Coding Agents

GPT-6 Astra costs exactly the same as Fable 5.1 on the API. But 62.7% vs. 99.9% on ARC-AGI-3 shows that the headline number belongs to the harness. With a reading strategy and decision matrix for agent builders.

DeepSeek-V4-Pro-0813 went GA on August 13, posting a Terminal-Bench 2.1 score of 87.9% — the first open-weight model to beat a frontier closed model on that benchmark. But the GA weights aren't on Hugging Face, a 264% price hike hits August 16, and several self-reported benchmarks still await independent replication.

Z.ai's GLM-5.3 keeps the same 743B base model as GLM-5.2 but extracts massive gains through post-training scaling alone. Terminal-Bench 3.0 jumps 514%, and unexpected cyber capabilities have already found 2,436 real-world vulnerabilities.

Grok 4.5 changes the model-routing question: when does a cheaper near-frontier model beat the benchmark leader on cost per accepted task?

Alibaba Qwen 3.7 Max changes the agent economics conversation because Alibaba did not ship another chat model. It shipped a long horizon agent backend with a 1M token context window, official Claude Code compatibility,

Codex 0.133 is not a feature checklist. It is the clearest sign yet that coding agents are becoming managed execution environments: they can see the product, pursue a durable goal, and carry teamspecific workflows

Vercel deepsec shows why AI-coded apps need repeatable security harnesses, second-agent revalidation, and controlled merge gates.

OpenAI’s Codex safety post turns coding-agent adoption into a control system: sandboxing, approvals, network policy, credentials, and telemetry.

The most productive AI coding setup in 2026 isn't one model — it's two. Here's how pairing Claude Opus 4.6 for architecture with Gemini 3.1 Pro for execution creates a dual-model AI coding stack that outperforms either alone.

A new study proves AGENTS.md files reduce AI coding agent runtime by 28.6% and token usage by 16.6%. We use AGENTS.md daily — here is the definitive practical guide.