DeepSeek V4 Flash Achieves Near-Frontier Performance at 3¢ Per Task, Undercutting All Major Rivals
Summary
Updated Aug 22 — DeepSeek V4 Flash gains vision capabilities and runs on a single DGX Spark workstation at 47 tokens per second.
- • DeepSeek V4 Flash scores 50 on the Artificial Analysis Intelligence Index — a 10-point jump over prior V4 Flash and 6 points higher than the larger DeepSeek Pro version
- • Priced at just 3 cents per task vs. 59 cents for GLM-5.2 and 36 cents for Meta Muse Spark — potentially the most cost-efficient capable model available
- • Performance ties Gemini 3.6 Flash and lands one point below GLM-5.2 and GPT-5.6 Luna despite being a small model
- • Uses 12% fewer tokens than prior iteration and shows a significant agentic performance jump on GDPval-AA, suggesting improved reasoning efficiency for multi-step tasks
Updated Aug 16 — DeepSeek V4 Pro GA (Aug 13) scores 53 on the Intelligence Index — 3 points above V4 Flash — at roughly 10x Flash's per-token price.
Updates
Vision capabilities added to V4 Flash
DeepSeek V4 Flash gains multimodal/vision capabilities, extending the model beyond text to handle image inputs — moving the cost-efficient flash-tier model into direct competition with multimodal offerings from US labs.
Runs on DGX Spark at 47 tok/s
V4 Flash confirmed to run on a single NVIDIA DGX Spark workstation-class device at 47 tokens per second, enabling on-premise enterprise deployment without data center hardware. Shifts the enterprise question from 'Which API?' to 'What has to leave our building?'
DeepSeek V4 Pro GA released August 13
DeepSeek-V4-Pro-0813 released August 13, 2026 via API, app, and web platform (Expert Mode); GA follows April 2026 preview. Features flexible reasoning effort levels and native OpenAI Responses API support.
1.6T-parameter MoE, 1M-token context
V4 Pro uses a 1.6-trillion-parameter Mixture-of-Experts architecture with 49 billion active parameters and a 1-million-token context window — significantly larger than V4 Flash (284B total, 13B active).
V4 Pro: $1.32/$3.96 per million tokens (peak)
Peak pricing: $1.32 per million input tokens, $3.96 per million output tokens; off-peak rates are halved. V4 Flash costs $0.14 input / $0.28 output at peak — making V4 Pro roughly 9–14x more expensive per token.
V4 Pro Intelligence Index 53 — now above Flashrevised
Artificial Analysis assigns V4 Pro GA a score of 53, placing it 3 points above V4 Flash (50). Earlier reporting (at time of V4 Flash launch) indicated V4 Pro at 44 — below Flash — but the V4 Pro GA release surpasses Flash on this benchmark.
Grok research provenance
Details compiled via Grok live web research sourcing Reuters, DeepSeek API Docs (pricing page), and DeepSeek GA release notes (news260813).
Official benchmark table from HuggingFace model card
DeepSeek-V4-Flash-0731 benchmark scores on agentic tasks: Terminal Bench 2.1 (82.7), NL2Repo (54.2), Cybergym (76.7), DeepSWE (54.4), Toolathlon-Verified (70.3), Agents Last Exam (25.2), AutomationBench Public (25.1), DSBench-FullStack (68.7), DSBench-Hard (59.6). Outperforms DeepSeek-V4-Pro on all listed tasks despite smaller activated parameter count.
Details
DeepSeek V4 Flash released
DeepSeek released V4 Flash (official model: DeepSeek-V4-Flash-0731), the production release superseding the preview version, with substantially enhanced agentic capabilities and a speculative decoding module attached.
Score: 50 on Artificial Analysis Intelligence Index
V4 Flash scored 50 on the AAII — a 10-point jump over prior V4 Flash and 6 points higher than the larger DeepSeek V4 Pro version.
Ties Gemini 3.6 Flash; one point below GPT-5.6 Luna
Performance matches Google's Gemini 3.6 Flash and falls just one point below GLM-5.2 and GPT-5.6 Luna on the AAII, despite being a small/flash-tier model.
3¢/task vs. 36-59¢ for top rivals
DeepSeek V4 Flash costs 3 cents per task compared to 59 cents for GLM-5.2 and 36 cents for Meta Muse Spark, positioning it as the most cost-efficient option at this performance tier.
12% fewer tokens than prior iteration
The model uses 12% fewer tokens per task than its predecessor, indicating genuine efficiency gains in generation and reasoning — not just a price cut.
Significant GDPval-AA agentic performance jump
A notable improvement on the GDPval-AA metric points to meaningfully better performance on multi-step agentic tasks — increasingly the critical metric for enterprise AI deployments.
DeepSeek V4 Flash model release: benchmark performance (AAII and agentic task benchmarks), pricing, and efficiency metrics vs. key competitors including GLM-5.2 and Opus-4.8.
What This Means
DeepSeek V4 Flash reinforces the Chinese lab's strategy of releasing small models that achieve near-frontier performance at a fraction of Western competitors' cost. At 3 cents per task — a 12x to 20x price advantage over nearest rivals at comparable performance — the model creates real pressure on Google, Meta, and OpenAI to match its efficiency. The combination of better benchmark scores, agentic performance gains, and lower token counts suggests this isn't just cost-cutting but genuine technical progress. For developers and enterprises building AI applications, V4 Flash could become the obvious default for cost-sensitive production workloads, extending DeepSeek's growing influence in the API market.