Grok 4.5 is Opus-class and a quarter of the price
xAI shipped Grok 4.5 to the public today, and Elon is calling it an Opus-class model. Look at the numbers and the framing is right. On SWE-Bench Pro it resolves a task in about 16,000 output tokens where Opus 4.8 burns 67,000 for the same work, a 4x gap in what you actually pay for. It beats Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1, loses by a few points on DeepSWE 1.1 and SWE-Bench Pro, and tops Harvey's legal agent benchmark, which is a strange flex for a model sold as a coder. Price is $2 per million in, $6 out, running around 80 tokens a second.
The benchmark trade is a coin flip. The price isn't. At roughly a quarter of frontier cost, the question stops being which model is smartest and becomes which one you can afford to run on every task, all day, inside an agent loop that fires thousands of calls. That is the whole game for autonomous agents, and it is where Grok 4.5 is aiming.
The part nobody should skip is where the training data came from. Grok 4.5 was trained alongside Cursor, the coding editor SpaceX agreed to buy for $60 billion last month. So xAI now owns the editor, watches millions of real engineers accept and reject completions inside it, and feeds that back into the model. That is a data flywheel no benchmark captures. The model gets cheaper and better because it lives inside the tool people already use to write code.
That is the real story of the day. The coding-model race just stopped being about the leaderboard and started being about who owns the workflow the model learns from. Full details at x.ai/news/grok-4-5.
← Back to all articles
The benchmark trade is a coin flip. The price isn't. At roughly a quarter of frontier cost, the question stops being which model is smartest and becomes which one you can afford to run on every task, all day, inside an agent loop that fires thousands of calls. That is the whole game for autonomous agents, and it is where Grok 4.5 is aiming.
The part nobody should skip is where the training data came from. Grok 4.5 was trained alongside Cursor, the coding editor SpaceX agreed to buy for $60 billion last month. So xAI now owns the editor, watches millions of real engineers accept and reject completions inside it, and feeds that back into the model. That is a data flywheel no benchmark captures. The model gets cheaper and better because it lives inside the tool people already use to write code.
That is the real story of the day. The coding-model race just stopped being about the leaderboard and started being about who owns the workflow the model learns from. Full details at x.ai/news/grok-4-5.
Comments