THE SHORT ANSWER
● Same accuracy — Within two percentage points on the 447-task pool - statistically equivalent inside a five-point margin
● Cheaper — mini-SWE-agent, by up to 3x per task on the same model
● Smallest first call — mini-SWE-agent at 829 tokens, against OpenCode 7,025 and Claude Code 16,581
● Caveat — On DeepSeek-V4-Flash the cost ranking reverses - mini-SWE-agent cost slightly more
Shandong University tested three coding-agent harnesses on 492 offline tasks from SWE-bench Verified, in isolated containers. mini-SWE-agent is roughly 100 lines of Python with a single bash tool and no context compaction. Claude Code and OpenCode are full products. The finding that matters: the 100-line harness kept up.
Side by side
| Claude Code | OpenCode | mini-SWE-agent |
|---|
| Accuracy, 447-task pool | Baseline | Up to 9 pts behind | Within 2 pts |
| Initial fixed context | 16,581 tokens | 7,025 tokens | 829 tokens |
| Cost on GLM-5.3-Flash | ~2x mini | Lowest | ~1x |
| Cost on DeepSeek-V4-Flash | Slightly lower | — | Slightly higher |
| Context compaction | Yes | Yes | No |
| Output truncation issues | Not reported | 32k limit cost it ~half its gap | Not reported |
| Lines of code | Product | Product | ~100 |
What the accuracy numbers actually say
On the 447-task pool, Claude Code and mini-SWE-agent sat within two percentage points, which the researchers judged statistically equivalent within a five-point margin on both Qwen models tested. OpenCode trailed by up to 9 points - but about half that gap on Qwen3.8-27B came from runs cut off by a 32,000-token output limit, which is a configuration problem rather than a reasoning one.
Then the result that should make anyone cautious about harness comparisons: on the 45 hardest tasks, a median 13% of tasks flipped between pass and fail when switching harness. Rerunning an unchanged setup produced the same median flip rate. A meaningful share of any harness comparison is noise, including this one.
Where the cost goes
Initial fixed context is the cleanest explanation. Claude Code opens every task carrying 16,581 tokens of scaffolding before it reads your code. mini-SWE-agent opens with 829. Multiply across steps and sessions and that is where a 3x gap comes from - not from thinking harder, but from carrying more.
Which is why the DeepSeek-V4-Flash reversal matters. On that model mini-SWE-agent cost slightly more per trial than Claude Code, presumably because no compaction means more total steps on a model that handles long context differently. The overhead is not universally wasteful; it is a bet that pays off on some models and not others.
How to decide for your own setup
The researchers' recommendation is to stop comparing published scores and compare three measurable things: first-call context size, per-step context growth, and step count. All three are visible in your own logs within an hour of running both.
Their summary sentence is worth keeping: a vendor's own harness on its own model buys no score advantage. Claude Code running Claude Opus 5 was included as a top-tier reference and did not break that pattern.