xAI dropped Grok 4.5 into the market on July 8, 2026, and it wasted no time making a statement. The model posted a 29.0% resolution rate on the SWE Marathon benchmark, a test designed to measure how well AI handles real-world software engineering tasks at scale.
To put that number in context, Claude Opus 4.8 came in at 26.0% on the same benchmark. Fable, another competitive entry, landed at 24.0%. Grok 4.5 cleared both by a meaningful margin.
What the benchmarks actually say
Grok 4.5 also posted an 83.3% resolution rate on Terminal-Bench 2.1, a separate evaluation focused on terminal-based development workflows.
Worth noting: independent verification of these figures is still catching up. Many results at this stage of a model launch are based on vendor evaluations, so the scores should be read as directional rather than definitive. That caveat applies equally to every model in this comparison.
The pricing angle is where things get interesting
Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens through the API.
xAI also claims roughly 2x better token efficiency compared to leading rivals. In plain terms: if a competing model takes 200,000 tokens to complete a task, Grok 4.5 allegedly handles the same task in around 100,000.
Kimi K3 enters the picture
One week after Grok 4.5’s launch, Moonshot AI announced Kimi K3 on July 16, 2026. The model features 2.8 trillion parameters and is built as an open-weight system, meaning developers can access and modify the underlying model weights rather than working solely through an API.
Kimi K3 is designed for agentic coding and reasoning tasks, positioning it as a competitor to top closed models from OpenAI, Anthropic, and now xAI. Kimi K3 is described as slower but inexpensive.
What this means for compute and infrastructure markets
Neither Grok 4.5 nor Kimi K3 includes blockchain integration or token utility, so the direct impact on crypto markets is limited. There is no native token play here, no on-chain governance angle, and no DeFi hook.
Kimi K3’s open-weight architecture adds another dimension. Open models tend to drive broader ecosystem experimentation, which can accelerate overall compute demand as more developers fine-tune and deploy custom versions. The 2.8-trillion-parameter scale also suggests significant inference costs even at competitive pricing.