DeepSeek pushed AI coding costs closer to zero on August 1, 2026, releasing DeepSeek-V4-Flash-0731 out of preview at $0.14 per million input tokens and $0.28 per million output tokens, pricing that undercuts most Western competitors by an order of magnitude while beating DeepSeek’s own larger flagship model on agentic coding benchmarks.
Quick facts
- DeepSeek-V4-Flash-0731 exited preview on August 1, 2026, priced at $0.14 per million input tokens and $0.28 per million output tokens.
- It scores 82.7% on Terminal-Bench 2.1 and 54.4% on DeepSWE, beating DeepSeek’s own 1.6-trillion-parameter V4-Pro-Preview model on agentic coding benchmarks.
- It ships with native Responses API support and Codex compatibility, plus DeepSeek’s own speculative decoding stack for 2-3x token throughput on H800 hardware.
- Legacy deepseek-chat and deepseek-reasoner API aliases have been retired in favor of deepseek-v4-flash.
- DeepSeek-V4-Pro’s full general availability remains pending, reportedly delayed into the August 10-20 window.
Why a smaller model beating a bigger one is the actual story
Per Axios’s reporting, the significant part isn’t just the price, it’s that a lighter, cheaper “Flash” model outperforming DeepSeek’s own much larger Pro-tier model on real coding benchmarks confirms something the industry has been circling for a while: post-training technique and data quality are starting to matter more than raw parameter count for practical coding tasks. That’s a direct threat to the pricing power of every lab still charging a premium primarily on the basis of model size.
The race to zero is now industry-wide, not just DeepSeek
DeepSeek isn’t cutting prices in isolation. OpenAI cut its own budget-tier Luna model from $1/$6 to $0.20/$1.20 per million tokens on July 30, 2026, and Anthropic’s Claude Sonnet 5 is running introductory pricing specifically to stay competitive during this window. Every major lab is now defending against the same pressure: a genuinely capable model at a fraction of the price forces competitors to either match on price or clearly justify a premium with capability that Flash-tier pricing can’t touch yet, like the hardest reasoning, judgment, and cybersecurity tasks.
What this actually means for anyone building on it
For agent and application builders running high-volume coding tasks, deepseek-v4-flash’s combination of price and Terminal-Bench score makes it a genuinely serious default option, not just a cheap fallback, for workloads where DeepSeek’s benchmarks hold up against your actual codebase. The retirement of the legacy deepseek-chat and deepseek-reasoner aliases means anyone still pointed at those needs to migrate their API calls to deepseek-v4-flash directly, a small technical task worth doing sooner rather than later given the old aliases are no longer the current model.
Common questions
Is DeepSeek-V4-Flash open-weight? DeepSeek’s prior model generations have been released as open-weight; the reporting reviewed here covers the API release specifically and doesn’t confirm an open-weight release date for this exact checkpoint.
Do I need to change my code to use it? If you’re using the legacy deepseek-chat or deepseek-reasoner model aliases, yes, those are retired and calls should point to deepseek-v4-flash directly. New integrations should use the current model name from the start.
When is DeepSeek-V4-Pro coming out? Reporting points to a general availability window between August 10 and August 20, 2026, though DeepSeek has not confirmed an official date; treat any specific date as unofficial until DeepSeek’s own channels confirm it.
Key takeaway
If cost per token is a meaningful line item in your AI spend, DeepSeek-V4-Flash-0731 is worth benchmarking against your actual workload now, not waiting for V4-Pro’s delayed general release. The gap between this model’s price and its coding benchmark scores is large enough that it changes the math for high-volume use cases specifically, even if it doesn’t touch the hardest reasoning tasks frontier labs still charge a premium for.
Leave a Reply