MODEL ECONOMICS · JULY 31 RELEASE
DeepSeek moved DeepSeek-V4-Flash-0731 into public API beta on July 31: MIT-licensed, 284B total / 13B active parameters, a 1M-token context window, and pricing of $0.14 per million input tokens and $0.28 per million output tokens, re-post-trained to 82.7 on Terminal-Bench 2.1, up from 72.1 on the April preview. Opus 5 scores 89.1 on the same benchmark at $5 / $25 per million tokens. This is a small working router: set what share of a monthly coding workload is boilerplate versus hard, and see what routing the boilerplate share to DeepSeek actually costs against an Opus-5-only baseline, and what quality you are trading for it.
Same monthly coding spend, two routing setups
Set what you spend today running everything through Opus 5, then set what share of that work is boilerplate, boilerplate routes to DeepSeek-V4-Flash-0731, everything else stays on Opus 5.
OPUS 5 ONLY
$1,200
$5 / $25 per Mtok
TIERED
$725
40% on DeepSeek at $0.14 / $0.28 per Mtok
SAVED
$475
40% less per month
Quality tradeoff
Blended Terminal-Bench 2.1 score at this split: 86.5 (Opus 5 alone: 89.1, DeepSeek-V4-Flash-0731 alone: 82.7). This is a task-count-weighted proxy, not a real accuracy measure, the point is that the tradeoff is small next to the cost drop.
Savings use DeepSeek-V4-Flash-0731's output-token price ($0.28/Mtok) against Opus 5's ($25/Mtok) as the routing ratio, since agentic coding runs output-heavy; the input-token ratio is cheaper still (2.8% of Opus 5's input price). Pricing and Terminal-Bench 2.1 scores are as published in DeepSeek's July 31, 2026 model card and API docs, and Anthropic's Opus 5 announcement. This is a spend estimate against a number you set yourself, not a live billing tool.
WHY THIS MATTERS FOR CLIENT WORK
Not every task a client pays for needs a frontier model. Boilerplate scaffolding, CRUD endpoints, internal scripts, and repetitive glue code do not need the model that scores 89.1 on Terminal-Bench 2.1, they need one that is correct and cheap. DeepSeek-V4-Flash-0731 is open enough to self-host and priced low enough that routing that share of the work off Opus 5 is close to free by comparison. The catch is the one DeepSeek states itself: its agentic benchmark numbers were run on a minimal harness mode that has not been released, so those scores are vendor-reported until I run my own tasks through it. And since DeepSeek is a China-based lab, running it through their own hosted API sends client work to their infrastructure, the same data-residency caveat that applied to Kimi K3 on this radar on July 29; the MIT license means self-hosting or a Western inference host is the actual fix, not the API endpoint DeepSeek ships by default.
Built 3 August 2026 · pricing and benchmark numbers sourced from DeepSeek's model card and API docs