DeepSeek V4.1 Flash Released: Benchmarks, Pricing, and What Happens to V4 Pro

DeepSeek officially released V4.1 Flash on September 10, 2026. The new model swaps out the old V4 Flash and V4 Flash Vision Exp models entirely, brings native image understanding into the base architecture, and cuts Flash API prices by roughly 11 to 57 percent depending on the token type. V4 Pro was originally slated for retirement on September 14, but DeepSeek reversed that decision and will continue offering it alongside V4.1 Flash. V4.1 Pro is still on the way and will likely replace V4 Pro when it arrives. Independent benchmarks from Artificial Analysis and Vals.ai confirm strong agentic performance at a fraction of the cost of competing models. V4.1 Flash’s Intelligence Index score of 40 also matches the original Qwen3.8 Max, though Alibaba’s updated Qwen3.8 Max 0902 has since pushed to 45, at more than twenty times the cost per task.

Release Status and What Changed

DeepSeek confirmed the launch in its official API changelog, describing V4.1 Flash as the smallest model in a new architecture family. It ships with native multimodal vision, higher throughput, and a design that scales up to larger models in the series. Alongside the weights, DeepSeek also published a full benchmark table and a technical report.

Three practical changes matter most for anyone using the API:

• Use the new model name. The correct name is now deepseek-flash. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work as aliases, but the underlying models are gone. Requests to those names now go to V4.1 Flash at the new rate.
• V4 Pro stays available, for now. DeepSeek originally planned to route all V4 Pro traffic to V4.1 Flash starting September 14, but reversed that decision after user feedback. V4 Pro remains on the API with unchanged pricing. V4.1 Pro is still coming and will likely replace V4 Pro when it ships.
• New Flash pricing is live, effective 04:00 UTC on September 10. Off-peak rates run 50% cheaper than peak, with peak hours at 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. The pricing tables are further down in this article.

DeepSeek ran a two-day beta preview from September 8 under the temporary model ID deepseek-v4.1-flash-expires-on-0910, capped at 20 concurrent requests. That beta window has now closed.

Architecture: Why It Is Faster and Cheaper

This is a ground-up redesign, not a tune-up of V4 Flash. The Hugging Face model card and the technical report lay out the details:

• Scale: 552B backbone parameters plus 196B Engram parameters (a conditional memory module using N-gram hashing), for roughly 748B total. The model activates only 8B per token during prefill and 16B during decode, keeping inference cheap despite the large total count.
• Mixture-of-Experts: 1 shared expert plus 384 routed experts per layer, with 6 routed experts active per token.
• Causal Encoder-Decoder (CED): The 40-layer network splits into a 20-layer causal encoder and a 20-layer decoder. Instead of recomputing KV cache per decoder layer, it projects decoder KV from the encoder’s final hidden states. This roughly halves prefill computation, which matters most when context windows are long and cache misses are frequent.
• Compressed Sparse Attention 2 (CSA2): Each layer runs in one of three modes: Full (compute KV and select attention indices from scratch), Reindex (reuse the shared KV but pick fresh indices), or Reuse (borrow both KV and indices from a prior layer). This cross-layer sharing is the main reason the cache shrinks so dramatically.
• FP4 KV caching: The main KV cache uses MXFP4 quantization, a 4-bit format. Combined with CSA2, this brings the global KV cache down to about 890 bytes per token, roughly one quarter of V4 Flash. That is a 437x reduction compared to DeepSeek V1.
• SWA Bounded Replay: Instead of persisting the full sliding-window attention state to SSD, the model reconstructs it by replaying only the most recent window of tokens. This cuts the persistent cache footprint to about one eighth of V4 Flash with negligible performance loss.
• DSpark speculative decoding: A lightweight 3-block drafter generates 5 draft tokens per forward pass, replacing the MTP module from DeepSeek V3. It is trained separately after pre-training and kept aligned during post-training.
• Native vision: A DeepSeek-ViT encoder trained from scratch handles images at arbitrary resolutions up to about 1344×1344 pixels. Text and image inputs are processed jointly from the start of pre-training, not bolted on afterward.
• Reasoning effort control: Continuously adjustable from 1 to 100 via the API, letting developers trade cost against accuracy. All instruct benchmark figures in this article use maximum effort.
• Training: Pre-trained from scratch on 45T multimodal tokens at 64K sequence length, with 1M token context support. Post-training uses standard SFT plus reinforcement learning, with no algorithmic novelty beyond established practice.
• Open weights: MIT-licensed, with a full technical report covering architecture and evaluation.

Behind the price cuts sits a real engineering change. DeepSeek says V4.1 Flash’s KV cache needs roughly one quarter of the high-bandwidth memory and one eighth of the SSD storage compared to V4 Flash. Compared to DeepSeek V1, the per-token cache is 437 times smaller. The technical report also shows that extending context from 4K to 1M tokens increases decode compute by only about 25%, meaning the model stays fast even at extreme context lengths. For agent workloads that pile up long context windows and hammer cached input tokens, that compression translates directly into lower bills. It is also why DeepSeek can serve more users at these prices without losing money on every request. For a deeper look at why cache hits drive agent costs, see our breakdown of DeepSeek V4 rates and cache economics.

Instruct Benchmarks: V4.1 Flash vs. Frontier Models

DeepSeek published instruct results at maximum reasoning effort with temperature 1.0 and top_p 0.95. Code agent tasks used DeepSeek Harness minimal mode with a 1M context window, visual agent tasks used the Claude Code harness with a 512K window, and remaining agent tasks used official scaffolds. The table below reproduces DeepSeek’s comparison against Opus 5, GPT-5.6 Sol, Kimi K3, GLM-5.3, V4 Pro, and V4 Flash.

The wins are not subtle. V4.1 Flash beats V4 Flash on every shared benchmark, often by wide margins: DeepSWE jumps from 54.4 to 74.2, Terminal Bench 2.1 climbs from 82.7 to 90.6, and CyberGym moves from 76.7 to 88.1. It also edges out V4 Pro across the board, though DeepSeek has since decided to keep V4 Pro available rather than retire it.

Against the frontier competition, V4.1 Flash leads on Terminal Bench 2.1, DeepSWE, CyberGym, HLE with tools, AutomationBench, and Agents Last Exam. It still trails Opus 5 on the newer Terminal Bench 3.0 and 4.0 tests, ProgramBench, and NL2Repo, and GPT-5.6 Sol pulls ahead on SEC-Bench Pro and ExploitGym. These are agent-heavy results, so real-world performance may differ from the numbers.

Base Model Scores and Multimodal Results

Below the instruct layer, the base model tells a more balanced story. DeepSeek compared V4.1 Flash Base against V4 Flash Base and V4 Pro Base under identical internal settings:

• Code and math leads: MMLU-Pro (74.1 vs 73.5 for Pro and 68.3 for Flash), BigCodeBench (60.6 vs 59.2 and 56.8), HumanEval (79.4 vs 76.8 and 69.5), and GSM8K (93.0 vs 92.6 and 90.8).
• Holds steady on Chinese knowledge: C-Eval at 92.1 matches V4 Flash and sits close to V4 Pro’s 93.1.
• Trails V4 Pro on knowledge recall: MultiLoKo (45.5 vs 50.9), SimpleQA Verified (42.3 vs 55.2), SuperGPQA (53.1 vs 53.9), MATH (61.1 vs 64.5), and LongBench-V2 (45.2 vs 51.5).
• New multimodal scores with no V4 equivalent: MMMU-Pro 56.5, CVBench 77.9, DocVQA 95.6, and RefCOCO average 86.0.

The base model picture is more nuanced than the instruct results. V4.1 Flash trades a bit of knowledge recall for stronger code, math, and a native vision pipeline. The gains show up where agent workflows need them most. The native image input continues the direction DeepSeek explored with the experimental V4 Flash Vision model, now built into the base model instead of a separate variant.

How It Compares to Other Close Rivals

Cross-vendor benchmark comparisons always come with a caveat: different testing harnesses, effort settings, and system prompts make one-to-one ranking imprecise. With that in mind, here is how V4.1 Flash stacks up against other close rivals on the benchmarks where figures are available. Independent scores from Artificial Analysis and Vals.ai are covered in the next section.

• GPQA Diamond: V4.1 Flash at 90.9 sits behind GPT-6 Astra (96.1), Gemini 3.8 Flash (95.3), Grok 4.6 (94.9), GPT-5.6 Sol (94.1), Claude Fable 5.1 (93.7), Muse Spark 1.3 (93.5), Kimi K3 (92.9 to 93.5), and V4 Pro (92.4), but ahead of V4 Flash (89.9) and GLM-5.3 (88.1).
• Terminal Bench 2.1: DeepSeek’s own harness puts V4.1 Flash at 90.6, leading Opus 5 (89.1), GPT-5.6 Sol (88.8), and Kimi K3 (88.3). Vals.ai’s independent harness gives it 74.53%, which ranks it second among open-weight models. The gap between 90.6 and 74.53% illustrates how much harness differences matter on this benchmark.
• DeepSWE: V4.1 Flash scores 74.2, essentially tying Opus 5 (74.0) and pulling ahead of GPT-5.6 Sol (73.0), Kimi K3 (67.5), GLM-5.3 (66.9), and both older DeepSeek models.
• Automation and agent work: V4.1 Flash tops its published set on both AutomationBench (54.8) and Agents Last Exam (31.8), beating Opus 5, GPT-5.6 Sol, Kimi K3, and GLM-5.3 on those same tests.
• Newer-generation tests: On Terminal Bench 3.0 and 4.0, Opus 5 still holds the lead (43.3 and 51.8 vs 30.0 and 31.2 for V4.1 Flash). The model is strong, but not a universal frontier leader.

For context on how Flash-class models have compared with Pro-class models on coding work, see Qwen3.8-Flash matching DeepSeek V4 Pro on coding benchmarks. For a direct head-to-head with another open-weight Flash model, see our GLM 5.3 Flash comparison. For pricing context across vendors, see our LLM API pricing comparison. For what Moonshot AI shipped beneath the K3 flagship that same week, see our coverage of Kimi K2.8 Preview and its 1M context window.

Independent Benchmarks: Artificial Analysis and Vals.ai

Independent testing is now available from two major evaluation platforms. These results use each platform’s own harnesses and scoring methods, separate from DeepSeek’s internal tests.

Artificial Analysis tested V4.1 Flash across their suite and published the following scores:

The standout figure is AutomationBench-AA at 68.9%, which Artificial Analysis reports as the highest score of any model from any provider on that benchmark. The AA-Omniscience score of -5 is worth noting: this metric measures knowledge reliability and hallucination, where 0 means equal correct and incorrect answers, and negative means more incorrect than correct. A -5 is close to neutral, suggesting the model is not heavily prone to hallucination on that specific test.

Vals.ai ran V4.1 Flash through their agent-oriented evaluation suite. Key results:

The Vals.ai results tell a cost-efficiency story that DeepSeek’s own benchmarks could not. V4.1 Flash is the new top open-weight model on the Vals Index (57.86%, ranking 15th of 56 overall), narrowly ahead of Kimi K3 (57.81%) at roughly $0.30 per test versus $6.47. It is also the cheapest model in the entire top 15.

On Code Migration, it leads all open-weight models at 45.62% (9th of 58), while the next open-weight competitor, GLM 5.3, scores 44.22% but costs $24.91 per test. It takes first place out of 34 models on SkillsBench (69.80% with skills, 61.66% without). On Vibe Code Bench, it ranks second among open-weight models at 84.74%, trailing Kimi K3 by 0.2 points but costing roughly 40x less ($0.41 vs $17.59 per task) and finishing about 5x faster (15 minutes versus 1.5 hours).

Compared to its predecessor, DeepSeek V4 Flash 0731, V4.1 Flash gains 4.3 points on the Vals Index, with the largest jumps on Legal Research Bench (+11.1 points), Vibe Code Bench (+10.0), and Terminal-Bench 2.1 (+7.5). Despite list prices two to four times higher, it costs less per test on most agentic benchmarks because it finishes tasks in far fewer tokens. It also roughly halves latency on Vibe Code Bench, EMB, Legal Research, and Harvey’s Legal Agent Benchmark.

Two scores stand out as weak spots: SRE Bench at 0.76% and Harvey’s Legal Agent at 6.67%. Both are niche, domain-specific evaluations where the model clearly struggles.

Vals.ai ran all evaluations at temperature 1 with default top-p and high reasoning effort. The model supports a 1M-token context window and up to 384k output tokens. Full details are available on the Artificial Analysis and Vals.ai leaderboards.

Flash Pricing From September 10

The new pricing took effect at 4:00 UTC on September 10 (noon in Beijing, midnight Eastern, 9 p.m. Pacific on September 9). Peak rates apply during 1:00 to 4:00 a.m. and 6:00 to 10:00 a.m. UTC on weekdays. Everything else counts as off-peak.

In U.S. dollars, per 1 million tokens:

Compared to the previous V4 Flash rates, the biggest win is on cached input. In yuan, that drops from 0.05 to 0.02 (a 60% cut), uncached input falls from 1.5 to 1 (about 33%), and output goes from 4.5 to 4 (roughly 11%). In dollar terms, the off-peak rates of $0.007, $0.22, and $0.66 drop to $0.003, $0.15, and $0.60. For agent workloads that hit the cache heavily, the savings add up fast.

The official DeepSeek API pricing documentation is the authoritative source for the live rates.

Pro Routing and V4 Pro Status

The old V4 Flash and V4 Flash Vision Exp models are gone, though their API names still work as aliases pointing to V4.1 Flash. V4 Pro, however, is staying put. DeepSeek originally planned to route all V4 Pro traffic to V4.1 Flash starting September 14, but reversed that decision in response to user demand. V4 Pro continues to operate with unchanged pricing.

When V4.1 Pro eventually ships, it will almost certainly replace V4 Pro. In the meantime, developers on V4 Pro can keep running without changes, or move workloads to V4.1 Flash where the benchmarks and pricing make sense. The price gap is real: V4 Pro off-peak rates of $0.022, $0.66, and $1.98 per million tokens compare to Flash’s $0.003, $0.15, and $0.60, cuts of roughly 86%, 77%, and 70%. But a retrained architecture can shift prompting behavior and tool calling patterns, so test before swapping production traffic.

Frequently Asked Questions

What model name should API users call now?

Use deepseek-flash. That is the official name going forward. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work as aliases and route to V4.1 Flash, but there is no reason to keep using them.

Is V4 Pro being retired?

Not anymore. DeepSeek originally planned to route all V4 Pro traffic to V4.1 Flash starting September 14, but reversed that decision after user feedback. V4 Pro remains available with unchanged pricing. V4.1 Pro is still expected and will likely replace V4 Pro when it arrives.

What are the new Flash prices?

Off-peak: $0.003 per million cached input tokens, $0.15 per million uncached input tokens, and $0.60 per million output tokens. Peak rates are double, during 1:00 to 4:00 a.m. and 6:00 to 10:00 a.m. UTC on weekdays.

Does V4.1 Flash support images?

Yes. V4.1 Flash handles images natively in the base model, so there is no separate Vision variant to worry about. The API vision guides already use deepseek-flash as the example model name.

Where are the full benchmark tables?

They are in the Hugging Face model card, the technical report, and the September 10 API changelog. Keep in mind that all the instruct scores above used maximum reasoning effort. Lower effort settings will score below these numbers.

Are independent benchmarks available yet?

Yes. Artificial Analysis and Vals.ai have both published V4.1 Flash scores on their own harnesses. Artificial Analysis reports 68.9% on AutomationBench-AA (the highest of any model they have tested) and an Intelligence Index of 40. Vals.ai ranks it as the top open-weight model on their Vals Index at 57.86%, ahead of Kimi K3, at roughly one twentieth the cost per test. The full results are in the independent benchmarks section above.

Conclusion

V4.1 Flash is a meaningful step forward, not just a minor revision. It beats both V4 Flash and V4 Pro across every shared benchmark, holds its own against frontier models on several agent and coding tasks, and costs less to run. DeepSeek claims it can handle over 95% of real-world tasks, and the benchmark spread supports that for coding, automation, and security work. Where it falls short, particularly on the newest-generation Terminal Bench tests and niche domain evaluations like SRE Bench and Harvey’s Legal Agent, frontier models still hold the edge.

Independent testing from Artificial Analysis and Vals.ai confirms the pattern. V4.1 Flash is the top open-weight model on the Vals Index and the highest-scoring model on Artificial Analysis’s AutomationBench-AA. The real surprise is the economics: despite higher list prices than V4 Flash, it often costs less per test because it finishes tasks in fewer tokens and with lower latency. For developers weighing the tradeoff between capability and cost, that combination is hard to beat. For workloads that need a higher intelligence score at a still-low price, Step 5 Preview at $0.71 per task is now the cheapest frontier-level alternative.

The bigger story may be what comes next. Between the smaller KV cache, the lower prices, and the eventual arrival of V4.1 Pro, DeepSeek is making its Flash tier the default choice for most workloads. Once V4.1 Pro arrives, it will likely take over from V4 Pro, but for now, V4.1 Flash looks like the sweet spot for cost-conscious agent development while V4 Pro continues to serve workloads that prefer its behavior. Budget-focused developers now have another option in this tier as well: Unisound’s U2-Flash undercuts DeepSeek’s Flash pricing during its launch promotion while claiming competitive scores against V4-generation checkpoints on coding and office benchmarks.