Build Detail

The first (and so far only) non-Anthropic/OpenAI/xAI/Moonshot vendor on this page: DeepSeek Harness 0.1.1-rc.2 running DeepSeek V4 Pro at high reasoning as the orchestrator, delegating to DeepSeek V4 Flash sub-agents — confirmed against the user’s own DeepSeek Platform usage screen, which shows exactly those two models for the day and nothing else. The brief is the usual one-shot instruction used across this whole page, with a harness-specific typo preserved verbatim in the transcript: “simulate user behaviour using the Playwright MPC.” The run did not stop once — single turn, 320 steps, start to finished commit in 1h 31m.

The delegation shape is clean and disjoint: three sub-agents — Admin UI, Storefront UI, Database Seeders — spawned together in the same step, each fenced to its own file scope so the parallel writers can’t collide, followed much later by a fourth, 2nd-Agent Verification, spawned alone after the orchestrator had already made its own bug-fix commits. That verification agent is the interesting part: it ran for 2 minutes 12 seconds, made 43 tool calls, every one of them bash or read — php artisan test, migrate:fresh, db:seed, route:list, pint --test, a tinker sanity check — and reported back “10/10 PASS” on what it called an “Execution checklist.” It never opened a browser, because none was available: the harness’s entire tool palette across all 5 session files is bash/read/write/edit/grep/glob/subagent_fork/todo_write plus a handful of orchestration primitives — no MCP client, no screenshot tool, nothing that could drive Chromium interactively. The orchestrator itself worked around this by shelling out to the playwright npm test framework (a scripted assertion suite, not an exploratory session), which is a materially different thing than what the brief asked for. The “full review meeting” the mission demanded, and the “verified by a 2nd agent” claim the run reports back, amount to a green Pest suite and a CLI smoke check — specs/testplan.md, the 143-test acceptance suite this whole page scores every build against, was never run by anyone in this session.

Economically this is easily the cheapest run on the page in real terms: $4.97, read directly off DeepSeek’s own official-vendor billing dashboard (821 requests, 229.0M tokens for the day — within 0.3% of this parser’s own count, confirming the screen is this exact run). Running the identical token volume through OpenRouter’s list price for the same two models instead would have cost $12.39 — DeepSeek’s own API is roughly 2.5× cheaper than routing through OpenRouter for itself, a genuinely useful data point for anyone deciding how to bill this model. 99.7% of all input tokens were cache reads, which is most of why the number is so low despite 320 orchestrator steps and 732 more sub-agent tool calls on top. No vision-model fallback occurred anywhere in the run, despite a vision-capable sibling model being advertised on the same DeepSeek platform page the user screenshotted — there was no browser tool to feed it a screenshot in the first place.

Net: the fastest and cheapest one-shot run measured here, produced by a harness with a narrower tool palette than every other agent on this page — no browser automation at all, which forced the “simulate user behaviour” and “verified by a 2nd agent” parts of the brief into their weakest form yet: a real Pest suite, and a verification pass that checked everything except what a user would actually see. Our own independent, browser-driven pass (below, and in specs/2026-08-25-deepseek-v4-pro.md) found what that gap predicts — see the Feature Tests page for the score the run’s own “10/10 PASS” missed.