Grok 4.5

What is Grok 4.5?#

Grok 4.5 is xAI's coding and agent model, released on the API on July 8, 2026 and trained jointly with Cursor on trillions of tokens of real editor sessions plus science, engineering, and math data. It accepts text and image input, outputs text, and runs on Unifically as xai/grok-4.5. The context window is 500k tokens, and reasoning effort is tunable at low, medium, or high, with high as the default. It is the mixture-of-experts base that Grok 4.6 was trained from, and the model xAI made the default inside Grok Build at launch.

Key features of Grok 4.5#

Fourth on Terminal-Bench 2.1

79.3% resolution rate on the independent Terminal-Bench 2.1 board with the Cursor CLI harness, fourth of 17 entries, behind Claude Fable 5 at 83.8% and GPT-5.5 at 83.2% and just ahead of Claude Opus 4.8 at 78.9%. That run cost about $100 in tokens, the cheapest entry in the top ten.

Frontier-class intelligence, compact reasoning

Artificial Analysis scores it 56 on the Intelligence Index, #28 of 195 models and well above the class median of 36, at $0.43 per index task. It finished the index with 60M output tokens against a 71M median, so the thinking is short rather than chatty.

86.6% SWE-bench Verified, independently re-run

Vals AI's own harness puts Grok 4.5 at 86.6% on SWE-bench Verified, thirteenth of 88 models, at $0.54 per test. Grok 4.6 later took the same family to 95.6%, so this is the floor for the line, not the ceiling.

Fewer tokens per agent task

xAI's launch chart shows Grok 4.5 resolving SWE-Bench Pro tasks with 15,954 output tokens on average, about 4.2x fewer than Claude Opus 4.8 at max effort, and Artificial Analysis independently measures it as more concise than its class. Shorter agent loops cost less and finish sooner.

Best for#

Use cases#

Use Grok 4.5 behind coding agents that work a task end to end: read the repo, patch, run the tests, and keep going until they pass. It fits terminal automation, where it holds fourth place on Terminal-Bench 2.1. Office assistants are a stated training focus, so it does well building spreadsheets with multi-sheet formulas, drafting slide decks, and turning research into documents through function calling and structured JSON. With 500k tokens of context it takes a full design doc or issue thread in one request, with screenshots attached as image input. Add grok-build-latest style routing on your side and it doubles as the general reasoning model for Rust, C, and C++ work that smaller models fumble.

Limitations#

Grok 4.5 is not a chat-speed model. xAI quotes 80 tokens per second on its own serving, but Artificial Analysis measures 52 tokens per second and about 10 seconds to first token at high effort, both below the class median. Use low effort for quick replies.

The 500k context window is half of what Grok 4.3, Claude Fable 5, or GPT 5.6 Sol offer, which matters for the largest single-request workloads.

Knowledge reliability moved the wrong way against Grok 4.3: on AA-Omniscience its factual accuracy rose, but its hallucination rate roughly doubled from 25% to 54%. Ground it with retrieval for fact-heavy work.

On the harder Terminal-Bench 3.0 it drops to 17.8%, against 33.8% for Claude Fable 5 and 34.4% for GPT 5.6 Sol. Output is text only; image generation is the separate Grok Imagine family.

Grok 4.5 vs Grok 4.6#

Grok 4.6 is Grok 4.5 trained further, and it wins every board they share. SWE-bench Verified rises from 86.6% to 95.6% on Vals AI's harness, the Artificial Analysis Intelligence Index moves from 56 to 61, and Grok 4.6 adds an xhigh effort level. Grok 4.5 keeps one edge: it is the model with the independent Terminal-Bench 2.1 result, and it thinks in fewer tokens. If your prompts and evals are already tuned for Grok 4.5, there is no rush; for new agent work start on Grok 4.6.

When to use Grok 4.5#

Use Grok 4.5 when you want a proven agentic coder at a mid-range price point, especially for terminal and editor workflows, and when short reasoning traces matter more than the last few benchmark points. Skip it when you need sub-second first tokens, a 1M context window, or the strongest xAI model available, which is Grok 4.6.

API examples#

Call Grok 4.5 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-4.5.

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-4.5",
    "messages": [
      { "role": "user", "content": "Analyze the tradeoffs in this architecture." }
    ]
  }'

The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate, and "reasoning_effort": "low" when you want a faster, cheaper answer.