DeepSeek’s V4 Flash has dominated model leaderboards since its July 31 rollout, but real-world testing tells a different story. Composio’s stress test drove V4 Flash through eight distinct agent setups — Codex, Claude Code, and OpenCode included — using 30 hard multi-step assignments that touched live services such as Google Sheets, Slack, Gmail, and GitHub. Each workflow ran eight times across the harnesses, 240 attempts in all, with 129 passing overall; only six of the 30 tasks were completed by every setup.
The outcome differences, which swung with the harness, provider stack, tool configuration, caching behavior, and retry handling, make the case that orchestration matters as much as the model itself.
The test lands as DeepSeek hikes prices for V4 Flash and V4 Pro by up to 1,100% depending on model, token type, and time of use. Off-peak, Flash pricing lands at $0.22 per million input tokens and $0.66 per million output tokens, with peak periods roughly doubling those rates. V4 Pro, meanwhile, is priced at $0.66 per million input tokens and $1.98 per million output tokens outside peak hours. Cache hits rise between 52% and 1,100%.
The move shifts the story away from ultra-cheap Chinese models as enterprises figure out where each model fits. Flash remains the most-used model on OpenRouter by weekly token volume.
