Open models caught Sonnet on coding tasks; the harness is where the gap moved

GLM 5.2 scored 91.9 on a run of nearly 1,000 coding-agent tasks. Sonnet 4.6 scored 90.8, and MiniMax M3 landed between them at 91.4. Tessl ran the evaluation internally against tasks drawn from its skills registry, scoring each model both with and without the relevant skill loaded, and put the dataset on Hugging Face (task-evals-for-skills, writeup). An open-weights model finishing ahead of Sonnet at slightly lower cost per task ($0.289 against $0.296) would have been the headline a year ago. The more useful finding is in the column breakdown: task completion barely separated the field, sitting at 96.9 to 97.8 for everything except Qwen, while the real spread showed up in instruction following, where GLM hit 87.4 and Sonnet 86.1. The models can mostly get the work done. Whether they do it the way you asked is where they diverge, and that is a harness problem as much as a weights problem.

Which is why most of the interesting work in the last day sat around the model rather than inside it. One developer running Claude Code posted the 4-tier routing stack he runs alongside it for calls that don’t justify frontier rates. DeepSeek V4 Flash sits in front as an orchestrator doing classification and triage at 2 to 5 seconds a decision; GLM-5.2 handles strategic analysis and, at max effort, multi-step reasoning; Opus 4.8 is reserved for high-stakes calls on sanitized inputs only. After tuning, 78% of calls resolve at the orchestrator or advisor tier, 18% reach deep reasoning, and 4% hit Opus. The tuning lesson is specific and worth stealing: the orchestrator kept treating long inputs as hard inputs, so a long summarization job got escalated when it shouldn’t have been. Route on reasoning depth, not token count.

Two posts converged on a related question, where project context should actually live. A Claude Code user argued that CLAUDE.md is the wrong home for most of it. It earns its keep for stable repo behavior, how to run tests, conventions, the folder nobody should touch, but it bloats and goes stale the moment you start pushing product decisions, customer reports, and open loops into it. His split keeps CLAUDE.md for repo behavior, treats issues and PRs as explicit source material, and adds a separate memory layer for people, decisions, and changes over time, with the agent expected to show which source it drew from. A Show HN project, MACCHA, pushes the same instinct into infrastructure: a 7-tier file-based context layer plus a working-memory engine with vector embeddings, confidence decay, and semantic conflict detection, shared across Antigravity, OpenCode, and Claude Code so all three read one persistent store. Both are bets that retrieval and provenance, not a fatter instructions file, are what keep an agent oriented across sessions.

NVIDIA Research made the sharpest version of the architecture argument with SpatialClaw, a training-free framework for spatial reasoning. Rather than handing a vision-language agent a fixed menu of structured tool calls, it gives the agent a persistent Jupyter kernel preloaded with perception primitives (SAM3 for segmentation, Depth-Anything-3 for 3D reconstruction) alongside NumPy, SciPy, and Matplotlib, then lets it write one Python cell at a time, read the output, and revise before committing to an answer. Across 20 spatial benchmarks it averaged 59.9% accuracy, 11.2 points over the prior best spatial agent, and it held that margin under one shared prompt and toolset across six VLM backbones ranging from 26 billion to 397 billion parameters. The portability across model sizes is the striking part: the gain came from the action interface, not from any single tuned model. It is the same wager as the code-mode retrieval work from earlier in the week, applied to perception instead of search.

The thread running through all four is that as the model field compresses, differentiation moves into the harness: how calls are routed, where memory lives, what interface the agent acts through. Tessl says it will run GLM against Opus next. Watch whether the instruction-following gap that separated the open models from Sonnet survives against a frontier model, or whether that too turns out to be a context-engineering problem rather than a weights one.