DeepSeek V4 Flash costs a tenth as much and passes 53.8% of agent tasks

The testing firm Composio put DeepSeek's V4 Flash through 30 deliberately difficult multi-step agent workflows and recorded a 53.8% pass rate: 129 of 240 runs ^[4][5]. That number belongs in any deck that proposes migrating production agents to V4 Flash on the strength of its pricing, which the model's promoters have priced at roughly a tenth of comparable models ^[3][11].

The sticker price is real. V4 Flash lists at $0.14 per million input tokens and $0.28 per million output tokens, which the report describes as undercutting comparable models by roughly tenfold ^[2][3]. The model launched on July 31, was called a "total monster" by developers, and went to the top of multiple leaderboards ^[1]. It is also still in public beta, with a broader adjustment to DeepSeek's API pricing scheduled for August 16, 2026 ^[10].

What Composio measured is closer to what operators actually buy. The 30 tasks used live tools -- Gmail, GitHub, Slack and Google Sheets -- and were run across eight agent harnesses including Claude Code, Codex and OpenCode ^[7][6]. Thirty workflows over eight harnesses is 240 runs, so each workflow was attempted once per harness ^[13]. Of those 240 attempts, 111 failed ^[12]. Only six of the 30 workflows were completed successfully by every harness tested, which is 20% of the task set that can be called reliable independently of integration choices ^[9][15].

The spread across harnesses is the more useful finding. Pi Agent completed 20 of 30 tasks, a 66.7% rate, while the aggregate across all eight sat at 53.8% ^[8][14][4]. Per Composio's framing, the same underlying model looks brilliant or mediocre depending on how it is wired into a workflow, and thoughtful integration can partially compensate for model limitations ^[8][16]. That cuts both ways: a procurement decision made on model price alone is being made on the smaller of the two variables.

Now the arithmetic that a migration memo should include. At a 53.8% pass rate, you need about 1.86 attempts per completed task, which puts the effective cost of a successful run at roughly $0.26 per million input tokens and $0.52 per million output tokens ^[17]. That is still well under a competitor at ten times the sticker price, so the cost case does not vanish on retries alone. It vanishes on the parts Composio did not price: failed runs consume compute, developer debugging time, and end-user trust, per the report's own accounting ^[11]. A workflow that touches a customer's Gmail or a production GitHub repo and fails 46% of the time is not a cheaper version of one that works ^[7][12].

Three things to watch. First, the August 16, 2026 API pricing change, which will tell you whether the tenfold gap is a launch posture or a durable position ^[10]. Second, whether anyone reruns the Composio suite once V4 Flash exits beta, since the beta label is DeepSeek's own signal that this is unfinished ^[10]. Third, your own harness: if Pi Agent's 66.7% is the ceiling in this test set, the honest question is not which model to buy but whether your integration is closer to the best harness or the worst ^[14][8].