Blog

Notes on building reliable AI products.

DeepSeek V4 Flash Vision release showing its low-cost hosted API beside the large open-weight checkpoint

DeepSeek V4 Flash Vision Is Open. The API Is Still Harder to Beat

DeepSeek released MIT-licensed V4 Flash Vision weights. Here is the cost, deployment reality, benchmark caveats, and agent-builder playbook.

Split API timeline showing the Assistants shutdown and the February 2027 transcription model cutoff

OpenAI’s Assistants API Is Gone. Four Transcription Models Are Next.

OpenAI shut down Assistants and set a February 2027 cutoff for four transcription models. Here is the migration plan builders actually need.

Tencent Hy4 Preview model branching into a hosted one-million-token API and a cluster-scale self-hosting path

Tencent Hy4 Preview: Why a 770B Open-Weight Agent Model Is Easier to Rent Than Run

Tencent’s Hy4 Preview offers 1M context, Apache-2.0 weights, and cheap agent APIs. Its real test is retrieval, caching, and route economics.

OpenAI model supply cable splitting across local and hosted Cursor workflows after the SpaceX acquisition

OpenAI’s Planned Cursor Cutoff Makes Model Access an M&A Risk

OpenAI plans to end managed model access in Cursor after SpaceX’s acquisition. Here’s what BYOK preserves—and why the 5% figure misleads builders.

Claude Code capacity gauge moving from a temporary 150 percent weekly allowance to a permanent 125 percent allowance with an 83.3 percent warning line

Claude Code’s New Weekly Limit Has an 83.3% Warning Line

Anthropic will reset Claude Code weekly limits to 125% of the old baseline on September 14. Here is the 83.3% threshold teams need to plan around.

An AI safety control board marked low risk while a coral path bypasses classifier and audit-log controls for 133 million exchanges

Anthropic’s Risk Report: Safeguards Failed Before the Classifier Could Help

Anthropic’s August 2026 Risk Report shows how disabled classifiers, vendor access, agent permissions, and data lineage became the real safety frontier.

Branching Claude token choices pass through a keyed watermark pattern into a likelihood gauge shaped by EU AI Act compliance

Claude's Global Watermark Measures Token Choice, Not Authorship

Anthropic will watermark future Claude text globally. Here is what the EU AI Act changes—and why builders need provenance workflows, not a detector boolean.

GLM-5.3 base model block feeding a conveyor of coding-agent sandboxes, verifiers, and training trajectories

GLM-5.3 Started Where Pretraining Stopped

Z.ai says GLM-5.3 improves coding and cyber-agent performance without a new base model. Here is what builders can verify, test, and trust.

Five illuminated AI model engines from OpenAI, Anthropic, xAI, Google, and DeepSeek feeding a central routing console for capability, speed, cost, context, and control

AI Models 2026: Release Updates and August Comparison

AI model release links checked October 5, plus an August 14 comparison of GPT, Claude, Gemini, Grok and DeepSeek with dated benchmark and pricing tables.

Desktop app events flowing through a timeline into local agent memory, live source recovery, and a reviewed automation

OpenAI Computer History Turns Desktop Activity Into Agent Memory

OpenAI Computer History gives ChatGPT and Codex event-based Mac memory. Here is how the data path works, what it risks, and how teams should test it.

Google Gemini 3.7 Flash agent relay connecting multimodal input, reasoning, tool use, verification, and a December 31 pricing gate

Gemini 3.7 Flash: Benchmarks, Pricing, and API Guide

Gemini 3.7 Flash is live. See benchmarks, API pricing, the Gemini 3.6 comparison, thinking levels, multimodal limits, and a migration guide.

A fleet of AI agents separated from protected evidence and authority layers by identity, quotas, and arbitration gates

Anthropic’s Agent Swarms Need an Operating System, Not a Better Group Chat

Anthropic’s multiagent study shows why agent swarms need quotas, ownership, independent arbiters, and stop conditions—not just more capable models.