DeepSeek released V4.1-Flash last night at a price that makes the rest of the frontier API market look mispriced: $0.003 per million cached input tokens off-peak, against $0.40 for GPT-5.6 Sol and $0.50 for Claude Opus 5. The model is a 552-billion-parameter mixture of experts with built-in image support and a one-million-token context window, and DeepSeek's own table puts it at 74.2 on DeepSWE v1.1, just ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. For anyone building agents, the first number matters more than the last.
### 💵 The price lives on the cached path
Off-peak, DeepSeek charges $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. During peak hours all three double. The peak window runs Monday through Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC. Everything else counts as off-peak.
That two-tier structure does more work than the headline number suggests. The cheapest rate DeepSeek advertises applies all weekend and on weekdays outside two blocks that convert to 09:00–12:00 and 14:00–18:00 in Beijing — the company's own working day. If your jobs run on a schedule, execution time has just become a pricing lever. If they run when your users are awake in Asia, it has not.
The gap matters most for the workload DeepSeek is plainly aiming at: an agent that reads the same repository, the same tool descriptions, the same system prompt and the same conversation history again and again, re-reading far more stored context than new. DeepSeek says cache-hit charges can make up a significant share of agent spend.
Take the simplified case. An agent caches a reusable 500,000-token prefix and hits it across 100 requests: 50 million cached input tokens. Setting aside cache writes, new uncached context and output, those reads cost roughly $0.15 on V4.1-Flash off-peak. At published cache-read rates the same reads cost $15 on Kimi K3, $20 on GPT-5.6 Sol and $25 on Claude Opus 5.
Two caveats belong with that arithmetic. It excludes cache writes, and the three-line price list above does not cover them. And no real workload achieves perfect reuse: total cost depends on response length, reasoning tokens, reruns, tool calls and how often tasks actually complete. The number worth measuring is cache-hit rate and cost per completed task, not prompt tokens multiplied by an advertised input price.
Even at peak — $0.006 cached in, $0.30 uncached in, $1.20 out — V4.1-Flash sits near the floor of the paid API market. At $0.15 in and $0.60 out, $0.75 per million tokens combined off-peak, only Meta's Contributor-tier Muse Spark and MiMo-V2.5 Flash come in lower on the current table. The peak combined rate of $1.50 draws level with MiniMax-M3 and LongCat's promotional pricing. The middle of the market sits around $4.50 to $8; the premium tier — Claude Fable/Mythos 5.1, GPT-6 Astra, GPT-5.6 Sol in fast mode — sits far above it.
Most buyers are not equipped to notice any of this. In a July VentureBeat Pulse Research survey of 170 companies with more than 100 employees, only 47% said they rigorously track AI compute cost and the return on it, which leaves 53% who do not. Another 12% had not yet looked at inference memory limits, including KV cache size, and 7% did not know the constraint existed. Even among organizations already running AI at production scale, only 56% track spend rigorously. Just 31% named cost per million tokens a top infrastructure success metric, behind availability and reliability at 51% and developer productivity and deployment speed at 39%, and only 22% put total cost of ownership among their main selection criteria.
DeepSeek has therefore priced hard against a line item most of the market admits it does not measure. That is a gift to the half that does.
### 🧩 "Flash" now means cheap to serve, not small
V4.1-Flash uses what DeepSeek calls a Causal Encoder-Decoder architecture: 40 Transformer layers split into a 20-layer causal encoder and a 20-layer decoder. Reading input — the prefill stage — activates 8 billion parameters per token. Generating a response activates 16 billion. It is not an 8-billion-active model end to end, and describing it that way would flatter it.
On top of that sit Compressed Sparse Attention 2, hierarchical sparse indexing and FP4 KV caching. DeepSeek says these shrink the global KV cache to 890 bytes per token, about a quarter of V4-Flash's, and cut persistent cache storage requirements to roughly an eighth. The technical report adds that the global KV cache is held in persistent storage for at least 72 hours, while the short-lived sliding-window attention cache lives in a distributed memory pool given 10% of server RAM and survives for minutes. When the long-lived cache is still there and the sliding-window state is not, V4.1-Flash rebuilds only the last attention window instead of reprocessing the whole history.
None of which makes the model smaller. V4-Flash had a 284-billion-parameter base and activated 13 billion. V4.1-Flash's base is 552 billion, up about 94%. Prefill activation fell from 13 billion to 8 billion, but decode activation rose from 13 billion to 16 billion. And 552 billion is only the base: the technical report lists a further 196 billion parameters in Engram conditional memory modules, accessed sparsely.
A commenter on the Hacker News launch thread doubted that a 552-billion-parameter model can still be called fast, arguing that the much larger total size makes local deployment considerably harder even if sparse activation and the cache design improve serving efficiency. That sharpens DeepSeek's claim rather than refuting it. The efficiency is real and it is an efficiency of serving, not of weights: DeepSeek says CED roughly halves prefill compute on sufficiently long sequences, while decode FLOPs per token rise only about 25% as context grows 256-fold, from 4,000 tokens to one million. For customers buying inference from DeepSeek, that translates into the price list. For teams planning to self-host the MIT-licensed weights, going from 284 billion to 552 billion parameters translates into a hardware bill. The release advises organizations interested in large-scale deployment to get in touch if they have around 2,000 GPUs and a storage cluster, which is a fairly direct statement about who the weights are for.
DeepSeek is also candid that the architecture creates reliability boundaries it has not fully characterized. Sparse selection errors in CSA2 and approximate state recovery in the SWA Bounded Replay mechanism could degrade the model in untested edge cases, particularly sparse retrieval over very long contexts and recovery after a cache restart. The company says it saw no systematic degradation in its own testing and plans further stress tests. Both of those failure modes sit on exactly the long-context, cache-heavy path the pricing is designed to sell.
### 📊 The benchmark lead is thinner than the table
The 74.2 on DeepSWE v1.1 is 0.2 points above Claude Opus 5 and 1.2 above GPT-5.6 Sol, and DeepSeek ran the evaluations itself. Treat that as a tie. DeepSeek also reports 88.1 on CyberGym and 54.8 on AutomationBench. Where the gaps are wide, they run the other way: Claude Opus 5 leads Terminal-Bench 3.0 by 43.3 to 30.0 and Terminal-Bench 4.0 by 51.8 to 31.2, and on DeepSeek's own table GPT-5.6 Sol leads GPQA Diamond and SEC-Bench Pro.
There is a cost hiding inside the headline scores. DeepSeek ran the comparison table at reasoning effort 100, the maximum. In its own tests, raising effort from 25 to 100 lifted DeepSWE v1.1 from 66.0% to 74.2% and Terminal-Bench 2.1 from 82.4% to 90.6%, while the model burned roughly 2.5 times more output tokens. DeepSeek says most of the gain arrives early: effort between 60 and 80 delivers the bulk of maximum-mode accuracy at less than half the token budget, and the final step to 100 stretches agent trajectories by 1.6 to 1.8 times for comparatively little. The public API exposes low, high and max, mapping to effort 50, 75 and 100. A team optimizing cost per completed task will probably find the answer is not max — which is to say the leaderboard configuration is the one you should least want to run in production.
Early third-party data points the same way. OpenDesign told VentureBeat that V4.1-Flash reached 98% of GPT-6 Astra's quality score at 1.4% of its cost on its own set of everyday design prompts. That is a narrow check on one workload, not a general evaluation, but it describes value for money rather than superior intelligence, and value for money is what this release is actually selling.
Here is the part of the launch I would argue with. DeepSeek has retired V4-Flash and V4-Flash-Vision-Exp and temporarily pointed those model IDs at V4.1-Flash. From 14 September, requests to deepseek-v4-pro route to V4.1-Flash as well, until a future V4.1-Pro arrives. The documentation states that requests to the old V4 Flash are now served by V4.1-Flash, and DeepSeek claims V4.1 Flash "fully surpassed V4 Pro in performance, cost, speed and overall time." Hacker News developers pushed back on the policy specifically, on the grounds that substituting a model behind an existing production identifier can invalidate regression tests even when the replacement is genuinely better or cheaper. They are right, and the claim and the complaint are not in conflict. "Better on our benchmarks" is not the same property as "behaves the same," and a migration you did not schedule is still a migration. A vendor confident enough to route a Pro tier onto a Flash model is making a statement about its lineup; teams with carefully tuned prompts and agent behaviour are the ones who pay for testing it.
### 🏦 A cheap model and an expensive fundraise
The launch lands while DeepSeek expands commercially. Reuters reported this week that the company has hired CITIC Securities to prepare for a possible listing on Shanghai's STAR Market, and that its current funding round could value it at up to 500 billion yuan, or about $75 billion.
Read against that, V4.1-Flash looks less like an attempt to win every benchmark than like a demonstration that DeepSeek can serve a capable model at scale for very little — the operating story a STAR Market listing would need to tell. The model is also being installed as the hub of the API lineup rather than added beside it, which is what a company does when it wants a single serving path to defend.
The question for developers is narrower than whether DeepSeek shipped another cheap frontier competitor. It is whether a given workload is input-heavy, repetitive and cacheable enough to exploit the 8-billion-parameter prefill path and the much smaller KV cache, and whether that saving survives the nearly doubled model size and the 16-billion-parameter decode path. Then whether maximum reasoning effort earns its 2.5x output tokens, or whether the middle setting delivers better cost per completed task.
For agent builders the useful metric was never intelligence per token. It is completed work per cached token, per second, per dollar — and DeepSeek has just set the reference price for the first of those at three tenths of a cent, on the condition that you run when its own servers are quiet.