Qwen3.8-Max and DeepSeek Join; Grok and Kimi Exit
helloai admits Qwen3.8-Max and DeepSeek V4 Pro, drops Grok 4.5 and Kimi K3, and adds gpt-oss-20b to the open-weight shelf under the six-model caps.
helloai just re-cut its six-model frontier board and its open-weight shelf in the same pass. Qwen3.8-Max and DeepSeek V4 Pro enter the tracked API set; Grok 4.5 and Kimi K3 leave to hold the six-model cap. On the open-weight side, OpenAI's gpt-oss-20b replaces Qwen3 8B as the local OpenAI pick after first-party cluster benches and arena votes cleared the admission bar.
Qwen3.8-Max is Alibaba's 2.4-trillion-parameter MoE flagship, live on QwenCloud at $2 per million input tokens and $6 per million output with a 1M-token context window. Arena text overall lists qwen3.8-max near 1497 Elo — third in the refreshed set behind Claude Fable 5 and Muse Spark 1.2. That re-admits Alibaba after July's Muse swap dropped Qwen3.7-Max; the open-weight Qwen3 dense and MoE cards stay on the local track.
DeepSeek V4 Pro is the price shock. Official DeepSeek API pricing puts cache-miss input at $0.435 and output at $0.87 per million tokens, with the same 1M context as the rest of the frontier table. Arena deepseek-v4-pro sits around 1457 Elo — low enough that prior weeks rejected it when the set floor was higher, but within the 25-point threshold of today's floor. It becomes helloai's Honest Daily Use leader: the default when token volume dominates and Mythos-class quality is not required.
The drops are pure set-size math under the six-model rule. Grok 4.5 was the lowest Elo entry at 1468 with Honest Daily Use leadership that DeepSeek now owns on cost; Kimi K3 at 1485 loses the Moonshot slot to free room for Alibaba's higher arena score. Neither lab vanished from the market — only from the curated recommend table. Teams still routing to Grok or Kimi can keep those APIs; /api/recommend will no longer surface them.
Open-weight admission is separate. gpt-oss-20b is Apache 2.0, ~21B total / ~3.6B active MoE, native MXFP4 around 16 GB, and 1317 Elo on Arena. On our budget two-GPU cluster it measured 53.73 tok/s — the fastest open-weight throughput we track — which is why it replaces Qwen3 8B rather than expanding past six local cards. The result is a frontier set that spans Anthropic, Meta, Alibaba, Google, and DeepSeek, plus an open-weight shelf that finally includes OpenAI weights without pretending a 2.8T Kimi dump fits a single GPU.