DeepSeek V4 Flash 0731 is the most cost-efficient agentic AI model available right now. For about $0.14 per million input tokens, you get a 1-million-token context window, native tool-calling, and independent benchmark scores that leapfrogged from 37 to 50 on the Artificial Analysis Intelligence Index — all without a single change to the underlying architecture or price.
Verdict: If you're building AI agents, coding assistants, or high-volume tool-using pipelines, V4 Flash 0731 is the best value-per-token option in 2026. It's not a frontier model — it trails DeepSeek V4 Pro, Claude Fable 5, and GPT-5.6 Luna on the hardest reasoning tasks — but for 60% less than the price of a comparable model, it wins on the agentic work that actually runs your automations. The upgrade shipped July 31, 2026 as a drop-in API change: set your model name to deepseek-v4-flash and you're on the new version.
Last verified: 2026-08-01 · Best value agentic model: DeepSeek V4 Flash 0731 · Intelligence Index: 50 (#3 open weights) · Price: $0.14/$0.28 per 1M tokens · Context: 1M tokens
Pricing and benchmarks change often — last checked 2026-08-01.
What is DeepSeek V4 Flash 0731?
DeepSeek V4 Flash 0731 is the official public-beta release of DeepSeek's efficiency-tier AI model, shipped on July 31, 2026 under the build name deepseek-v4-flash-0731 (DeepSeek API Change Log). It's a 284-billion-parameter Mixture-of-Experts (MoE) model with only 13 billion active parameters per token — meaning it delivers large-model quality at small-model speed and cost. The "0731" refers to the release date.
The upgrade is purely in post-training: same architecture, same parameter count, same 1-million-token context window — but materially better at agentic tasks, coding, and tool use. DeepSeek confirmed that V4-Flash-0731 "keeps the same architecture and size as V4-Flash-Preview, only re-post-trained" (DeepSeek API Change Log).
What changed from the preview to the official release?
The official release improves on the April 2026 preview across every agentic benchmark DeepSeek tested, with zero architecture changes. The upgrade includes:
- Post-training fine-tuning — significantly stronger agent, coding, and tool-calling performance.
- Native Responses API support — adapted specifically for Codex-style coding agents (DeepSeek API Change Log).
- Full interface compatibility — continues to support OpenAI ChatCompletions and Anthropic-style APIs, so switching is a configuration change, not a re-integration.
Important scope: Only the Flash API was upgraded. The DeepSeek V4 Pro API and the models on DeepSeek's app and website remain unchanged. If you're using the DeepSeek chat web interface, you are NOT using the 0731 build — it's API-only for now.
How much does DeepSeek V4 Flash cost?
DeepSeek V4 Flash is one of the cheapest serious AI APIs on the market, and the 0731 release did not change pricing:
| Component | Price per 1M tokens | Source |
|---|---|---|
| Input tokens | $0.14 | DeepSeek pricing |
| Output tokens | $0.28 | DeepSeek pricing |
| Cache-hit input | $0.003 | Artificial Analysis |
| V4 Pro input (for comparison) | $0.435 | DeepSeek pricing |
| V4 Pro output (for comparison) | $0.87 | DeepSeek pricing |
Flash costs roughly one-third what V4 Pro costs. For context, OpenAI's GPT-5.6 Luna scores 51 on Artificial Analysis Intelligence Index (one point above Flash) but costs substantially more — DeepSeek reports the new Flash is approximately 60% cheaper than GPT-5.6 Luna at comparable intelligence (the-decoder.com).
Caveat: "Cheap per token" is not "cheap per task." Agentic loops that run many tool calls and reasoning steps can rack up significant token usage. Cap reasoning effort and tool iterations on simple calls.
How good is V4 Flash at agentic and coding tasks?
The 0731 release delivers its biggest gains in agentic benchmarks. These are DeepSeek-reported scores using their own harness settings (DeepSeek API Change Log) — treat as vendor figures and validate on your own workloads:
| Benchmark | Flash Preview | Flash 0731 | What it tests |
|---|---|---|---|
| Terminal Bench 2.1 | 61.8 | 82.7 | Agentic command-line / software tasks |
| Cybergym | 38.7 | 76.7 | Security / CTF-style problem solving |
| DeepSWE | 7.3 | 54.4 | Software engineering |
| NL2Repo | 39.4 | 54.2 | Repo-level code generation |
| Toolathlon (verified) | — | 70.3 | Breadth and reliability of tool use |
| DSBench-FullStack | — | 68.7 | Multi-file coding competence |
| DSBench-Hard | — | 59.6 | Hard data-science tasks |
| Agent Last Exam | — | 25.2 | Deliberately hard agentic reasoning |
| Automation Bench (Public) | — | 25.1 | Agentic SaaS workflows |
Sources: DeepSeek API Change Log, OrcaRouter
The Artificial Analysis Intelligence Index independently measured the upgrade: V4 Flash 0731 (Reasoning, Max Effort) scores 50, up from 37 for the April preview build — a 13-point jump that moved it to #3 among 101 open-weight models on the index (Artificial Analysis). For comparison, OpenAI GPT-5.6 Luna scores 51, GLM 5.2 scores 51, and Kimi K3 leads open weights at 57.
How does V4 Flash compare to V4 Pro?
DeepSeek's V4 line uses a two-tier "cheap-default-plus-premium-flagship" ladder. Here's how they stack up:
| Feature | V4 Flash 0731 | V4 Pro |
|---|---|---|
| Architecture | 284B / 13B-active MoE | Larger flagship model |
| Context window | 1,048,576 tokens (1M) | Same V4 family |
| Max output | 384K tokens | Same V4 family |
| Input modality | Text-only | Text-only |
| Price (in/out per 1M) | $0.14 / $0.28 | $0.435 / $0.87 |
| Best for | High-volume agents, coding, long docs | Hardest reasoning, frontier-level tasks |
| Status | Public beta (July 31, 2026) | Available; official Pro upgrade coming soon |
Simple routing rule: Send the hardest 10-20% of tasks to V4 Pro; route everything else to V4 Flash. The intelligence gap is narrow (50 vs. higher), but the cost gap is 3x.
How do you start using V4 Flash 0731?
The upgrade is a drop-in API change. If you're already calling the Flash model, you're on the new version — the model name deepseek-v4-flash now points to the 0731 build.
- Get an API key from platform.deepseek.com.
- Set the model name to
deepseek-v4-flashin your existing OpenAI-compatible client. - Verify you're on 0731 — if you were using a preview snapshot, update to the latest model identifier. The API calling method is unchanged (DeepSeek API Change Log).
- Try the Responses API — V4 Flash 0731 natively supports the Responses API format and is specifically adapted for Codex-style coding agents. Both OpenAI ChatCompletions and Anthropic-style interfaces continue to work.
- Use third-party providers — V4 Flash is also available through OpenRouter (
deepseek/deepseek-v4-flash) and OrcaRouter (0% markup) if you want routing, fallbacks, or simplified billing.
For agent frameworks: If you're building with a multi-model agent system, create a separate profile per model so you can plug in new releases the moment they drop. This pattern — swapping models via config rather than re-integrating — is how production agent stacks absorb releases like 0731 within hours, not weeks.
What are the limitations of V4 Flash 0731?
Honest caveats from primary sources:
- Not a frontier model. It trails V4 Pro, Claude Fable 5, and top-tier models on the hardest reasoning tasks. DeepSeek themselves position it as the efficiency tier, not the flagship.
- Not multimodal. Input is text-only — no native image input (Artificial Analysis).
- Vendor benchmarks. The agentic scores above use DeepSeek's own harness settings. Independent Artificial Analysis scores are available (50 Intelligence Index) and are the more conservative figures to cite.
- API-only upgrade. The DeepSeek app and website still run the older model. This upgrade is only accessible via the API.
- Open weights coming soon. DeepSeek has indicated open-weight release is planned (OrcaRouter). Once released, this will be the second-highest open-weight score behind Kimi K3.
What this means for you
For developers and small businesses running AI agents: V4 Flash 0731 is the sweet spot. At $0.14/$0.28 per million tokens with a 1M context window, it can handle entire codebase analysis, long-document processing, and multi-step agent workflows in a single pass — at a fraction of what frontier models charge. Switch your model config to deepseek-v4-flash, run your own evaluation on your actual workload, and route the hardest cases to V4 Pro or a frontier model.
For teams building multi-agent systems: The pattern that matters isn't chasing each new model — it's having infrastructure where new models plug in as config changes. If you have a solid agent operating system, absorbing V4 Flash 0731 takes minutes. If your agent code is hardcoded to one model, you'll always be playing catch-up.
How does V4 Flash fit into the 2026 AI landscape?
The 0731 release reflects a broader 2026 trend: models improving through post-training rather than scale. DeepSeek didn't build a bigger model — they trained the same 284B MoE architecture to be a sharper agent. That means the cost curve is bending down even as capability rises.
This matters for how you think about your AI token cost optimization strategy: the cheapest model that completes the task is now also a very capable one. The gap between budget and flagship models is narrowing, and post-training upgrades like 0731 are the reason. For a deeper look at how mixture-of-experts models enable this, see our guide to serving AI agents at scale with sparse attention.
For teams evaluating open-weight models specifically, V4 Flash 0731 (once weights are released) and Kimi K3 are now the two leading open-weight options. Our Kimi K3 open-source model guide covers the rival that currently leads on the Intelligence Index.
FAQ
Q: Is DeepSeek V4 Flash 0731 a new model or an update?
A: It's a post-training update to the existing V4 Flash model. The architecture (284B total / 13B active MoE parameters) and the 1-million-token context window are unchanged. DeepSeek retrained the same model to be significantly better at agentic, coding, and tool-calling tasks. If you were already calling deepseek-v4-flash, you're automatically on the new version.
Q: How much does DeepSeek V4 Flash cost per million tokens?
A: $0.14 per million input tokens and $0.28 per million output tokens, with cache-hit input dropping to $0.003 per million. That's roughly one-third the price of DeepSeek V4 Pro ($0.435/$0.87) and approximately 60% cheaper than OpenAI's GPT-5.6 Luna, which scores only one point higher on the Artificial Analysis Intelligence Index (51 vs. 50).
Q: Can V4 Flash 0731 replace a frontier model like Claude Fable 5 or GPT-5.6 Luna?
A: For most agentic and coding workloads, yes — it's fast, cheap, and handles 1M-token contexts natively. For the hardest reasoning tasks (frontier-level math, cutting-edge coding competitions), you'll want V4 Pro, Claude Fable 5, or GPT-5.6 Luna. The practical approach is routing: Flash as the default, Pro or a frontier model for the hardest 10-20% of tasks.
Q: Does V4 Flash 0731 support the Responses API and Codex?
A: Yes. The official release adds native Responses API support and is specifically adapted for Codex-style coding agents, while continuing to support OpenAI ChatCompletions and Anthropic-style interfaces. This means you can drop it into existing OpenAI-compatible or Anthropic-style integrations without changing your client code.
Q: When will DeepSeek V4 Pro get upgraded?
A: DeepSeek has confirmed that the V4 Pro official upgrade is "coming soon" but has not specified a date. The July 31 release only upgraded the Flash API. The company has also indicated that open weights for V4 Flash 0731 will be released soon, which would make it the second-highest open-weight model on the Artificial Analysis Intelligence Index, behind Kimi K3.
Q: Is the DeepSeek app and website using V4 Flash 0731?
A: No. The 0731 upgrade is API-only. The models on DeepSeek's app and website remain unchanged. To use the latest version, you need to call the API directly with the model name deepseek-v4-flash, or go through a hosting provider like OpenRouter or OrcaRouter.
Discussion
0 comments