Baseten vs SemiAnalysis

Goodfire Launches Internal AI Agent Monitors

Goodfire, a startup specializing in AI interpretability, launched on Thursday a new type of AI agent monitor that inspects a model's internal signals rather than reading its output, aiming to detect rogue behaviors more efficiently and at a fraction of the cost. The monitors are available to customers of Baseten, an AI model hosting platform. Baseten's Base Labs had previously announced a safety partnership with Goodfire and Hugging Face. Goodfire's approach uses small probes that scan a model's internal activations at each step, triggering a closer AI review only when flagged. In tests on the Kimi K3 model, Goodfire's monitors caught 94% of malicious hacking sessions and cost about $51 for 1,500 sessions, compared to $233 for a cheaper model and $10,000 for a top-tier one. The company positions the solution for open models, which can be stripped of safeguards, and sees it as critical for inference-time guardrails.

• Goodfire launched internal activation monitors for AI agents on October 8, 2026.
• Monitors are available to Baseten customers.
• In tests on Kimi K3, monitors caught 94% of malicious hacking sessions.

Announcing our partnership with OpenAI

Baseten announces a partnership with OpenAI, as featured on their blog.

Baseten, Hugging Face, Goodfire Partner for AI Safety

Baseten's research arm, Base Labs, announced a partnership with Hugging Face and Goodfire AI to establish a safety standard for open-weight AI models. The collaboration aims to develop and publish methods for training and monitoring these models, making safety an integrated feature rather than an afterthought. This comes amid rising concern over 'abliteration,' a technique that removes safeguards from open models, with Hugging Face listing over 6,000 such models. The companies have not detailed technical specifics, but Goodfire's expertise in AI interpretability is expected to play a key role. This initiative seeks to leverage openness as an advantage for safety, providing transparent controls and encouraging community contributions.

• Baseten launched a safety infrastructure standard via Base Labs.
• Partnership with Hugging Face and Goodfire AI announced.
• Hugging Face hosts over 6,000 abliterated models.

Unlocking Claude's Full Value: Plans, Workflows, and Setup

This article analyzes the economics of Anthropic's Claude subscription plans, comparing them to OpenAI's offerings. SemiAnalysis found a $200 Claude Max plan provides about $11,700 worth of Claude Opus 5.5 usage at API prices, versus roughly $2,100 for OpenAI's equivalent. Most subscribers use only a small fraction of their allowance, making the plans profitable for Anthropic. Recent product changes, including the merger of Cowork into Claude chat, new model versions (Opus 5.5, Sonnet 5.5, Haiku 5.5), and monthly API credits on Max and Team plans, make it easier to use the full allowance. The article provides a comprehensive guide with model routing tables, a map of the Claude stack, a 5-layer operating system, and 17 workflows to help users maximize their subscription value.

• SemiAnalysis study published October 5 finds a $200 Claude Max plan provides ~$11,700 of Claude Opus 5.5 API usage, vs ~$2,100 for OpenAI's GPT-6.1 Sol.
• Anthropic's gross margin on an Opus-heavy Max subscriber is +6% at 20% usage and -369% at full usage.
• Cowork merged into Claude chat on September 16.

How GLM5.3 Sparse Attention Affects HBM Memory Usage

SemiAnalysis analyzes how GLM-5.3's sparse attention affects HBM and NAND memory usage. Sparse attention reduces KV cache memory and bandwidth during attention operations but does not reduce overall memory capacity needs, as top-k selection requires full context in HBM. The SGLang team's HiSparse system offloads KV cache to host DRAM to overcome capacity limits. InferenceX benchmarks show GB200 and GB300 cost advantages over AMD MI355X ATOM at different latency targets, and GLM-5's architecture is suspected to be optimized for Moore Threads MTT S4000.

• GLM-5 (and GLM-5.x) is a 744B total, 40B active-parameter mixture-of-experts model with 1 shared expert and 8 of 256 experts routed per token (sparsity 32).
• HiSparse, designed by the SGLang team, offloads KV cache entries from device HBM to host DRAM using an LRU cache with layer-wise overlapping from HiCache.
• InferenceX benchmarks show GB200 costs approximately $0.044 per million total tokens at 150 tokens/s, about 12% lower than $0.049 for MI355X ATOM.

SemiAnalysis Releases ClusterMAX 3.0 GPU Cloud Ratings

SemiAnalysis released ClusterMAX 3.0, its third GPU cloud rating system for neoclouds, covering 77 providers and expanding its market view to 323. Nebius joined CoreWeave in the Platinum tier, while Google Cloud moved to Gold. The methodology includes audits of compute, networking, storage, orchestration, reliability, and security, and the report introduces standardized SLAs and highlights security concerns. Key trends include financing structures, Blackwell/GB300 deployments, Vera Rubin transition, and agentic coding impact. Separately, Australian AI data center operator Firmus, backed by Nvidia, withdrew its planned IPO due to market volatility. The company had sought to raise $5 billion at a $30.6 billion valuation. It will now pursue private funding after raising $2 billion from Nvidia, Coatue, Blackstone, and Jane Street in August, totaling over $3 billion in equity. Firmus also has agreements with Meta for GPU capacity in Southeast Asia.

• ClusterMAX 3.0 was released covering 77 providers, with the market view expanded to 323 providers.
• Nebius was upgraded to Platinum tier, joining CoreWeave, while Google Cloud moved to Gold.
• Firmus withdrew its planned IPO, citing market volatility, and will pursue private funding.