DeepSeek V4 Flash 0731 is the official, re-post-trained release of DeepSeek's Flash model, and it scores 50 on the Artificial Analysis Intelligence Index — 10 points above the April preview it replaces. Same 284-billion-parameter architecture, same $0.14-per-million-token API price, but a very different model on agent and coding work.
Short answer: DeepSeek V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index (up from 40), roughly tying Gemini 3.6 Flash and sitting one point behind GPT-5.6. It costs $0.14 per 1M input tokens and $0.28 per 1M output on DeepSeek's own API. The gains come entirely from re-post-training — the 284B-parameter, 13B-active architecture didn't change.
Last updated: August 1, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
In my testing, I pulled the same coding and agent prompts I'd already run against the April Flash preview and pointed them at the 0731 build through DeepSeek's API, then cross-checked the published Intelligence Index numbers against my own timing and cost logs. The gap between the two Flash versions was bigger than I expected for what DeepSeek describes as a pure post-training update with no architecture change. Here's what the numbers actually mean and where the model earns its keep.
What you'll need
An API key from platform.deepseek.com if you want to reproduce any of this yourself — the model name is deepseek-v4-flash, and DeepSeek's own docs confirm "0731" is just the dated version string behind that name, not a separate model you select. You don't need anything to just read the benchmark numbers, but if you're going to verify pricing or run your own eval, budget a few dollars of prepaid balance and an OpenAI-compatible client. A basic sense of what the Artificial Analysis Intelligence Index measures helps too — it's a blended score across reasoning, coding, and math evals, not a single benchmark.
Step-by-step: reading the DeepSeek V4 Flash 0731 numbers
1. Confirm you're looking at 0731, not the April preview
DeepSeek shipped an initial V4 Flash preview in April 2026, then replaced it with this July 31 release. Both answer to the model name deepseek-v4-flash in the API, so check the version string in the response metadata or in DeepSeek's changelog if you need to know which one served a given request.
2. Check the Intelligence Index, not a single benchmark
Artificial Analysis puts 0731 at 50 on its Intelligence Index, up from 40 for the preview — a 10-point jump from post-training alone. That places it one point behind GPT-5.6 (max, 51), roughly level with Gemini 3.6 Flash (50), and seven points behind the open-weights leader, Kimi K3 (max, 57).
3. Look at the agent benchmarks separately
The Intelligence Index barely hints at what actually changed. On Terminal-Bench 2.1, 0731 jumped from 61.8 to 82.7. Cybergym went from 38.7 to 76.7, and DeepSWE — a harder agentic coding eval — moved from 7.3 all the way to 54.4. Those are the numbers that matter if you're routing agent or coding traffic to this model.
4. Price it against what you're replacing
At $0.14 per 1M input tokens (cache miss) and $0.28 per 1M output, 0731 costs the same as the preview did — DeepSeek didn't reprice the upgrade. Against DeepSeek's own V4 Pro ($0.435 input, $0.87 output), Flash still runs at roughly a third of the cost.
5. Run your own eval before trusting the vendor numbers
MarkTechPost’s writeup flags something worth repeating: all of DeepSeek's benchmark numbers for this release are vendor-reported on an unreleased test harness. Artificial Analysis's Index score is independently run, which is why I treat that number as the more trustworthy anchor and the vendor's own agent numbers as directional.
6. Factor in cache pricing if you're sending repeated prompts
Input tokens that hit DeepSeek's prompt cache bill at $0.0028 per 1M — a fraction of the $0.14 cache-miss rate. If your workload reuses a system prompt across many calls, which is the normal pattern for agent loops, your real cost per request lands well under the sticker price.
Example prompts you can copy
These are the prompts I used to sanity-check the benchmark claims against real output, not synthetic evals:
- "Given this failing test and the file it covers, find the root cause before proposing a fix: [paste]."
- "Write a script that does [task], then list the two most likely edge cases it misses."
- "Here's a multi-step task: [describe]. Break it into an ordered plan before executing any of it."
- "Return the answer as JSON matching this schema, no prose: [schema]."
- "Compare these two approaches to [problem] and state which one you'd ship, with a reason."
That last one is where the difference from the April preview showed up most: 0731 gave a direct pick with a stated tradeoff far more often, where the preview tended to hedge and describe both options without committing.
Common mistakes to avoid
The biggest one: assuming a higher Intelligence Index score means the model got smarter across the board. It didn't — DeepSeek re-post-trained the same 284B/13B-active weights specifically for agentic and coding tasks, so the jump is concentrated in tool use, terminal work, and multi-step coding, not general knowledge or open-ended reasoning. Second, comparing 0731's benchmark numbers to a competitor's self-reported figures instead of a shared, independent leaderboard — vendor numbers use different harnesses, and Artificial Analysis's Index is the one apples-to-apples source I'd trust here. Third, forgetting that the price didn't move: some people assume a capability jump this size comes with a cost increase, and it doesn't — $0.14/$0.28 held steady from the preview. Fourth, skipping the cache-hit discount in cost projections, which in my testing made the real per-request cost on a repeated-prompt agent loop noticeably lower than the quoted rate. Fifth, judging the model only on the Intelligence Index number and ignoring that Artificial Analysis separately flagged it as unusually verbose in its reasoning traces — that verbosity itself adds to your output-token bill.
DeepSeek V4 Flash 0731 vs. the field
| V4 Flash 0731 | V4 Flash (April preview) | V4 Pro | GPT-5.6 (max) | |
|---|---|---|---|---|
| Artificial Analysis Intelligence Index | 50 | 40 | Higher (full model) | 51 |
| Input price (per 1M tokens) | $0.14 (cache miss), $0.0028 (cache hit) | $0.14 | $0.435 (cache miss) | Not DeepSeek-comparable pricing |
| Output price (per 1M tokens) | $0.28 | $0.28 | $0.87 | Not DeepSeek-comparable pricing |
| Terminal-Bench 2.1 | 82.7 | 61.8 | Not published for this eval | — |
| Architecture | 284B total, 13B active | 284B total, 13B active | 1.6T total, 49B active | Closed |
Numbers confirmed on DeepSeek’s own API pricing page and in Artificial Analysis’s DeepSeek V4 Flash 0731 writeup as of July 31, 2026. On Artificial Analysis's separate GDPval-AA v2 agent benchmark, 0731 posted an Elo of 1559 against 1189 for the preview — the second-highest open-weights score behind Kimi K3 (max, 1687) and ahead of GLM-5.2 (max, 1510).
Tools that make this easier
If you haven't actually set up DeepSeek V4 Flash yet — API key, base URL, first call — my full DeepSeek V4 Flash setup guide covers that end to end; this article is the numbers behind the model, not the setup steps. For the plain consumer chat app instead of the API, see how to use DeepSeek, and for wiring any DeepSeek model into a coding tool like Cursor or Claude Code, how to use DeepSeek for coding walks through the base-URL swap. If you're deciding whether DeepSeek belongs in your stack at all, my DeepSeek vs. ChatGPT test compares output quality and cost head to head, best AI models in 2026 puts DeepSeek next to the rest of the field, and my Claude Opus 5 review covers the closed-model alternative if price per token isn't your main constraint.
Frequently Asked Questions
Is DeepSeek V4 Flash 0731 a new model?
No, not architecturally. It's the same 284B-total, 13B-active mixture-of-experts model as the April preview, re-post-trained on agentic and coding data. DeepSeek's own materials describe it as a post-training-only upgrade with no new weights design.
How much does DeepSeek V4 Flash 0731 cost?
$0.14 per 1M input tokens on a cache miss, $0.0028 per 1M on a cache hit, and $0.28 per 1M output tokens through DeepSeek's own API — unchanged from the April preview's pricing.
What is DeepSeek V4 Flash 0731's Intelligence Index score?
50 on the Artificial Analysis Intelligence Index, up from 40 for the previous Flash version. That puts it roughly level with Gemini 3.6 Flash and one point behind GPT-5.6.
Is DeepSeek V4 Flash 0731 better than V4 Pro?
Not on raw reasoning — Pro is the larger, more capable model. But 0731 closes a meaningful part of that gap on agentic and coding tasks specifically, at about a third of Pro's per-token price, which is why I'd default to Flash first and only reach for Pro on tasks that actually need it.
Where can I check the latest DeepSeek V4 Flash 0731 benchmarks?
Artificial Analysis runs its Intelligence Index independently and updates it as new models ship, which makes it a more reliable comparison point than any single vendor's own benchmark post.