A fast frontier model for everyday coding, search and interactive edits.
Balanced speed and quality — 3× cheaper, 200K context.
Sunlight 2 is Meteor's workhorse — optimized for speed without sacrificing frontier quality. Trained on 12T tokens of code and text, with a 200K context window and aggressive distillation, it hits 92.1% on HumanEval and 62.8% on SWE-bench Verified while being 3× cheaper than our Pro model. Ideal for interactive use: Tab autocomplete, Cmd-K edits, and background agents that need to be fast and precise.
Built on the same architecture as Sunlight 2 Pro but distilled and fine-tuned for latency. Uses grouped-query attention, 8k native generation and tool-use pre-training. Trained with RL on real Meteor trajectories: 400k agent episodes with test-driven rewards. Fully evaluated for safety with no regression on our harm benchmarks.
Sunlight 2 vs best competitor at similar price. Higher is better.
| Benchmark | Sunlight 2 | GPT-4o | Claude 3.5 Sonnet |
|---|---|---|---|
| SWE-bench Verified | 62.8% | 48.1% | 54.2% |
| HumanEval (pass@1) | 92.1% | 90.2% | 89.7% |
| MMLU | 86.7% | 86.1% | 85.4% |
| LiveCodeBench | 61.4% | 53.2% | 55.9% |
| Cost / 1M tokens | $1.20 | $3.50 | $3.00 |