Deepseek gives away the agent harness and charges more for the loop it runs

Deepseek moved its flagship out of testing, released its proprietary agent software as open source, and announced higher API prices in the same breath ^[1]. The scaffolding is now free to take, while the tokens it burns cost more, and the steepest increase lands on cache hits, which is the single thing an agent loop does most ^[1][12][13].

Start with the money, because it has a deadline. The new rates begin August 16 at 4:00 p.m. UTC ^[6], after an end-of-June announcement that named neither figures nor a date ^[7]. Time-of-day pricing itself is not new at Deepseek; it ran night discounts on V3 and R1 from February 2025 ^[8]. What is new is the level. Off-peak V4-Pro input goes from $0.435 to $0.66 per million tokens and output from $0.87 to $1.98 ^[10], which is roughly 52 percent more for input and 128 percent more for output ^[2][1]. Peak rates double that, to $1.32 and $3.96 ^[11], so peak output is about 4.6 times the old rate ^[4]. Cache hits move from $0.003625 to $0.022 off-peak and $0.044 at peak ^[12], about 6.1 times and 12.1 times the old price ^[3]. Deepseek's own framing is that the cache discount shrinks from roughly one-hundred-twentieth of the input price to one-thirtieth, and that agents repeatedly reading the same files are the worst affected ^[13].

Put a workload against it. An agent run that re-reads 200 million cached tokens over its lifetime cost about $0.73 before; it now costs $4.40 off-peak and $8.80 at peak ^[7]. Deepseek says this partially undoes a price cut it made in May, and that cache hits will end up costing more than they did before that cut ^[14]. The increase arrives while the company raises capital and prepares an IPO ^[15].

The peak windows are 1 a.m. to 4 a.m. and 6 a.m. to 10 a.m. UTC, which Deepseek ties to the Chinese workday ^[9], or about 9 a.m. to noon and 2 p.m. to 6 p.m. in China ^[6]. That leaves seven peak hours and seventeen off-peak hours a day ^[5], and European teams sit in the cheap band for most of the afternoon ^[9]. It also means your unit economics now depend on when your cron jobs fire.

The giveaway is real. Deepseek Harness v0.1 ships as a Developer Preview under MIT, pitched against OpenAI's Codex and Claude ^[16], built on a plugin system called Cordis in which tools, sandboxes, sessions and the UI are all swappable ^[17]. A continuous session log records every prompt, tool call and result, and runs can be resumed, branched and replayed ^[18]. A minimal mode reduces the setup to a shell and a file editor, which is what Deepseek uses for its own benchmark runs ^[19]. It launches through npx into a local web interface, and Deepseek warns of compatibility problems ^[20]. The project is led by Cui Tianyi, who joined from Jane Street in March 2026 ^[21]; 712 projects signed up for the beta within three days of the early-August call ^[22].

On capability, Deepseek's table shows Terminal Bench 2.1 rising from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7, with wins over Claude Opus 4.8 on several agent benchmarks ^[23], gains of 15.8 and 49.9 points ^[9]. Artificial Analysis has V4-Pro moving 45 to 53 on its Intelligence Index, level with GLM-5.2 and behind Muse Spark at 57, Qwen 3.8 Max at 58, Kimi K3 at 60 and Claude Opus 5 at 63 ^[24], a ten-point gap to the top ^[8]. Weights for the new build are not out; the April preview is still what sits on Hugging Face ^[25]. The update was partly a response to V4 Flash, whose 0731 build nearly matched the Pro Preview on the same index at a fraction of the price ^[26].