Today's News
Latent.Space
Stacklok, founded by Kubernetes co-creators Craig McLuckie and Joe Beda, released Mecatl as an open-source agent harness that decouples the agent loop from tool execution, state management, and execution environments—enabling enterprises to manage agentic workloads centrally on Kubernetes rather than on individual desktops. The company open-sourced both Mecatl and ToolHive (MCP server platform) while building a proprietary "enterprise spine" control plane handling identity, authorization, policy, and auditing across multiple clusters and clouds—the monetization layer that mirrors how Kubernetes became the standard while companies monetized operational complexity.
McLuckie explicitly frames Stacklok's mission as helping large enterprises (banks, semiconductors, telcos) reduce dependency on frontier labs like OpenAI and Anthropic, similar to how Kubernetes reduced cloud provider lock-in. This positions agent infrastructure as the next competitive battleground: the operational layer that decouples workloads from model providers, potentially reshaping how enterprises manage AI workload economics and vendor relationships.
Anthropic via Techmeme
Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million input/output tokens (up to 100K tokens), matching GPT-6 Luna's sticker price while delivering 26-point Intelligence Index gains year-over-year (43 vs. 17 on Haiku 4.5). The model includes effort settings, adaptive thinking, and 1M context window. Simultaneously, Anthropic halved Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens, reducing agentic work costs by ~20%, and introduced $100–$500 monthly API credits for Max and Team subscribers usable across all models.
Haiku 5.5 introduces tiered pricing—5x cheaper under 100K tokens—explicitly positioning it as a subagent for high-volume work alongside Opus or Sonnet. On benchmarks, Haiku 5.5 (max effort) scores 43 on Artificial Analysis Intelligence Index, ahead of GLM-5.3 Flash (42) and GPT-6 Luna (38), and reaches 72.4% on OSWorld computer-use tasks (vs. 15.7% on Haiku 4.5). The pricing cuts and subscription credits signal Anthropic's strategy to capture the inference waterfall—where agents repeatedly call trained models, making more work economical to delegate as inference costs drop—and lock in developer adoption through bundled API spend.
OpenAI via Liberty's Highlights
OpenAI launched Dots, always-on agents powered by GPT-6 Astra with their own cloud computer, full filesystem access, memory inspection, and nightly reflection cycles. Users can view, edit, and download agent memories and instructions; the system learns from feedback over time and works 24/7 across ChatGPT, Slack, and Teams. Meta's Muse implements similar architecture with a "dreaming" process that reviews conversations and synthesizes behavioral improvements. This convergence on persistent context/workspace architecture signals industry consensus on agent stickiness and the economics of delegation.
Concurrently, Meta and Microsoft are cutting internal Claude usage by approximately 33-50% as their internal models approach performance parity with Anthropic's offerings—a concrete signal that Meta's larger models (including unreleased Watermelon) have reached cost/control thresholds where dogfooding benefits outweigh remaining capability gaps. The pullback suggests imminent release of stronger Meta models and potential architectural changes to Muse, while signaling competitive model progression and infrastructure strategy across the hyperscaler-lab ecosystem.
Epoch AI via Nathan Benaich
OpenAI and Anthropic combined reached a $105B annualized revenue run rate by summer 2026, up from $30B at year start—explosive growth reflecting booming inference demand as agents repeatedly call trained models, making more work economical to delegate as inference costs drop 47% per quarter since 2023. This signals the "inference waterfall" is now the industry's primary revenue engine, reshaping software economics as the cost of delegating work to AI agents continues to plummet.
