The Official DeepSWE Number Is Footnoted to DeepSeek Harness

DeepSeek released its V4-Pro model and an open-source developer tool, DeepSeek Harness, on August 13. The official DeepSWE score for V4-Pro is 62.7, up from 12.8 for the preview, with the footnote indicating tests were run inside the harness. The harness features a plugin system with revertible effects, named Cordis, and peak API pricing is $3.96 per million tokens.

On August 13, DeepSeek shipped four things in one day. V4-Pro left preview. Peak and off-peak API prices start at 16:00 UTC on August 16. An 88-page draft paper went up on spatiotemporal composability. And deepseek-harness dropped as a developer preview, MIT license, one command: npx @deepseek-ai/dsh web . The model headlines ate the English YouTube tab. I stayed on the harness. Official figure for DeepSeek-V4-Pro-0813: DeepSWE 62.7, against 12.8 for the preview. The footnote says public code-agent sets were tested inside DeepSeek Harness in minimal mode. Fable 5 is still at 70.0 on the same grid. The loop is a plugin. Unload is written as revertible effects. They named that implementation Cordis. Peak output on V4-Pro is $3.96 per million tokens. I put the footnote, the plugin topic, and the peak-hour bill into a four-minute video. The August 13 write-up is still the plumbing version: DeepSeek's Harness Is the Price Signal https://dev.to/reidmarlow/deepseeks-harness-is-the-price-signal-17bl . This is the scoreboard version. If you already run Claude Code with a DeepSeek key taped on, does the plugin topic change anything for you?