DeepSeek V4 Flash 0731 — Ranking, Cost, and Similar Models
Snapshot: 2026-08-01, based on the official release on 2026-07-31. Prices and benchmark values change quickly; the linked primary sources are the authority for current values.
The model in one paragraph
DeepSeek V4 Flash 0731 is the official public-beta release of DeepSeek's efficiency-tier model, shipping on 2026-07-31. It keeps the identical sparse mixture-of-experts architecture of V4 Flash Preview — 284B total parameters, 13B active at inference, 1M-token context, text in/out only — and was only re-post-trained. Despite the unchanged size, a post-training round sharply improved agentic, coding, and tool-calling ability, taking its Artificial Analysis Intelligence Index score from 40 to 50 and its agentic Elo from 1189 to 1559. Pricing is unchanged at **$0.14 / \(0.28 per 1M input/output tokens**, with a distinctive **98% cache-hit discount** (\)0.0028 per 1M).
Intelligence ranking
On the Artificial Analysis Intelligence Index (v4.1), max-effort / adaptive reasoning configurations:
| Rank | Model | Intelligence Index |
|---|---|---|
| 1 | Kimi K3 (max) — open-weights frontier | 57 |
| 2 | Claude Opus 4.8 | 56 |
| 3 | GPT-5.6 Terra (max) | 55 |
| 4 | Claude Sonnet 5 | 53 |
| 5 | Muse Spark 1.1 (xhigh) | 51 |
| 6 | GPT-5.6 Luna (max) | 51 |
| 7 | GLM-5.2 (max) | 51 |
| 8 | DeepSeek V4 Flash 0731 | 50 |
| 9 | Gemini 3.6 Flash | 50 |
| 10 | DeepSeek V4 Pro | 44 |
| 11 | DeepSeek V4 Flash (previous) | 40 |
Reading of the numbers:
- One point behind GPT-5.6 Luna (max, 51) and within one point of GLM-5.2 (max, 51) and Muse Spark 1.1 (xhigh, 51).
- In line with Gemini 3.6 Flash (50).
- 7 points behind the open-weights frontier set by Kimi K3 (max, 57).
- 6 points ahead of DeepSeek's own V4 Pro (44) — the efficiency model outranks the premium model on this index.
Index breakdown vs the predecessor
DeepSeek V4 Flash 0731 improves on every Intelligence Index evaluation over the previous V4 Flash:
| Evaluation | V4 Flash 0731 | Previous V4 Flash | Delta |
|---|---|---|---|
| GPQA Diamond | 91% | 90% | +1 |
| AA-LCR | 66% | 63% | +3 |
| Humanity's Last Exam | 37% | 32% | +5 |
| SciCode | 50% | 45% | +5 |
| CritPt | 17% | 8% | +9 |
| Intelligence Index | 50 | 40 | +10 |
Agentic evaluations (Artificial Analysis):
| Evaluation | V4 Flash 0731 | Previous V4 Flash | Delta |
|---|---|---|---|
| GDPval-AA v2 (Elo) | 1559 | 1189 | +370 |
| Terminal-Bench 2.1 | 79% | 62% | +17 |
| τ³-Bench Banking | 31% | 23% | +8 |
Once weights are released, the 1559 Elo on GDPval-AA v2 will be the second highest open-weights score, behind Kimi K3 (max, 1687) and ahead of GLM-5.2 (max, 1510).
Agentic work ranking
Artificial Analysis' GDPval-AA v2 measures agentic real-world work tasks:
| Model | GDPval-AA v2 Elo |
|---|---|
| Kimi K3 (max) | 1687 |
| DeepSeek V4 Flash 0731 | 1559 |
| GLM-5.2 (max) | 1510 |
Official agent benchmarks (DeepSeek)
DeepSeek reports these first-party agent results for V4-Flash, claiming results far exceeding V4-Pro-Preview (tested with the DeepSeek Harness minimal mode, max effort, topp 0.95, temperature 1.0):
| Benchmark | Score |
|---|---|
| Terminal Bench 2.1 | 82.7 |
| Cybergym | 76.7 |
| Toolathlon verified | 70.3 |
| DSBench-FullStack (internal) | 68.7 |
| DSBench-Hard (internal) | 59.6 |
| NL2Repo | 54.2 |
| DeepSWE | 54.4 |
| Agent Last Exam | 25.2 |
| Automation Bench (Public) | 25.1 |
Hallucination / AA-Omniscience
The AA-Omniscience Index improvement is driven entirely by fewer hallucinations, not higher accuracy:
| Metric | V4 Flash 0731 | Previous V4 Flash | Delta |
|---|---|---|---|
| AA-Omniscience Index | -16 | -23 | +7 |
| Hallucination rate | 84% | 95% (report: -12) | -11 |
| Accuracy (correct %) | 37% | 37% | 0 |
At 84%, the hallucination rate is comparable to GPT-5.6 Terra (max, 85%) and Mistral Medium 3.5 (82%).
Token efficiency
The improved model also uses fewer output tokens in the evaluation run:
| Model | Output tokens for Intelligence Index run |
|---|---|
| DeepSeek V4 Flash 0731 | ~206M |
| Previous DeepSeek V4 Flash | ~234M |
That is a 12% reduction in output tokens for a 10-point higher score.
Cost
First-party list prices
| Metric | Price |
|---|---|
| Input (cache miss) | $0.14 / 1M tokens |
| Output | $0.28 / 1M tokens |
| Cache hit input | $0.0028 / 1M tokens (98% discount) |
The 98% cache-hit discount is significantly more aggressive than the ~90% offered by most of the industry, and is a key driver of DeepSeek's low cost per task.
Cost per task vs comparable models
Artificial Analysis reports that even after OpenAI's 80% price cut on GPT-5.6 Luna, V4 Flash 0731's cost per task on the first-party API is ~60% lower than GPT-5.6 Luna (max), a model of comparable intelligence.
The model lands on the Artificial Analysis Pareto frontier for Intelligence vs Cost per Task.
Reference task arithmetic
reference task = 20,000 input tokens + 5,000 output tokens
no prompt-cache discount, no tool-call fees, no retries
| Model | Input $/1M | Output $/1M | Reference task cost |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | $0.14 | $0.28 | $0.0042 |
| Gemini 3.6 Flash | $1.50 | $7.50 | $0.0675 |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.0500 |
| GPT-5.6 Terra | $2.50 | $15.00 | $0.1250 |
| Claude Sonnet 5 | $3.00 | $15.00 | $0.1350 |
| Claude Opus 4.8 | $5.00 | $25.00 | $0.2250 |
These are arithmetic estimates, not observed costs for completed tasks. They answer "what would this token mix cost?", not "how many tokens will the model need to finish my task?".
How the model is served
- Model name:
deepseek-v4-flash— the API name is unchanged and points to the 0731 build. - API: OpenAI-compatible Chat Completions and Anthropic-format endpoints, plus native Responses API support and a specific Codex adaptation.
- Scope of the release: only the V4-Flash API upgraded. The V4-Pro API and the APP/WEB models are unchanged; the official V4-Pro release follows soon.
- Weights: expected to be released as open weights in the coming weeks.
- OpenRouter serves it as
deepseek/deepseek-v4-flash-0731at \(0.14/\)0.28.
Caveats
- The Artificial Analysis token totals are full evaluation-suite runs, not per-task traces; they cannot be converted to tokens-per-task without the suite's task count and accounting rules.
- First-party agent benchmark scores use DeepSeek's own harness and settings; third-party numbers (Terminal-Bench 2.1 at 79% vs 82.7 official) differ by methodology.
- Public third-party eval data for this build is only days old and will firm up as more providers and evaluators pick it up.
Bottom line
- Choose DeepSeek V4 Flash 0731 when you want GPT-5.6 Luna-class intelligence (50 vs 51) at roughly 60% lower cost per task, especially for agentic workloads where the GDPval-AA v2 Elo of 1559 ranks second among open-weights models.
- Choose Gemini 3.6 Flash as the closest-price comparable if you need multimodal input.
- Choose Kimi K3 (max) or Claude Opus 4.8 when raw intelligence on this index matters more than cost.