On July 8, 2026, xAI announced Grok 4.5 — positioned as their strongest model yet for coding, agentic workflows, and knowledge work. If you build software for a living, this release is worth paying attention to, even if you are not planning to switch tools tomorrow.
What is Grok 4.5?
Grok 4.5 is xAI’s frontier model trained on large-scale datasets spanning coding, science, engineering, and math. According to the official announcement, training ran across tens of thousands of NVIDIA GB300 GPUs, with heavy investment in data filtering, deduplication, and domain-focused curation.
The model is available today in several places engineers already work:
- Grok Build (default model) — x.ai/cli
- Cursor — all plans
- xAI API — model name
grok-4.5 - Microsoft Office add-ins — Word, PowerPoint, and Excel
Note: xAI states Grok 4.5 is not yet available in the EU via API or products; EU availability is expected mid-July 2026.
Why engineers should care
Real-world coding benchmarks
xAI published results on several software engineering benchmarks — DeepSWE, Terminal Bench 2.1, SWE Bench Pro, and others. Grok 4.5 lands near the top on multiple leaderboards, especially on agentic and terminal-style tasks. Independent verification always matters, but the direction is clear: these models are being optimized for actual engineering work, not just chat.
Token efficiency
One standout claim: on SWE Bench Pro, Grok 4.5 averages about 15,954 output tokens per resolved task versus roughly 67,020 for Opus 4.8 (max) — about 4.2× fewer tokens for comparable work. If that holds in production, it directly affects latency, cost, and how long you can keep an agent loop running.
Speed and pricing
Grok 4.5 is served at fast-model speeds (~80 TPS) and priced at $2 / 1M input tokens and $6 / 1M output tokens on the xAI API. Combined with higher token efficiency, the cost-per-completed-task math looks competitive against other frontier models.
Quick API example
If you already have an xAI API key, switching models is a one-liner:
curl https://api.x.ai/v1/responses
-H "Content-Type: application/json"
-H "Authorization: Bearer $XAI_API_KEY"
-d '{
"model": "grok-4.5",
"input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
}'
The developer docs also recommend setting a prompt_cache_key (Responses API) or x-grok-conv-id header (Chat Completions) for reliable prompt caching — worth reading before running long multi-turn agent sessions.
Training alongside Cursor
Interesting detail from the launch: Grok 4.5 was trained alongside Cursor. That partnership signal matters — the model is not abstractly “good at code”; it is being shaped inside the IDE workflows many of us already use daily.
My take
As someone who spends most of the day thinking about distributed systems, manufacturing software, and developer productivity, I see Grok 4.5 less as a chatbot upgrade and more as another step toward usable coding agents — models that can plan, execute, use tools, and finish tasks with fewer tokens and less hand-holding.
I have been using Grok 4.5 heavily for a while now — including at Tesla before the public release. On real workloads: Go microservice refactors, multi-file debugging in Cursor, API design, documentation, and long agentic coding sessions. It has already earned a place in my daily workflow.
I am genuinely happy it is publicly available now. The broader engineering community finally gets access to a model many of us have been productive with behind the scenes. If you have been waiting to try a frontier coding agent, this is a great time to start — free Grok 4.5 usage is being offered for a limited time in Grok Build and Cursor.
Links
- Official announcement — Introducing Grok 4.5
- API documentation
- xAI Console (API keys)
- Grok Build CLI
What are you planning to build with Grok 4.5? Drop a comment or reach out on LinkedIn.