consistent reasoning and tool use — delivered while keeping a million tokens of context in view.
When to use it
- Whole-repository work: large codebases, logs, and design docs in one window.
- Long-document analysis: contracts, research packets, and multi-file corpora.
- Latency-sensitive volume: turn reasoning to none for fast, cheap direct answers.
- Cost-aware agents: long tool loops where context size matters more than peak reasoning.
CrossModel exposes Grok 4.3 through an OpenAI-compatible /v1/chat/completions API. Current pricing is available in the model catalog.