DeepSeek's Cheap Model Just Beat Its Own Flagship. V4-Flash-0731 Outscored V4-Pro-Preview on Agent Benchmarks at $0.14 a Million Tokens.
On July 31, DeepSeek moved V4-Flash out of preview and shipped it as an official release, tagged 0731. It scored 82.7 on Terminal-Bench 2.1, a benchmark that grades an agent on actually completing real terminal tasks, not answering multiple-choice questions about them. DeepSeek's own V4-Pro-Preview, the more expensive model from the same lab, scored 72.1 on the same test. The budget tier beat the flagship tier, from the same company, on the benchmark that most closely resembles what an agent actually does when you point it at your codebase.
I've written about DeepSeek's pricing moves twice this year on this blog (the April V4-Pro launch and the permanent price cut in May), and both times the story was "cheaper, and close enough on quality to matter." This one is different. This isn't "close enough." This is the cheap model winning outright on the metric that predicts whether an agent ships working code instead of confidently wrong code.
What actually changed on July 31
V4-Flash-0731 isn't a new model architecture. It's the same 284-billion-total, 13-billion-activated mixture-of-experts setup that shipped as a preview back on April 24. DeepSeek redid the post-training stage and left the base architecture alone. That matters because it means the jump from 72.1-ish preview performance to 82.7 on GA isn't a bigger, more expensive model doing the work. It's the same-sized model getting meaningfully better at agentic tasks through better fine-tuning, which is the cheapest possible way for a lab to improve a model and the hardest kind of improvement to sustain once you've squeezed it once.
Pricing on the GA release: $0.14 per million input tokens on a cache miss, $0.0028 per million on a cache hit, and $0.28 per million output tokens. Compare that to what frontier US labs charge for agentic coding work and you're looking at roughly two orders of magnitude difference, not a 20% discount. I want to flag one honest caveat here: DeepSeek has announced a peak/off-peak pricing scheme that would apply a 2x multiplier during high-demand windows, and as of this writing that scheme isn't in effect yet with no confirmed start date. Today's price is not guaranteed to be next month's price, and if you're building cost projections around this model, build in room for that.
Why Terminal-Bench 2.1 is the number to trust
Static benchmarks (the kind that ask a model to pick the right answer from four options, or generate a code snippet that gets checked against a unit test in isolation) are easy to game and don't tell you much about whether a model can run as an autonomous agent. Terminal-Bench 2.1 puts a model in an actual shell, gives it a real task (set up an environment, fix a failing build, debug an actual error), and scores whether it gets to a working end state. It's closer to what you're actually doing when you point Claude Code, Cursor, or an open-source agent harness at a real repo and ask it to fix something.
A model scoring 82.7 versus 72.1 on that specific test isn't a marginal quality gap. It's the difference between an agent that mostly finishes the job and one that needs you to step in and finish it yourself a meaningful fraction of the time. For a solo operator running agent tasks unattended (CI fixups, scheduled content generation, background data processing), that gap is the whole ballgame. A model that fails silently or half-finishes a task costs you more in cleanup time than the token savings are worth, no matter how cheap the tokens are.
What this does to the routing decision
Most solo operators I know run a rough mental model: expensive model for anything customer-facing or judgment-heavy, cheap model for background grunt work, and accept that the cheap model will occasionally embarrass you. That mental model assumed a real quality tax on the cheap tier. V4-Flash-0731 beating its own lab's pricier preview model on the benchmark that predicts real-world agent reliability breaks that assumption, at least for this specific pairing.
The practical move: if you've got a routing layer (LiteLLM, OpenRouter, a custom wrapper) that sends "important" tasks to an expensive model and everything else to a cheap fallback, it's worth re-running your actual task set against V4-Flash-0731 rather than assuming the old price-to-quality curve still holds. I did this on three of my own background jobs (a weekly content research pass, a PR description generator, and a log-triage script) and V4-Flash-0731 matched or beat the more expensive model I'd been defaulting to on two of the three, at roughly 4% of the cost.
The honest counter-take
Benchmark scores from a lab about its own model deserve some skepticism, even when the number itself is a specific, third-party-style benchmark rather than a marketing claim. DeepSeek picked which benchmark to headline, and Terminal-Bench 2.1 happens to be one where the newer post-training run does well: that's not necessarily representative of every task category. I also haven't seen independent, reproduced numbers from a source outside DeepSeek's own blog post as of this writing, which is a real gap; I'm treating the 82.7/72.1 comparison as DeepSeek's self-reported figure until a third party reproduces it.
There's also the pricing-stability question I flagged above. A model that's this cheap because a lab is subsidizing adoption, with an unannounced 2x peak surcharge sitting in the wings, isn't the same commitment as a model priced sustainably. If you build a product with hard dependency on this exact price point, you're making a bet on DeepSeek's pricing strategy staying stable, not just on model quality.
What I'd actually do
Test it against your actual workload before you trust the benchmark. Pull three or four of your recurring agent tasks (the ones you already run daily or weekly through whatever model you currently default to) and run them through V4-Flash-0731 for a week. Compare failure rate, not just cost. If it holds up on your specific tasks the way it held up on Terminal-Bench, move your background and non-customer-facing agent work to it and keep the savings. Don't move anything customer-facing over on benchmark numbers alone. Run your own eval first, because a lab's headline benchmark and your actual use case are not the same test.
Author
Lukas
@lukcombinator