DeepSeek shipped V4 Flash 0731 on July 31 with a claim that stopped people mid-scroll: a 284-billion-parameter MoE model, activating only 13 billion parameters per token, that beats its own larger flagship - V4-Pro-Preview - on every agentic benchmark the company published. The price did not move. DeepSeek's pricing page held at $0.14 per million input tokens on a cache miss and $0.28 per million output tokens, with a 2,500-request concurrency limit. That combination - smaller, faster, cheaper, and apparently better at agent tasks - is worth examining closely, because the story is real, but also messier than the launch posts let on.
What actually changed on July 31
The 0731 release is the official version of DeepSeek V4 Flash, superseding the April preview. The model card is explicit: architecture and parameter count are unchanged. Every performance gain comes from a new post-training pipeline focused on coding, agents, reasoning, and tool use.
That matters more than it sounds. Most labs ship new-architecture releases and call them improvements. DeepSeek is saying: we took the same weights, ran them through a better post-training process, and the resulting model is now materially stronger on the tasks that matter for agent deployments. Weights ship under an MIT license on Hugging Face, and the API adds native OpenAI Responses API support for Codex CLI users.
The stable API model name remains deepseek-v4-flash, so existing integrations receive the new snapshot without changing the route.
That is either convenient or unsettling depending on your perspective - your calls started returning a different model on August 1 without any change on your end.
The benchmark numbers, and the asterisks they carry
On Terminal Bench 2.1, an agentic terminal-use benchmark, Flash-0731 scores 82.7, up from 61.8 on the preview and ahead of V4-Pro-Preview's 72.1. On DeepSWE, the jump is sharper: 7.3 to 54.4 - a more than 7× improvement in a single post-training cycle, and well clear of V4-Pro-Preview's 12.8.
Those are striking numbers. Here is what cuts through the hype.
The DeepSWE score of 54.4 was obtained using DeepSeek's own Harness framework, which the company acknowledged in the same changelog would be released "soon" - meaning no third party can yet replicate the test.
The 82.7 Terminal-Bench 2.1 score was also obtained with this proprietary scaffolding. Independent evaluations using standard harnesses - SWE-bench's official scaffold, open-source agent frameworks - may produce materially different numbers.
Alongside the version confusion runs a second claim: that V4 Flash matches Claude Opus 4.8 at a fraction of the price. DeepSeek published a nine-row comparison against Opus 4.8. Opus 4.8 leads all nine.
The closest the two models come is on Agents' Last Exam, where V4-Flash-0731 scores 25.2 versus Opus 4.8's 25.7 - a gap of half a point. That is genuinely close. On everything else, the gap is real.
The non-obvious insight here: harness variance across evaluations can produce score variance that exceeds model-to-model differences. A 5-point gap between two models on Terminal-Bench may mean less than a 15-point gap caused by differences in retry logic, tool selection, prompt formatting, and context management. When you see DeepSeek's nine-row table, you are reading system-level measurements, not a clean model comparison.
Beagle in action#eng-infra, 9:41amWhat the pricing actually means for agent workloads
This is where V4 Flash 0731 makes a genuinely strong case, independent of how benchmark disputes resolve.
At $0.28 per million output tokens - roughly a third of V4-Pro's output price - seed-stage startups, indie developers, and internal platform teams can run agent loops at this price without a GPU budget.
To make that concrete: a coding agent that burns 20 million output tokens per day (roughly 400 pull-request-sized tasks) costs $5.60 at Flash pricing. At V4-Pro rates that is $17.40. At GPT-5.6 Luna's post-cut rate of $1.20/M output - which OpenAI announced the day before this release - the same workload runs $24.00. OpenAI cut GPT-5.6 Luna's price by 80% to $0.20 input and $1.20 output per million tokens the day before DeepSeek's announcement. Flash's output token price is still less than a quarter of that.
| Model | Input $/M | Output $/M | Output at 20M tok/day |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | $0.14 | $0.28 | $5.60 |
| DeepSeek V4 Pro | $0.435 | $0.87 | $17.40 |
| GPT-5.6 Luna (post-cut) | $0.20 | $1.20 | $24.00 |
| Claude Opus 4.8 | ~$3.00 | ~$15.00 | $300.00 |
Prices sourced from DeepSeek's pricing page and Artificial Analysis as of August 9, 2026. Claude Opus 4.8 pricing is approximate.
At 115 tokens per second, V4 Flash 0731 is notably fast for its size tier
- which matters for agent loops where latency compounds across tool calls. Time to first token on DeepSeek's API is 1.15 seconds, competitive against the 1.85-second median for open-weight models of similar size.
One flag to log: DeepSeek's pricing page now warns that overall API pricing will rise in the near future and that the increase is expected to be significant. It does not disclose a multiplier, model-by-model rates, or an effective date.
DeepSeek has also indicated that billing will double inside two Beijing-time windows per day, though no effective date has been published. Teams treating today's Flash pricing as a permanent baseline are taking a risk.
Self-hosting reality check
The weights are MIT-licensed and ungated, but every expert stays resident in memory even though only 13B activate per token. DeepSeek's vLLM example serves it on a single 4×GB300 node. Unsloth's dynamic GGUFs put the lossless 8-bit build at 162 GB and a 3-bit build at 103 GB, needing roughly 110 GB of combined RAM plus VRAM.
That is not a laptop. "Open weight" here means you can run it - but running it at production throughput requires hardware most teams do not have idle. For most teams the API is the practical path.
DeepSeek V4 Flash 0731: common questions
What is DeepSeek V4 Flash 0731?
DeepSeek V4 Flash 0731 is the official release of DeepSeek V4 Flash, superseding the preview version, with substantially enhanced agentic capabilities. The architecture is unchanged from the April preview - 284B total parameters, 13B active via MoE - and all gains come from a new post-training pipeline focused on coding, tool use, and agent workflows.
Does V4 Flash 0731 beat Claude Opus 4.8?
No. V4 Flash 0731 does not close the gap to the frontier entirely, and Opus 4.8 still leads on every benchmark DeepSeek published. The closest contest is on Agents' Last Exam, where the two models are within half a point. On Terminal-Bench and most coding evals, Opus 4.8 leads by meaningful margins.
Can I self-host DeepSeek V4 Flash 0731?
Yes, the weights are MIT-licensed and published on Hugging Face. Unsloth's dynamic GGUFs put the lossless 8-bit build at 162 GB and a 3-bit build at 103 GB, needing roughly 110 GB of combined RAM plus VRAM. For most teams, the API at $0.14/M input is the practical path.
Why do the benchmark scores have caveats?
The key headline scores - Terminal-Bench 2.1 and DeepSWE - were obtained with DeepSeek's own Harness framework, which is proprietary scaffolding that no third party can yet replicate. The benchmark itself may be credible; the harness is a variable no one outside DeepSeek can control for yet.
Is V4 Flash 0731 a good fit for Slack-based agent workflows?
At $0.14/M input and $0.28/M output with native Responses API support and 2,500 concurrent requests, it is priced right for high-volume internal tooling. A teammate like Beagle would pair well with a model at this price point for draft-and-approve loops where token volume is high and latency tolerance is a few seconds. The pricing-rise warning from DeepSeek is worth factoring into any long-term infrastructure plan.