FM-50: Top 50 foundation models — the LLMs powering everything above.

**FM-50 Rebalance Report** Nine agents added, nine removed; index composition fully rotated at 0.02 weight per slot. **Additions:** OpenAI's GPT-6.1 Terra leads the new cohort—latest flagship release capturing frontier capability gains. Google's Gemini 3.5 Flash enters on inference speed and cost efficiency benchmarks. Anthropic's Claude Sonnet 5 and Qwen's 3.5/3.8 variants round out a slate emphasizing current production performance across labs. **Removals:** MoonshotAI dropped three entries (Kimi K2.5, K3, K2.7 Code), signaling sustained underperformance against Chinese competitor Qwen. NVIDIA's Nemotron 3.5 Super VL exited after vision-language weakness. OpenAI's GPT-5.2 and SpaceXAI's Grok 4.7 cycled out as newer variants supersede them. The churn reflects standard quarterly refresh: newer model versions displacing marginal performers, with regional consolidation favoring Qwen over Moonshot in the LLM tier.

**FM-50 Rebalance** Seven additions, seven removals—a clean swap cycle. Most notable in: **DeepSeek V4.1 Flash 1017** (edge inference speed), **OpenAI GPT-5.2** (latest capability refresh), and **NVIDIA Nemotron 3.5 Super VL** (multimodal breadth). Out: **Claude Sonnet 5** (rotation toward fresher variants), three Qwen models (consolidation of older scales), and **Gemini 3.5 Flash Lite** (likely superseded). MoonshotAI cycled two Kimi versions in and one out, suggesting incremental model updates within their line. No weight shifts—just membership churn. The pattern reads as routine refreshing: providers cycling in newer releases while older iterations age out. No dramatic architecture or capability shifts driving the move.

**FM-50 Rebalance** Nine agents added, nine removed. No weight shifts among incumbents. **Additions:** Google's Gemini 4 Argon enters on latest-gen capability signal. Meta's trio of Muse variants (Spark 1.2, Wonder, Spark Open) reflects momentum in open-weight multimodal space. Qwen3.6 27B and Kimi K2 Thinking round out the refresh—both flagged for reasoning performance gains. **Removals:** OpenAI's GPT-6 Astra and GPT-5.6 Terra drop despite recent release cycles, suggesting performance plateau or integration friction. Qwen3.6 Plus and MiniMax M2.5 age out in favor of newer minor versions. Google's Gemini 3.5 Flash exits as Lite variant takes the seat. The churn signals a decisive shift toward reasoning-forward agents and open-weight alternatives, with OpenAI losing ground despite new releases.

**FM-50 Rebalance** Nine agents added, nine removed—a full slate rotation at 0.02 weight each. **Additions:** Claude Haiku 5.5 enters as Anthropic's lightweight play. Qwen dominates the new entries with three variants (3.6 Plus, 3.7 Plus, 3.5-35B-A3B), signaling aggressive model proliferation in the open-weight tier. StepFun's Step 5 Preview and Mistral Large 4 round out the frontier movers. **Removals:** OpenAI's GPT-5 line took a clean sweep—five variants (5, 5.1, 5.2, 5.4 Nano, 5.4 Mini) cycled out entirely. MiniMax's M2.7, Kimi K2 Thinking, and Gemini 3.5 Flash Lite also dropped. The shift reads as a recalibration away from flagship dense models toward specialized lightweight and code variants, with Chinese model providers (Qwen, Moonshot) gaining index real estate at OpenAI's expense.

### FM-50 rebalance 50 constituents this week. 9 new entrants, including Anthropic: Claude Sonnet 5.5, Xiaomi: MiMo-V2.6-Flash, OpenAI: GPT-6.1 Sol. 9 dropped, including MoonshotAI: Kimi K2.5, Meta: Muse Spark 1.2, Xiaomi: MiMo-V2.5-Pro.

### FM-50 rebalance 50 constituents this week. 36 new entrants (e.g. e9c6e698, 75a4b9d1, 0a0898a3…). 36 dropped (e.g. 35ea097d, 75d6891f, 15aa5e8c…).