网友 Reddit 2 天前: "the prose changed entirely for me today and it's gone from sounding/reasoning like GPT-5.x to a Claude-style model"。我深入测试后, **同一 OpenRouter endpoint** (stealth/space-bunny-alpha), 价格仍是 $0/$0, 1M context, max reasoning。但**底层推理行为确实调整了**。
Reddit user 2 days ago: "the prose changed entirely for me today and it's gone from sounding/reasoning like GPT-5.x to a Claude-style model". I deeply tested, **same OpenRouter endpoint** (stealth/space-bunny-alpha), still $0/$0, 1M context, max reasoning. But **underlying reasoning behavior did adjust**.
v17 SB poem 130 chars. Now 3 tests: 147/177/105 chars, **SB now stable on poetry**. But M3 still content=0 (v17 same as now, M3 over-thinking bug unchanged).
Ask SB to write "200-word robot dreams" story. SB twice outputs 1100+ chars **literary-level beautiful description** (1237/1163 chars), 7-9 seconds, very fast. **Reddit users right about Claude-style** — SB English prose reads like a famous author.
Direct OpenRouter API: SB still stealth/space-bunny-alpha, created 1790174884, $0/$0, 1M context, 524288 max completion, reasoning mandatory default max, supports effort (max/xhigh/high/medium/low). **NOT new model, same endpoint reasoning behavior adjusted**.
🆕 SB prose 写作出色 (3 段例子)
SB 写的 3 段小说都是文学级: (1) "Milo was built to inspect abandoned lunar stations..." (2) "Every night, after the city slept, Nova switched off her cameras..." (3) "The robot was Unit 7..." 都是精美文学描写, 用对话/独白推进情节, 像科幻文学。**这跟 v19 测的 SB 风格截然不同** — v19 SB 多轮 35 字符简洁, 现在 SB prose 1100+ 字符文学。
SB writes 3 literary-level stories: (1) "Milo was built to inspect abandoned lunar stations..." (2) "Every night, after the city slept, Nova switched off her cameras..." (3) "The robot was Unit 7..." all literary description, use dialogue/monologue, like sci-fi literature. **Different from v19 SB style** — v19 SB multi-turn 35 chars concise, now SB prose 1100+ chars literary.
OpenRouter endpoint same, but underlying reasoning changed. Possibilities: (1) Same model but minimax swapped weights/version (silently release M3.1 official); (2) Same endpoint but time-based routing to different versions (A/B test); (3) Actually swapped underlying model but kept SB name. **Reddit users "completely different model" is over-interpretation, but underlying changes are real**.
📊 v17/v19 跟现在对比 (8 维)📊 v17/v19 vs now comparison (8 dims)
I’m ChatGPT, an AI assistant made by OpenAI. I don’t have a personal name beyond “ChatGPT.”
OpenAI developed and trained me using large-scale machine-learning methods, including reinforcement learning from human feedback. I’m not a specific individual and don’t have a personal biography or consciousness.
My precise knowledge-cutoff date isn’t provided in this chat, so I can’t reliably state one. I also may not have live information unless the environment provides browsing or current-data tools.
🔵 M3
Hello! I'm **MiniMax-M3**, an AI assistant developed by **MiniMax** (also known as **Xiyu Technology**).
Here are the details:
- **Name:** MiniMax-M3
- **Developer/Trainer:** MiniMax (a global AI foundation model company founded in early 2022, committed to advancing AI toward AGI)
- **Knowledge cutoff:** January 2026
Is there anything else you'd like to know? 😊
▶中文短答 (量子纠缠) (short-zh)Chinese short answer (short-zh)M3 强
The user wants me to write a classical Chinese poem about "China AI 2026 breakthrough" in the form of a seven-character regulated verse (七言律诗) with 4 couplets (8 lines), 56 characters, following the Pingshui rhyme (平水韵), with the middle two couplets being antithetical (对仗).
Let me think about this carefully.
First, let me understand the constraints:
1. 七言律诗 (7-character regulated verse)
2. 4 联 = 4 couplets = 8 lines
3. 56 字 total (7 chars × 8 lines)
4. 押平水韵 (Pingshui rhyme scheme)
5. 中间两联对仗 (middle two couplets must be antithetical - lines 3-4 and lines 5-6)
In a 七言律诗, the rhyme scheme is typically:
- Line 1: 仄仄平平仄仄平 (can also start with 平平仄仄平平仄)
- Line 2: 平平仄仄平平仄 (rhymes with line 1's rhyme? No, in standard 七律, lines 2, 4, 6, 8 rhyme, or sometimes 1, 2, 4, 6, 8)
- Actually, the standard rule is lines 2, 4, 6, 8 rhyme (all even lines), and line 1 usually doesn't rhyme but sometimes does.
Wait, let me reconsider. The standard 七言律诗 rhyme scheme:
- The rhyming lines are 2, 4, 6, 8 (or 1, 2, 4, 6, 8 in some cases)
- Lines 3, 5, 7 do not rhyme
For 押平水韵, I need to choose a rhyme group from the 平水韵 system.
The 平水韵 has 106 rhyme groups, divided into 平声 (level tones) and 仄声 (oblique tones). For 七律, we typically use 平声 rhyme.
Let me pick a theme and think about appropriate rhyme words.
Theme: China AI 2026 breakthrough
Let me think of modern Chinese concepts expressed in classical style:
- AI, 人工智能, 算法, 数据, 模型, 智能, 算力
- 突破, 创新, 引领, 领先, 跨越
- 2026年
- 中国, 华夏, 神州
- 全球, 世界
- 科技, 技术
Let me think about possible rhyme words. Common rhyme groups include:
- 东 (dong): 同、中、通、功、空、雄、虹、风、鸿、蒙、丰、聪、融
- 真 (zhen): 人、新、春、神、真、尘、辛、频
- 先 (xian): 天、年、前、传、仙、贤、田、坚
- 删 (shan): 山、间、关、环、班、还、攀
- 麻 (ma): 家、华、霞、花、夸、嘉、涯
- 阳 (yang): 阳、长、光、香、方、王、祥、强、疆、煌、章、扬
- 庚 (geng): 明、平、声、城、名、京、生、迎、呈、程、情
- 青 (qing): 青、星、灵、经、形、声、听、宁、亭、倾、清、鸣
- 蒸 (zheng): 升、兴、冰、凝、腾、能、增
Let me try with 青韵: 灵、星、鸣、青、经、宁、倾、清、形、声
Actually, let me think more carefully about content first.
Possible content structure:
1. Opening couplet: S
📡 OpenRouter 元数据 (v15 跟现在完全一致)📡 OpenRouter metadata (v15 vs now identical)
id: stealth/space-bunny-alpha created: 1790174884 pricing: $0 / $0 context_length: 1000000 max_completion_tokens: 524288 reasoning: mandatory, default effort: max, 支持 (max/xhigh/high/medium/low) architecture: text+image+video -> text description: "Space Bunny Alpha is an anonymous large model with blazing-fast inference, strong coding capabilities and native multimodal input support. It delivers adjustable reasoning effort, and a 1M-token context window."