DeepSeek’s V4.1 Flash Is Cheap to Run. Its Benchmark Claims Are Not.

On September 10, 2026, DeepSeek released V4.1 Flash — a 552-billion-parameter mixture-of-experts model with a one-million-token context window and a global KV cache footprint of 890 bytes per token. That last number is the one that matters. It represents roughly one-quarter of what DeepSeek V4-Flash consumed at the same context length, achieved through a new attention architecture called Compressed Sparse Attention 2. The pricing that dropped with the release — $0.15 per million uncached input tokens and $0.60 per million output tokens off-peak — makes the compression’s implications concrete: running long-context agentic workloads at frontier model scale just got substantially cheaper.

The benchmark table DeepSeek chose to headline the launch tells a different story. It leads with a 74.2% score on DeepSWE v1.1, against Claude Opus 5’s 74.0% — a 0.2-point margin the company presents as evidence of matching frontier performance. That is not wrong exactly. But a margin of 0.2 points on a software engineering benchmark, against a model priced at roughly 20× the output cost, is not the same thing as being a general-purpose peer. The harder benchmarks tell you more: on Terminal Bench 3.0, V4.1 Flash scores 30.0 against Opus 5’s 43.3. On Terminal Bench 4.0, it’s 31.2 against 51.8. On HLE, a knowledge benchmark, it’s 36.8 against 56.3. These numbers appear in DeepSeek’s own technical report, which includes the acknowledgment that on knowledge-heavy evaluations, both Western flagship models lead by a clear margin. The press kit chose not to open with that.

What CSA2 Actually Does

The compression architecture is worth understanding on its own terms. Compressed Sparse Attention 2 assigns each attention layer one of three operating modes — Full, Reindex, or Reuse — determining how that layer handles the key-value cache representing prior context.

A Full-mode layer builds a fresh global KV cache and performs its own selection of which earlier positions matter for the current query, using sparse attention to pick the top 512 entries from a candidate pool of up to 16,384 positions. This is the most expensive mode and the most accurate: it looks at everything it can and selects the relevant slice. A Reindex-mode layer shares an existing cache but makes its own independent selection from it, paying for the selection cost without paying for cache construction. A Reuse-mode layer shares both the cache and the previous layer’s selections entirely — adding no memory overhead beyond what the layer above already computed.

Layering these three modes across a 552B model’s attention stack drives the KV cache from V4-Flash’s roughly 3,560 bytes per token down to 890. Weight quantization compounds the savings: MoE expert weights are stored in FP4, with attention and normalization layers staying in FP8. Combined, the global KV cache for a one-million-token context works out to roughly 3.56 gigabytes per concurrent session under V4-Flash, and about 890 megabytes under V4.1 Flash. On a GPU fleet where KV cache competes directly with active request buffers and decode batch sizes, the 4× reduction maps directly to throughput: four long-context sessions where previously one fit, without adding hardware.

The persistent KV cache — the portion offloaded to SSD or host memory for very long sessions — compresses further, to roughly one-eighth of V4-Flash’s footprint. DeepSeek’s “8x smaller” headline refers specifically to this tier. Whether your workload sees that improvement depends on whether your inference architecture offloads to slower storage at all. For most sub-million-context deployments, the relevant figure is the 4× reduction in active HBM consumption, not the 8× figure for persistent storage.

The Reasoning Dial

V4.1 Flash ships with what DeepSeek calls controllable reasoning effort, a parameter from 1 to 100. This is not a simple token budget. The model doesn’t use different weights at different settings; it uses the same weights throughout, with a different operating point baked in during training through reward weighting. Lower effort settings applied stronger penalties during training for generating additional reasoning tokens; higher settings applied weaker penalties. The result is a continuous tradeoff between token generation and answer quality, rather than the binary think-or-don’t-think modes that have dominated the lab-to-market cycle so far.

At Max (100), V4.1 Flash hits the benchmark numbers DeepSeek published — GPQA Diamond at 90.9%, DeepSWE at 74.2%, Terminal Bench 2.1 at 90.6%. The realistic comparison point for production deployments is more likely effort 50 or 75, where latency comes down substantially. DeepSeek has not published benchmark numbers at lower effort settings, which makes evaluation harder for anyone building a cost-calibrated pipeline. What the design offers is a way to allocate inference compute proportionally: route hard queries to effort 100, route simpler classification or retrieval tasks to effort 25, and let the model do less work when less work is sufficient. Whether that calibration holds under real workloads is something builders will have to measure themselves — the technical report doesn’t benchmark it.

Who This Is Actually For

The case for V4.1 Flash isn’t the DeepSWE number. It’s the context window price. At $0.60 per million output tokens off-peak, serving a 1M-token context is cheaper than it has ever been at frontier model tier. Claude Opus 5 runs at approximately $25 per million output tokens. Most long-context agentic workloads don’t require Opus 5’s breadth on knowledge tasks; they require reliable tool use, code generation, and context tracking across many documents or a long session. That’s the workload profile where V4.1 Flash is genuinely competitive: Automation-Bench at 54.8% versus Opus 5’s 50.3%, CyberGym at 88.1% versus GPT 5.6 Sol’s 84.5%, DeepSWE at 74.2%. The cost difference — 20× on output tokens — is large enough that for pipelines running thousands of long-context requests daily, it determines feasibility, not preference.

The compression tradeoff is real and documented. Sparse attention means the model doesn’t examine every token in context — it selects from a candidate pool and approximates the rest. SWA Bounded Replay, the mechanism for reconstructing context from the compressed cache, produces results that can vary depending on where a cache hit occurs. In practice, very long sessions with critical information scattered at unpredictable positions may behave differently than they would under full attention. For retrieval tasks with well-organized context, this is unlikely to matter. For synthesis tasks that require integrating evidence spread across thousands of positions, it is worth testing explicitly before routing production traffic.

Starting September 14, DeepSeek will route all V4 Pro API traffic to V4.1 Flash at Flash prices — a migration that existing Pro API users don’t have to opt into. That’s a substantial cost reduction for anyone currently paying Pro rates. It’s also a change in model behavior. V4 Pro and V4.1 Flash are not the same model with the same characteristics. Anyone using V4 Pro for tasks where knowledge depth or harder terminal-class reasoning matters should test V4.1 Flash against those workloads before the migration completes rather than after.

The architecture advance in V4.1 Flash — CSA2 with three layer modes, 890 bytes per token at 1M context, 70-86% price cuts versus V4 Pro — is a real infrastructure achievement. The benchmark table DeepSeek headlined it with is less useful: a 0.2-point DeepSWE lead is a tie, not a win, and it sits next to a 13-point Terminal Bench 3.0 deficit and a 20-point Terminal Bench 4.0 deficit that belong in the same press release. The model is genuinely strong for specific agentic workloads at a price point that changes what’s feasible to build. That’s the story. The comparison benchmarks are a marketing decision, not an engineering one.

AI-generated editorial illustration · TemperatureZero · September 10, 2026