DeepSeek
Open-source LLM optimized for advanced reasoning and code
Start new thread

DeepSeek-V4-Flash-0731 - Frontier agent intelligence at Flash prices

by•
DeepSeek-V4-Flash-0731 is the official release of V4-Flash, featuring a massive leap in agentic capabilities. It outperforms V4-Pro (Preview) on key benchmarks, natively supports the Responses API, and is fully adapted for Codex CLI.

Add a comment

Replies

Best
Hunter
📌

Hi everyone!

This is another moment.

DeepSeek-V4-Flash-0731 keeps the same architecture and size as the , but the jump in agentic performance is hard to treat as a normal update.

Terminal-Bench 2.1 moved from 61.8 to 82.7. DeepSWE went from 7.3 to 54.4 (simply CRAZY). Flash now beats V4-Pro Preview on every benchmark shown in DeepSeek’s release table, while activating far fewer parameters.

under MIT too, so this is not just an API release.

Then there is the price. V4 Flash currently costs $0.14 per million uncached input tokens and $0.28 per million output tokens. .

Top-tier intelligence is becoming extremely cheap. I mean extremely cheap.

I think we are entering a different phase of AI. When intelligence at this level is almost free and , what will you build? How far can your imagination go when the cost of trying is no longer the main constraint??

P.S. one slightly crazy hint👀👀:

The number I care about isn't $0.14 per million, it's cost per completed task. A cheaper model that needs two retries on an agent run costs more than a pricier one that lands it first, which is why 61.8 to 82.7 on Terminal-Bench is the line that actually moves my bill. Where it gets interesting is cached input pricing, since on long agent loops most of my spend is context I'm re-sending, not new tokens.

61.8 to 82.7 on Terminal-Bench and 7.3 to 54.4 on DeepSWE for a model that's "the same architecture and size" as the preview is the kind of jump that makes me want to know who ran the eval, not just what it scored. are these numbers reproduced by anyone outside DeepSeek yet, or is it still first-party only? weights being MIT and downloadable makes independent verification actually possible here, unlike a closed API release, so I'd rather wait for someone else's harness to confirm it than take the release table at face value.

 Right question, and I would point it at one number. 7.3 to 54.4 on DeepSWE sitting next to 61.8 to 82.7 on Terminal-Bench, same architecture, same size. That asymmetry is the tell. A 7.3 baseline is not a model that reasons badly. It is a model falling out of the agent loop. So a lot of that 47 point delta is probably tool-call and format adherence rather than new capability. Which matters because adherence gains are harness shaped. They transfer if your loop looks like theirs and quietly do not if it does not. Has anyone seen the failure breakdown on the old 7.3? Format versus reasoning is the whole question.

Is this becoming a price match, Oai also reduced a lot on their frontier models..

Where does it land on long-horizon agent tasks vs. raw benchmarks? That gap is usually where the cheaper tiers fall short.

thanks to deepseek for gpt price cut.

I'm really enjoying using DeepSeek.

This is likely the most cost efficient model right now, works better than kimi 2.7code, and def better option than gemini flash. 5.6 Luna and Grok build 0.1 are both very good as well

I think the biggest shift isn't that intelligence is getting cheaper—it's that experimentation is becoming cheaper. When the cost of trying hundreds of ideas approaches zero, the competitive advantage shifts from access to AI to identifying the right problems worth solving.