with a 552-billion-parameter mixture-of-experts backbone, native vision, a 1-million-token context window and an architecture built to make repeatedly reading large contexts cheaper.
For developers evaluating the model for coding agents and other long-running workflows, however, the headline API rate only tells part of the story.
During off-peak hours, DeepSeek prices V4.1-Flash at $0.003 per million input tokens on a cache hit, $0.15 per million on a cache miss and $0.60 per million output tokens. Peak rates are double those figures.
That distinction matters because an agent repeatedly working against the same repository, tool definitions, system instructions or conversation history may reread far more cached context than fresh context. DeepSeek itself says cache-hit charges can account for a significant portion of agent costs.
DeepSeek’s established peak window runs Monday through Friday from 01:00–04:00 UTC and 06:00–10:00 UTC; all other hours are off-peak. For schedulable agent jobs, then, the time a workload runs is itself another cost-control lever.
That makes V4.1-Flash a particularly useful test case for a problem VentureBeat’s own research suggests many enterprises still cannot see clearly.
In VentureBeat Pulse Research’s July 2026 survey of 170 enterprises with more than 100 employees, only 47% said they rigorously track AI compute cost and ROI — meaning 53% do not. Meanwhile, 12% said they have not yet addressed inference-memory limits such as KV-cache capacity, and another 7% were not aware of the constraint. Even among organizations operating AI in production at scale, rigorous cost tracking reached only 56%.
The implication for builders is straightforward: comparing models on uncached input price alone can obscure one of the fastest-growing components of agent economics.
DeepSeek’s off-peak cached-input rate is unusually low relative to the frontier APIs against which it benchmarks V4.1-Flash.
OpenAI lists GPT-5.6 Sol at $4 per million regular input tokens, $0.40 for cached input and $20 for output. Anthropic charges $5 for standard input, $0.50 for Claude Opus 5 cache hits and $25 for output. Moonshot AI’s Kimi K3 is $3 for cache-miss input, $0.30 for cache-hit input and $15 for output.
DeepSeek’s comparison is not perfectly apples-to-apples because its lowest rates require off-peak usage. Even at peak, however, V4.1-Flash rises only to $0.006 per million cached input tokens, $0.30 for uncached input and $1.20 for output.
For a simplified builder calculation, consider an agent that retains a 500,000-token reusable prefix and hits that cache across 100 requests. That represents 50 million cached input tokens. Ignoring cache-write charges, new uncached context and output, those cache reads would cost about $0.15 on V4.1-Flash off-peak, compared with $15 on Kimi K3, $20 on GPT-5.6 Sol and $25 on Claude Opus 5 at their published cache-read rates.
Real workloads will not achieve perfect cache reuse, and total cost depends on output length, reasoning tokens, retries, tool calls and task success rates. But that is precisely why developers evaluating agent workloads should measure cache-hit ratio and cost per completed task, rather than multiplying total prompt tokens by a model’s headline input price.
DeepSeek-V4.1-Flash now sits near the absolute low end of the global paid API market while also looking more capable than its pricing tier suggests.
, it ranks behind only Meta’s Contributor-tier Muse Spark and MiMo-V2.5 Flash in the current chart; even at its
, it still ties with MiniMax-M3 and LongCat’s promotional pricing, far below the mid-tier cluster around
and dramatically below premium frontier models such as Claude Fable/Mythos 5.1, GPT-6 Astra and GPT-5.6 Sol Fast mode.
The strategic point is that DeepSeek is using V4.1-Flash not just as a cheap “flash” model, but as the new center of gravity for its API: its documentation says legacy V4 Flash requests are now served by V4.1-Flash, and that V4.1 Flash has “comprehensively surpassed V4 Pro in performance, cost, speed, and total time,” with V4 Pro set to route to V4.1 Flash after Sept. 14, 2026 until a future V4.1 Pro arrives.
VentureBeat’s Pulse data suggests that instrumentation has not caught up. Just 31% of surveyed enterprises named cost per million tokens as a primary infrastructure success metric, while uptime and reliability led at 51% and developer productivity and deployment speed reached 39%. Only 22% selected total cost of ownership as a top buying criterion.
DeepSeek’s technical report also reveals what sits underneath those cache-hit economics. The company says global KV remains in its persistent cache with a guaranteed lifetime of at least 72 hours. Short-lived sliding-window-attention KV, by contrast, is kept in a distributed memory pool provisioned from 10% of host DRAM and given a minute-scale lifetime. When the long-lived global cache survives but the SWA state does not, V4.1-Flash reconstructs…
The architectural distinction is also important. V4.1-Flash uses what DeepSeek calls a
architecture, splitting its 40 Transformer layers into a 20-layer causal encoder and a 20-layer decoder. The design activates
That corrects an easy shorthand to get wrong: V4.1-Flash is not an “8B-active” model across the entire inference cycle. It activates 8B during prefill and 16B during decode.
The model combines that architecture with Compressed Sparse Attention 2, hierarchical sparse indexing and FP4 KV caching. DeepSeek says those techniques bring the global KV cache down to 890 bytes per token, approximately one-quarter the size of V4-Flash’s, while its persistent cache storage requirement falls to roughly one-eighth.
For input-heavy agent loops, those are potentially meaningful serving improvements. A coding agent that continually rereads a repository is exactly the kind of workload where prefill and KV-cache efficiency can dominate.
But “Flash” no longer means small. DeepSeek’s previous V4-Flash used a 284B backbone with 13B active parameters. V4.1-Flash increases the backbone to 552B, a roughly 94% jump. Prefill activation falls from 13B to 8B, but generation activation actually rises from 13B to 16B.
And 552B describes only the backbone: DeepSeek’s technical report separately lists 196B parameters in its sparsely accessed Engram conditional-memory modules.
That tradeoff has prompted skepticism among developers discussing the release. One commenter in discussion around the
whether a 552B model is “not really flash anymore,” noting that its much larger total footprint makes local deployment considerably more demanding even if its sparse activation and cache architecture improve serving efficiency.
The criticism is useful, but it does not negate DeepSeek’s efficiency claim so much as define it more precisely.
V4.1-Flash is not more compact than V4-Flash in total weights. It is designed to use less compute while ingesting tokens and substantially less memory for retaining long contexts. For builders buying inference from DeepSeek, that can translate into lower API pricing. For teams intending to self-host the MIT-licensed weights, the jump from 284B to 552B makes the hardware equation considerably less friendly.
The technical report provides an important counterweight to the “it just got bigger” critique. DeepSeek says CED effectively halves prefill computation for sufficiently long sequences, while single-token decode FLOPs increase by only about 25% as context length expands 256-fold from 4K to 1 million tokens. In other words, V4.1-Flash is larger in weights, but DeepSeek has engineered the amount of work associated with a growing context to scale muc…
There is a corresponding caveat. DeepSeek explicitly says the new architecture creates robustness boundaries it has not fully characterized. The company says potential CSA2 sparse-selection errors and the approximate state reconstruction used by SWA Bounded Replay could cause capability degradation in untested edge cases, particularly around sparse retrieval over very long contexts and cache-resumption boundaries. DeepSeek says it has not observe…
DeepSeek’s own release even tells organizations interested in large-scale deployment to contact the company if they have resources on the order of 2,000 GPUs plus a storage cluster.
The benchmarks are promising, but workload shape matters more
It also reports V4.1-Flash at 88.1 on CyberGym and 54.8 on AutomationBench.
Those are DeepSeek-run evaluations, and the results are not uniformly dominant. Opus 5 leads V4.1-Flash 43.3 to 30.0 on Terminal-Bench 3.0 and 51.8 to 31.2 on Terminal-Bench 4.0, while GPT-5.6 Sol leads on GPQA Diamond and SEC-Bench Pro in DeepSeek’s table.
There is another cost variable hidden inside those headline scores: reasoning effort. DeepSeek ran the comparison table at its maximum effort setting of 100. In the company’s own tests, increasing effort from 25 to 100 raises DeepSWE v1.1 from 66.0% to 74.2% and Terminal-Bench 2.1 from 82.4% to 90.6%, but consumes roughly 2.5 times as many output tokens. DeepSeek says the gains are front-loaded: effort levels between 60 and 80 recover most of max…
That creates another practical routing decision for builders. DeepSeek exposes public API presets of low, high and max, corresponding to reasoning-effort values of 50, 75 and 100. A team optimizing for cost per successful task may therefore find that the model’s best operating point is not the max-effort configuration behind its leaderboard numbers.
Early third-party evidence supplied to VentureBeat points toward the same price-performance thesis rather than a clean intelligence lead.
OpenDesign said V4.1-Flash reached 98% of GPT-6 Astra’s quality score at 1.4% of its cost on the company’s set of everyday design requests. That is a narrow third-party workload, not a general model evaluation, but it reinforces the case for testing successful work per dollar rather than benchmark position alone.
There is another production caveat builders should notice. DeepSeek has retired V4-Flash and V4-Flash-Vision-Exp and temporarily routes their existing identifiers to V4.1-Flash. It also says that beginning Sept. 14, calls to
will be routed to V4.1-Flash until V4.1-Pro arrives.
separate line of criticism from developers on Hacker News
, who argued that changing the underlying model behind an existing production identifier can invalidate regression testing even when the replacement is nominally better or cheaper. For teams with tightly tuned prompts and agent behavior, model migration remains something to test rather than assume.
A price-performance release at a consequential moment
The launch also arrives as DeepSeek’s commercial ambitions appear to be widening.
reported this week that the company has tapped CITIC Securities as it prepares for a potential listing on Shanghai’s STAR Market. Reuters also reported that a current fundraising could value DeepSeek at as much as 500 billion yuan, or roughly $75 billion.
That context raises the stakes around a release whose core proposition is less about winning every benchmark than about reducing the cost of running capable models at scale.
For developers, V4.1-Flash therefore poses a more specific question than whether DeepSeek has produced another cheap frontier challenger. The useful test is whether a workload is input-heavy, repetitive and cacheable enough to exploit its 8B-parameter prefill path and much smaller KV cache — and whether those savings outweigh a model footprint that has nearly doubled and a decode path that now activates 16B parameters.
The technical report adds one more variable to that equation: whether the workload benefits enough from maximum reasoning effort to justify roughly 2.5× the output-token consumption, or whether a mid-tier effort setting delivers a better cost-per-task result.
For agent builders, that is the benchmark that matters: not just intelligence per token, but completed work per cached token, per second and per dollar.