Dropstone is not a foundation model. It is a runtime that turns open-weight models into functional agents. The version number tracks the integration cycle, not the model weights. Each generation, we evaluate the leading open-weight models on public coding benchmarks and deploy whichever performs best, on infrastructure that does not trust the model any more than the last one. Dropstone 1.6 is that evaluation, six months into the platform's evolution from a coding-focused tool into a general-purpose agent that plans, calls APIs, browses, writes, and executes multi-step tasks. This post covers what changed, what the numbers say, and where we still trail. Dropstone 1.6 ships with new backends across all three tiers. Heavy now runs on GLM-5.2, a 744-billion-parameter model with roughly 40 billion active per token. Pro moves to Kimi K2.7 Code. Fast moves to DeepSeek V4 Flash. GLM-5.2 launched in June 2026 and advanced from GLM-5.1's 58.4% on SWE-bench Pro to 62.1% in a single step, the largest one-generation jump we have swapped in on the Heavy tier to date. GLM-5.2 is now the highest-scoring open-weight model we track. On SWE-bench Pro it resolves 62.1% of issues, ahead of GPT-5.5 at 58.6%. On Terminal-Bench 2.1 it reaches 81.0%, the first open-weight model we have measured above 80% on that benchmark. On FrontierSWE it scores 74.4%, roughly one point behind Claude Opus 4.8. We are not going to soften the rest of that comparison. Claude's frontier models still lead. Fable 5 scores 80.3% and Opus 4.8 scores 69.2% on SWE-bench Pro, and the gap is largest on the hardest tier of tasks, where it runs to 18.2 points. GLM-5.2 closed most of the gap this cycle. It did not close all of it, and we are not claiming otherwise. The model swaps every cycle. The runtime guarantees do not. Inference runs exclusively on US-hosted infrastructure. Every tool call, file edit, shell command, and API call sits behind an approval gate before it executes, regardless of which model is behind the wheel. Credits pool weekly across every tier: 500 on Free, 23,000 on Pro, 122,500 on Max. Heavy carries a 1M-token context window, dual thinking-effort control, and output up to 131,072 tokens per turn. Commercial and enterprise sessions are excluded from training entirely, with no exceptions. Consumer sessions can improve the underlying models, but only with consent, and opting out is available at any time. What we retain from consumer sessions is text only. Raw images are not kept. This cycle we started routing real, execution-verified coding sessions back into refining the open model behind Dropstone. Call it a data flywheel: more usage, better signal, a better next model. We have not yet measured a capability gain from it, and we should not be credited for one until we have. It is a direction, not a result. We will report back once there is a number worth reporting. Dropstone Max is $75 a month and delivers comparable or better results than Claude Code at $100, Codex at $200, and Cursor at $200. GLM-5.2 delivers its SWE-bench advantage at roughly one sixth of the output cost of the models it is closing the gap on. The economics are the same argument they were at 1.5: the cost curve on open-weight frontier models keeps bending down, and Dropstone passes that down on every tier, not just the top one. The base models behind every tier are trained by third parties. We cannot audit their training data, and we do not pretend to. The Claude frontier gap is real, not a rounding error, and it is widest on the hardest tasks. Pro and Fast do not yet share a common capability axis with Heavy, so cross-tier comparisons in this report should be read carefully. Some of the Terminal-Bench comparisons this cycle mix vendor-reported and independently-run scores, and we have labeled which is which rather than blend them. The cost advantage assumes current US-hosted inference rates, which can move. And the data flywheel, as above, is unproven. Dropstone 1.6 is live today across Fast, Pro, and Heavy. Full benchmarks and methodology are published in our technical report at https://blankline.org/research/dropstone-1-6-technical-report.