The AI model market has segmented along cost, speed, and accuracy dimensions, making older open-weight models competitive long after their release.

OpenRouter data shows that GPT-OSS 120b, a year-old open-source model, still serves 36% of the daily tokens of Anthropic’s newer Claude Opus 4.8. GLM 5.

2 now serves 495 billion tokens per day, outpacing frontier models. Anthropic shipped Opus 5 to contest the space that Moonshot’s Kimi 3 targets, while Poolside launched Laguna S 2.

1 for the US mid-market. The key technical driver is mixture-of-experts (MoE) architecture. The 118-billion-parameter Laguna S 2.

1 activates only 8 billion parameters per token, enabling it to decode at the same speed as a dense 26B model on Apple M5 Max hardware.

Replacing Gemma 4 26b with Laguna in a local agent stack reduced tool-call failure rates from 29.4% to 20.1%, a 7 percentage point gain.

Segmentation shows healthy competition: frontier models still handle the hardest tokens, but lower tiers now achieve higher quality ceilings.