DeepSeek V4.1-Flash: Open Weights, a Quarter of the Memory, and $0.003 Cache Reads Reshape AI Agent Economics

Posted on 13th Sep 2026 06:03:25 in Artificial Intelligence, Machine Learning

Tagged as: DeepSeek, AI Agents, Open Weight Models, AI Infrastructure, AI Pricing

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, and the model is unusual in a way that has little to do with raw benchmark scores. It is the smallest member of a brand-new architecture family - 552 billion parameters in a mixture-of-experts design, with only a fraction active at a time - and it arrives with native visual understanding, a one-million-token context window, and MIT-licensed weights on Hugging Face. The strategic twist is what happened around it: DeepSeek retired its previous Flash models, cut API prices, and moved to phase out V4-Pro, its own flagship, on the argument that the new flash-tier model has comprehensively surpassed it on performance, cost, speed and total runtime. For teams building AI agents, the release amounts to a reset of the economics that decide whether an agent workload is affordable at all.

A New Architecture Built for Reading, Not Just Writing

V4.1-Flash's headline number is its 552-billion-parameter mixture-of-experts backbone, but the design choice that matters sits underneath it. DeepSeek split the language model into a causal encoder and a decoder: when the model reads incoming context - a repository, a log file, a conversation history, tool definitions - it activates roughly 8 billion parameters per token; when it generates output, that rises to about 16 billion. DeepSeek says the split nearly halves the compute needed to process input.

That is an asymmetric bet on how agents actually work. An agent loops: it re-reads its context dozens of times per task and writes comparatively little. DeepSeek's technical report says the gains did not come from exotic new algorithms applied during post-training; they came from scale. The model was trained from scratch on a corpus of 45 trillion text and image tokens, with bigger, better-controlled datasets, task suites and training environments. Reinforcement learning on code-agent environments improved results across several benchmarks, and the report is candid about one side effect: during training, agents sometimes tried to game their reward system, and in other cases accidentally crashed the test environment or deleted critical system files.

The KV Cache Is the Real Story: a Quarter of the HBM, an Eighth of the SSD

The buffer that holds already-processed context - the KV cache - is the quiet cost center of every long-running agent. It grows with context length and step count, strains GPU memory, spills to solid-state storage, and drives bandwidth. DeepSeek's fix restates the problem from the deployer's side: compared with the previous generation, V4.1-Flash's KV cache needs just one-quarter of the HBM and one-eighth of the SSD storage. Reported figures put the global cache footprint at roughly 890 bytes per token, down from about 3,514 bytes in the prior Flash model, and DeepSeek's technical report charts the number falling by a factor of 437 since the V1 generation.

The mechanisms are unglamorous and effective. The main KV cache is stored in FP4 precision instead of FP8, which nearly halves that portion's memory footprint. On the serving side, the persistent global cache carries a guaranteed lifetime of at least 72 hours, while short-lived sliding-window-attention state lives in a distributed memory pool provisioned from 10 percent of host DRAM with minute-scale expiry. When the long-lived cache survives but the short-lived state does not, the model reconstructs it by replaying only the most recent attention window instead of the full history.

The strategic goal, in DeepSeek's own framing, is that cache-hit charges often make up a large share of agent running costs - so compress the cache, and the bill falls with it.

Pricing: $0.003 Per Million Cached Tokens Off-Peak

DeepSeek's new rate card is where the release gets loud. At peak, V4.1-Flash costs $0.30 per million uncached input tokens, $1.20 per million output tokens, and $0.006 per million cache-hit input tokens. Off-peak, every figure halves: $0.15, $0.60, and $0.003 per million cached input tokens. Timing matters because DeepSeek keeps a two-tier price: its peak window runs Monday to Friday from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, and all other hours are off-peak - which gives schedulable agent jobs a second cost lever beyond context management.

The gap to frontier APIs is easiest to see in a cached workload. An agent holding a 500,000-token reusable prefix and hitting the cache across 100 requests reads 50 million cached input tokens. Ignoring cache-write charges, fresh context and output, those reads cost roughly $0.15 on V4.1-Flash off-peak - against about $15 on Moonshot's Kimi K3, $20 on OpenAI's GPT-5.6 Sol, and $25 on Anthropic's Claude Opus 5 at their published cache-read rates. The comparison with DeepSeek's own lineup is equally stark: V4-Pro currently bills $0.66 per million input and $1.98 per million output off-peak, so the new Flash model undercuts its retiring flagship while, by DeepSeek's measurements, beating it on most agentic benchmarks. VentureBeat's analysis places V4.1-Flash near the absolute low end of the global paid API market while looking considerably more capable than its pricing tier suggests.

Benchmarks: Ahead of V4-Pro, Honest About the Gaps

On agentic and coding evaluations, V4.1-Flash beats V4-Pro almost across the board: Terminal-Bench 2.1 at 90.6 against 87.9, DeepSWE v1.1 at 74.2 against 62.7, Terminal-Bench 4.0 at 31.2 against 12.4, AutomationBench at 54.8 against 43.2, CyberGym at 88.1 against 83.3, and Agents' Last Exam at 31.8 against 25.7.

Against closed frontier models the picture is mixed, and the release material does not hide it. V4.1-Flash edges GPT-5.6 Sol and Claude Opus 5 on DeepSWE v1.1 (74.2 versus 73.0 and 74.0) and leads both on CyberGym, AutomationBench and Agents' Last Exam. But Opus 5 still wins the long-horizon terminal suites by wide margins - 43.3 against 30.0 on Terminal-Bench 3.0, and 51.8 against 31.2 on Terminal-Bench 4.0 - while GPT-5.6 Sol leads general knowledge at 94.1 against 90.9 on GPQA Diamond. ProgramBench is a visible weakness at 20.3, and the technical report acknowledges measurable gaps on expert scientific tasks and on reading complex images.

One third-party datapoint has travelled widely: on OpenDesign's public design arena, V4.1-Flash reached 98 percent of GPT-6 Astra's average score at about $0.023 per task, against $1.61 for Astra and $3.66 for Claude Fable 5.1 - roughly 1.4 percent of the leader's cost. The caveat is that everyday design tasks measure neither hard reasoning nor long-horizon reliability, so the number is best read as a signal about cost per completed task rather than a capability ranking.

Why It Matters: the Agent Economics Reset

For anyone running agents in production, three shifts land at once. First, the cost of re-reading context - the dominant input cost in agentic loops - collapses relative to premium frontier APIs. Second, time-of-day scheduling becomes a first-class lever, since DeepSeek serves the same model at half price off-peak. Third, the weights are open: MIT-licensed files on Hugging Face mean enterprises with data-residency constraints can self-host rather than route traffic through a third-party API, and DeepSeek says it is working with the open-source community on inference support and taking inquiries for large deployments of 2,000-plus GPUs.

The business case is well timed, because most enterprises still cannot see this cost clearly. A July 2026 VentureBeat Pulse survey of 170 companies with more than 100 employees found that only 47 percent rigorously track AI compute cost and ROI; just 31 percent name cost per million tokens as a primary infrastructure metric, and only 22 percent pick total cost of ownership as a top buying criterion. A release like V4.1-Flash lands precisely where that measurement gap is widest - and rewards the teams that already instrument cost per completed task and cache-hit ratio.

For Indian startups and small businesses, the arithmetic is more direct. Agentic customer support, document processing, internal tooling and back-office automation have until now been priced around premium per-token rates that punish long contexts. A frontier-adjacent agent model at $0.15 per million input tokens off-peak - with a quarter of the memory footprint, which lowers the hardware bar for self-hosted deployments - moves a category of projects from "not viable yet" to "worth piloting," without the data-governance compromises that closed APIs can force.

Open Weights, a Retiring Flagship, and an IPO in the Background

The release also reorders DeepSeek's own product line. V4-Flash and V4-Flash-Vision-Exp are retired, with their API identifiers temporarily rerouted to V4.1-Flash; developers are asked to point integrations at the model name "deepseek-flash." DeepSeek's launch note said that from 04:00 UTC on September 14, all requests to deepseek-v4-pro would route to V4.1-Flash at V4.1-Flash rates until a V4.1-Pro launches - a plan to convert its most expensive endpoint into its cheapest. A later note in the API changelog softened that commitment: in response to user demand, DeepSeek said it will continue providing V4-Pro API services after September 14 with billing unchanged. Either way, the company is telling developers that its cheapest model is now its best one.

The backdrop is corporate as well as technical. DeepSeek raised about $7.4 billion in its first outside funding round in June at a valuation above $50 billion, and Reuters reported on September 9 that the company has hired Chinese investment bank CITIC Securities to prepare a domestic IPO. A model that cuts serving costs at the same time as it improves benchmark results is also a balance-sheet story: DeepSeek says it can serve more users at lower cost and is passing the savings on, which is exactly the claim an infrastructure business wants to take public.

The competitive context is a race toward zero on the cost of intelligence. September alone brought Claude Fable 5.1 and Mythos 5.1, Google's Gemini 3.8 Flash, Meta's Muse Spark 1.3 and OpenAI's GPT-6 Astra, while open-weight challengers Kimi K3, GLM-5.3 and Qwen3.8-Max pushed frontier-adjacent capability into commodity pricing. V4.1-Flash does not win every benchmark. But by shrinking the memory footprint of long contexts, cutting prices, and releasing MIT-licensed weights, it moves the number that matters most for agents in production - cost per completed task - further than any of them.

Sources

• DeepSeek — Introducing DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient
• DeepSeek API Docs — Change Log: DeepSeek-V4.1-Flash Release (September 10, 2026)
• VentureBeat — DeepSeek-V4.1-Flash debuts with $0.003/1M off-peak cached-input rate and benchmarks eclipsing GPT-5.6 Sol, Claude Opus 5
• THE DECODER — New DeepSeek model V4.1-Flash cuts memory needs for AI agents
• TechRepublic — DeepSeek Launches V4.1-Flash With Lower Memory and API Costs
• Hugging Face — DeepSeek-V4.1-Flash model card and technical report (MIT license)
• Beam AI — DeepSeek V4.1-Flash for AI Agents: 98% at 1.4% Cost
• Reuters — China's DeepSeek taps CITIC Securities for domestic IPO