xAI's flagship reasoning model, with native tools and real-time search
Reasoning effort is adjustable across low / medium / high, trading latency and cost against depth per request.
Overview
Grok 4.5 is xAI's flagship reasoning model, released on July 8, 2026. It continues the Grok 4 line's focus on native reasoning and real-time information: the model thinks before answering, calls tools on its own, and can pull fresh context from the web and 𝕏 through xAI's Live Search rather than relying only on pretraining.
Trained in xAI's Memphis data centers, Grok 4.5 targets hard, tool-driven work — agentic coding, research, and automation — while staying reachable through an OpenAI-compatible API. Reasoning effort is adjustable across low / medium / high (default high), so latency and cost can be traded against depth per request. Its knowledge cutoff is February 1, 2026.
Key capabilities
| Dimension | Detail |
|---|---|
| Context window | 500,000 tokens |
| Max output | 500,000 tokens |
| Input modalities | Text, image |
| Output modalities | Text |
| Reasoning | Adjustable effort: low / medium / high (default high) |
| Tools | function calling, structured outputs, streaming, vision, web search, 𝕏 search, code execution |
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.
Agentic coding
xAI's launch post leads with agentic coding (vs Opus 4.8)
DeepSWE 1.0 runs inside each provider’s own harness (less neutral); DeepSWE 1.1 uses the shared mini-swe-agent harness.
xAI's launch announcement leads with agentic coding rather than a reasoning-splits table. Against Claude Opus 4.8, Grok 4.5 posts 83.3% on Terminal-Bench 2.1 (Opus 4.8: 78.9%) and 64.7% on SWE-Bench Pro resolve rate (Opus 4.8: 69.2%). On DeepSWE it reads differently by harness: 62.0% on DeepSWE 1.0 run inside each provider's own harness, but 53.0% on DeepSWE 1.1 under the neutral mini-swe-agent harness (Opus 4.8: 59%). The provider-vs-neutral gap is worth keeping in mind when reading any single coding number.
Token efficiency
Same SWE-Bench Pro results, with far fewer output tokens
xAI's own data: lower token consumption directly lowers per-task agentic coding cost.
The efficiency story is where Grok 4.5's cost case sits. On SWE-Bench Pro, xAI reports an average of 15,954 output tokens per task versus 67,020 for Opus 4.8 at max effort — roughly 4.2x fewer output tokens for comparable resolve rates. For high-volume agentic coding, fewer output tokens per task translates directly into lower cost per task.
Real-time search and tools
Beyond coding, Grok 4.5's differentiator is acting on fresh information. Through Live Search it retrieves current context from the web and 𝕏; with native function calling, structured outputs, and code execution it drives multi-step tool loops — search, read, run, and revise. Vision input lets it fold screenshots, diagrams, and charts into the same workflow, and the 500K-token window keeps large repositories and long research packets in a single context.
When to use it
- Agentic coding: tool loops that read the repo, run commands, and iterate on failures.
- Cost-sensitive engineering: comparable resolve rates at markedly fewer output tokens per task.
- Real-time research: questions where fresh web and 𝕏 context beats a static snapshot.
- Long-context work: large codebases, logs, and research documents in one window.
CrossModel exposes Grok 4.5 through an OpenAI-compatible /v1/chat/completions API. Current pricing is available in the model catalog.