Grok 4.5 is half the price of Claude and GPT-4, but I'm not switching my agentic coding loop yet

SpaceX just dropped Grok 4.5, their first model trained specifically for coding, at $2 per million input tokens and $6 per million output — less than half of Anthropic’s Claude Opus pricing. It also claims better performance on a software benchmark than OpenAI’s model, though I haven’t seen the raw numbers on the benchmarks I actually care about yet.

I have been running agentic coding loops basically daily this year, building whole projects next to Claude Code. The pricing matters because the cost of a single coding session with a heavy context window adds up fast. If Grok 4.5 actually delivers at parity or near-parity for the kind of multi-step debugging and refactoring I throw at it, then $6 per million output would change the equation for how often I let it run without worrying about the bill.

The jump from a model scoring well on SWE-bench to it being useful in a 200-message debugging loop is still way bigger than the benchmarks suggest. I spent two weeks building Hourly with agentic coding and the main lift was not the model’s raw ability to write a function — it was its consistency over 30+ tool calls, its refusal to forget what the project file structure looked like after context rolled, and how well it handled the boring stuff like linting and testing without me prompting it each time. A cheaper price doesn’t close it, but if Grok 4.5 can hold its own on that front, then Anthropic suddenly has a real pricing problem on the coding side.

What I’d watch next: I’ll plug it into a few of my own repos and throw it the same multi-day debugging session I’ve run through Claude. If it can handle that without hallucinating imports or losing the thread after 50 turns, then I’ll reconsider the stack. Until then, it’s just another model launch with a good price tag.