DeepSeek V4 Flash 0731 pushes agent performance down to sub-$0.30 per million tokens

On July 31, 2026 DeepSeek released DeepSeek-V4-Flash-0731, the official V4 Flash drop after preview. On Hugging Face the checkpoint is MIT-licensed at 304B parameters (167 GB). Artificial Analysis lists the reasoning-max variant at 284B total and 13B active parameters with a 1M-token context. The vendor promises substantially enhanced agentic capabilities and ships DSpark speculative decoding with the checkpoint.

Official Flash beats DeepSeek's own Pro preview on agent benches

DeepSeek's table shows a jump over Flash preview and V4-Pro preview: Terminal Bench 2.1 at 82.7 (preview 61.8, Pro 72.1), Toolathlon-Verified 70.3, Cybergym 76.7, DeepSWE 54.4. On Agents' Last Exam it scores 25.2, just under Opus-4.8 at 25.7. Code-agent tasks ran in the minimal mode of the still-to-be-released DeepSeek Harness at max reasoning effort.

Simon Willison flagged it as a release that punches well above its weight: Artificial Analysis ranks it ahead of MiniMax M3 (428B), and on the Intelligence Index vs. cost chart Flash 0731 (max) sits deep in the attractive upper-left quadrant. AA score 50, roughly 124 output tokens/s, pricing about $0.14 / $0.28 per million input and output tokens. Willison also showed that default reasoning produced a weak pelican drawing, while high reasoning recovered a much better result.

For teams with an agent harness, unit economics of a run change

The practical point is not another screenshot ladder. It is the cost of long tool-use loops: debugging, terminal agents, multi-step coding. When an open-weight model holds agent benches near the closed frontier while API pricing is an order of magnitude cheaper than the tip of the market, buyers have to ask whether premium spend is still justified only because a vendor logo looks safer on slides.

MIT weights also open a self-host path through vLLM and SGLang with built-in DSpark. That matters for teams that refuse to ship code and logs into someone else's SaaS but still need agent throughput. Reasoning effort levels low/high/max and a recommended 384K max output on high/max show the model is built for long deliberation, not chat autocomplete alone.

Vendor benches and a pelican test are not a purchase proof

Model-card numbers are strong, but some sets are internal (DSBench) and code agents run inside a vendor harness that is not public yet. That is the standard trap: scores do not transfer 1:1 into your OpenHands, Codex-like runner, or internal orchestrator. Willison's pelican test is a small but useful reminder that default sampling can look cheaply dumb until you raise reasoning effort.

The second risk is operational. Early user reports praise price and agentic debugging, and already ask whether DeepSeek can hold latency and pricing once cheap Flash becomes a traffic magnet. A Pareto bargain is fragile when the whole market of agent jobs stamps on it at once.

Independent harness runs outside DeepSeek's table will decide adoption

Three signals matter. First, reproduction of Terminal Bench, Toolathlon, and real coding-agent jobs outside the official minimal harness. Second, API price and rate-limit stability after the first adoption wave. Third, whether self-host recipes on GB300/vLLM keep promised throughput beyond blog happy paths.

Until that lands, Flash 0731 is a strong A/B candidate against expensive closed models, not an automatic full-stack swap. When outside teams start reporting comparable agent results at fractional cost in their own harness, the buying default changes, not just the leaderboard.

Lilith's verdict

DeepSeek is again moving the front line where agent-harness budgets break: when Flash holds Terminal Bench around 83 % at fractions of a dollar per million tokens, every expensive closed model has to justify each extra cent.