Meta and Reflection Debut Massive Open Models Alongside Frontier API Shifts
Meta's Muse Spark Open and Reflection Beam 1.0 expand US open-weights, while Gemini, Anthropic, and OpenAI roll out aggressive pricing and architectural updates.
The landscape for both open-weight architectures and hosted frontier models has shifted rapidly with the release of several massive mixture-of-experts (MoE) systems and aggressive price reductions across hosted APIs. Leading the US open-weights releases are Meta's Muse Spark Open and the newly unveiled Reflection Beam 1.0, both offering multi-hundred-billion parameter scale with large context windows, while proprietary providers adjust latency, pricing, and agent capabilities.
Open-weights push scale with Muse Spark and Reflection Beam
Meta introduced Muse Spark Open, an 816-billion total parameter mixture-of-experts model with 42 billion active parameters per token. The model incorporates 62 layers, grouped-query attention with 16 key-value heads, and routes between 8 of 256 available experts. Built with a 1.8-billion parameter perception encoder, it handles text, image, video, and audio inputs with a 1-million-token context window. While full BF16 weights weigh roughly 1.6TB, Meta's NVFP4 checkpoint quantizes the experts alone, cutting the size down to 530GB—allowing it to run on hardware configurations such as four Blackwell Ultra GPUs, eight H200s, or two 512GB M5 Ultra Mac Studios.
On Artificial Analysis's Intelligence Index, Muse Spark Open posted a score of 39, ranking highest among US open-weights systems and ahead of NVIDIA’s Nemotron 3.5 Super VL at 26. Hosted inference platforms including Together, Fireworks, and Baseten have priced the model at $1.25 per million input tokens and $4.25 per million output tokens, with $0.15 for cached inputs. However, performance remains varied across benchmarks: while it tops Code Arena WebDev among US open models with 1511 points and leads on AA-Omniscience, it ranked last among the top nine open models on Artificial Analysis’s four agentic evaluations.
Concurrently, Reflection released Reflection Beam 1.0 under an Apache 2.0 license. Beam 1.0 is a 501-billion total, 23-billion active parameter MoE with a 1-million-token context window. It scored 30 on the Artificial Analysis index, making it the second-highest ranked US open model. The release includes BF16, FP8, and NVFP4 checkpoints, the open Mirror CLI training harness, and day-zero serving in vLLM.
Frontier updates and the emergence of decision models
Proprietary providers have responded with expanded capabilities and sharper pricing. Google expanded availability of Gemini 4 Argon to Pro subscribers and AI Studio Build, increasing the output token limit from 64K to 1 million tokens, shared across the context window, with introductory pricing of $2 per million input and $10 per million output tokens running through November 30. In agentic testing, four Gemini 4 Argon agents completed Minecraft end-to-end without human intervention, beating the Ender Dragon in just over 27 hours across 2,783 generated programs.
Anthropic adjusted its lightweight tier with Haiku 5.5, pricing it at $0.10 per million input and $0.50 per million output tokens for prompts up to 100K tokens—a reduction of roughly 90% compared to Haiku 4.5. On Terminal-Bench 4.0, Haiku 5.5 achieved a 39.2% score at maximum effort, compared to 16.4% for GPT-6 Luna. OpenAI also rolled out updates to GPT-6 Astra Pro, incorporating parallel agent execution across discrete tasks within ChatGPT.
Alongside general models, dedicated one-pass classifiers known as decision models are gaining adoption. Liquid's d1-ultra reached an index score of 65 on Artificial Analysis's Decision Index at a latency of 92 milliseconds and a cost of $0.042 per million input tokens with free output. Perplexity similarly published pplx-decider v1.2, releasing open-weight 27B and 4B Flash models aimed at low-latency routing and reranking tasks.
What it means for developers
The dual expansion of high-parameter open models and highly specialized small models fundamentally changes deployment economics. For self-hosted workloads, running top-tier open models like Muse Spark Open still demands enterprise-grade hardware, requiring at least 530GB of VRAM even with NVFP4 expert quantization. Furthermore, Meta's custom license mandates a "Built with Muse Spark Open" attribution, alongside commercial licensing agreements for model-as-a-service entities exceeding $20M in annual revenue.
At the same time, the steep cost cuts on models like Haiku 5.5 and the emergence of sub-cent decision models such as pplx-decider allow teams to build multi-tier architectures. In these pipelines, fast classifiers handle initial triage and tool delegation before routing complex inputs to larger models. For engineering teams evaluating these tradeoffs across vendors, developers can try top AI models cheaply through one API at https://apixoai.online.
As open architectures match proprietary systems in sheer parameter count, real-world utility increasingly hinges on specific operational profiles. Teams building coding and text analysis workflows can capitalize on expanded 1M context windows, but lower scores in agentic benchmarks for models like Muse Spark Open suggest that agent-heavy systems still demand rigorous, workload-specific evaluation.
Source: [AINews] High Taste Testers Ep. 1 — Latent Space. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
