Beam's 501B-parameter MoE model achieves 3-4x inference efficiency gains

Anthropic's subscription plans deliver 5x+ more API value than OpenAI; OpenAI halves $200 plan limits

SemiAnalysis conducted systematic limit testing across Anthropic and OpenAI subscription tiers and found Claude subscriptions offer dramatically superior API-equivalent value—particularly for mid-tier models like Opus 5.5 versus GPT-6.1 Sol. Despite subscriptions representing only ~10% of overall revenue, they consume over 40% of inference compute and lower blended revenue per megawatt by ~$36M for Anthropic. OpenAI responded by halving the $200 plan's API-equivalent value (with grandfathering until October 29) while introducing a $500 tier that offers only 21% more capacity and actually decreased value for Sol-class models. The moves signal OpenAI is reducing subscription subsidies to improve margins, contrasting with Anthropic's more gradual approach of tiering value by model capability. SemiAnalysis notes that subscription limits can change silently at any time—they even detected an A/B test where one account had ~20% lower limits than others—making real-time tracking essential for accurate financial modeling of AI labs.

Reflection launches Beam: 501B-parameter open-weight MoE model with 3-4x inference efficiency versus GLM-5.2

Reflection unveiled Beam, a 501B-parameter sparse mixture-of-experts model with 23B active parameters, trained on 23.8 trillion tokens and reinforced with high-compute RL across ~1 million tasks on ~10,500 NVIDIA GB300 GPUs over four weeks. The model achieves reasoning performance comparable to GLM-5.2 while using 3-4x less inference compute, with 80.9 on SWE-bench Verified and strong coding and agentic capabilities. Beam was trained with a controllable length penalty that rewards efficient reasoning, and users can adjust a reasoning effort parameter to trade off response length against performance. Full weights, technical report, and developer artifacts will release under Apache 2.0 later this month. This represents a concrete US-trained open alternative, though independent analysis places it around GLM-5.2 level and below DeepSeek V4 Flash on some benchmarks, signaling continued competitive intensity in the open-weight model space.

OpenAI DevDay 2026: Decisions API gains vision support; GPT-6.1 Sol priced at $2/$10 per million tokens

OpenAI's DevDay announcements included a vision-enabled Decisions API for structured classification tasks like thumbnail selection across video frames, and GPT-6.1 Sol pricing at $2 input / $10 output per million tokens—a significant cost reduction versus Astra ($10/$50). Sol is positioned for cost-conscious teams building on OpenAI's platform primitives including the Agents API, Codex, and plugin monetization. Developer John Lindquist demonstrated eight concrete use cases showing Jev (OpenAI's decision model) as a fast, inexpensive classification and routing engine rather than a generative chatbot—processing decisions in milliseconds at near-zero cost (73 cents across 23 dev runs), with chess analysis running 10x faster and 4x cheaper than comparable LLMs. The framing matters: developers treating Jev as an if/else statement replacement rather than prompt engineering target unlock previously impractical workflows. Kath Korevec, OpenAI's Product Lead, demonstrated ChatGPT Sites with Plugin Insights for personalization, showing how fast models enable temporary but valuable personal software.

OpenAI faces capacity constraints: halves subscription limits, speeds up GPT-6 models by ~50%

OpenAI paused new $200 subscription sign-ups and halved usage limits across subscription plans, then increased default speed for GPT-6 Astra and GPT-6.1 Sol from ~30 to ~50 tokens per second across all surfaces. Codex lead Tibo pledged meaningful improvements or resets daily for 28 days. This signals infrastructure strain and competitive pressure from Anthropic, which SemiAnalysis reports delivers 5x+ more API-equivalent value per dollar on comparable subscription tiers.

Introducing Beam: Reflection's 501B-Parameter Open-Weight MoE Model with High-Compute RL

Reflection AI published a technical report on Beam, a 501B-parameter sparse MoE open-weight model designed for coding and agentic workloads. Main takeaways:

• Beam achieved competitive performance with GLM-5.2 and approaches Qwen 3.8-Max on coding/agentic tasks while requiring 3-4x less inference compute, demonstrating efficiency gains in the open-weight frontier
• Reflection conducted one of the largest-scale RL runs to date: 10.5K NVIDIA GB300 GPUs over 4 weeks generating 100M+ rollouts at 256K context, with 1.3B sandboxes and one million curated environments
• Model uses controllable reasoning effort parameter allowing users to trade token efficiency for capability, with RL improving solution quality while reducing completion lengths early in training
• Beam demonstrates cross-domain generalization: trained on reasoning/coding/terminal tasks, it organically learned web browsing and tool use without explicit training, suggesting broader agentic capability transfer
• Infrastructure innovations enabled frontier-scale RL: asynchronous execution with policy staleness handling, 12-second median weight distribution across fleet, 71 inference incidents recovered without job termination, 99.99% trainer batch fullness