DeepSeek's Cheaper V4.1-Flash Beats Its Own Pro Model on Agent Tasks

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, and the most consequential line in the announcement was not about the model's size. It was the claim that the company's cheapest current model now outperforms its own flagship on the work that matters most to agent builders. That claim is unusual enough to check closely, and the early independent evidence largely backs it up, with some important exceptions.

Related reading: DeepSeek's V4 Pro Is Its Best Model Yet, and Its Priciest by Far · Google's Gemini 3.8 Live Reasons Out Loud While It Talks.

The new model is a 552-billion-parameter mixture-of-experts system built on a new Causal Encoder-Decoder design. DeepSeek says it activates just 8 billion parameters while reading input and 16 billion while generating output, a split aimed at workloads where the prompt is large and the answer is short. It accepts images natively, carries a one-million-token context window, and can emit up to 384,000 tokens. The API model name is deepseek-flash, and the weights are downloadable under an MIT license.

What DeepSeek shipped

The release, announced on DeepSeek's site and mirrored in its API change log, replaces the previous V4-Flash line outright. The company retired V4-Flash and V4-Flash-Vision-Exp, then kept the old identifiers alive as compatibility aliases that route to the new model.

Key specifications, as published by DeepSeek:

• 552B parameters, mixture-of-experts, with a 552B backbone.
• 8B active parameters for input, 16B for output, split across a 20-layer encoder and a 20-layer decoder.
• 1M-token context and a 384K-token maximum output.
• Native image input (JPEG, PNG, GIF, WebP). V4 Pro does not accept images.
• Model ID deepseek-flash, with MIT-licensed weights downloadable on Hugging Face.
• KV cache of roughly 890 bytes per token, which DeepSeek says is one quarter of V4-Flash's HBM use and one eighth of its SSD storage.

The architecture is the reason the price is low. A coding or research agent spends most of its tokens re-reading context: a repository, a system prompt, a growing tool transcript. Activating fewer parameters during that input phase cuts compute, and a smaller cache makes it cheaper to keep that context resident. DeepSeek's release notes say cache-hit charges "often account for a large share of agent costs," so the company is pricing the model around the part of the bill its architecture shrinks.

The price list, and the cache gap

DeepSeek kept the peak and off-peak structure it introduced in August, and made the new rates effective at 04:00 UTC on September 10. According to the official Models & Pricing page, off-peak rates per million tokens are $0.003 for cached input, $0.15 for uncached input, and $0.60 for output. Peak rates are exactly double, and peak hours run weekdays from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC. Weekends stay off-peak all day.

The striking number is the 50-fold spread between a cache hit and a cache miss on input. On most API bills, cached reads are a rounding error. Here they are the difference between a cheap workload and an expensive one. TechBriefly worked an illustrative example: a 500,000-token reusable prefix reused across 100 requests, or 50 million cached input tokens, would cost about $0.15 on V4.1-Flash off-peak, against roughly $15 on Kimi K3 and $25 on Claude Opus 5 at published cache-read rates. Those are list-price comparisons rather than measured bills, and they assume the cache actually hits.

Against V4 Pro, V4.1-Flash is cheaper on every line. V4 Pro lists $0.022 for cached input, $0.66 for uncached input, and $1.98 for output off-peak. Flash is therefore roughly three to four times cheaper, with the widest gap on cache hits. The two models also differ in throughput: V4.1-Flash carries a concurrency limit of 2,500, while V4 Pro's is 500.

The retirement DeepSeek announced, then withdrew

Pricing is not the only reason the release drew attention. When DeepSeek launched V4.1-Flash, it said it would phase out V4 Pro and, from 04:00 UTC on September 14, route every deepseek-v4-pro request to V4.1-Flash at Flash rates. That plan did not survive contact with users.

DeepSeek has since changed course. Its live pricing page now states: "In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged." V4 Pro keeps its own model ID, its own rates, and its 500-request concurrency limit. A future V4.1 Pro model has been mentioned but not dated.

The interesting inversion is not that a cheaper model got better. It is that a flagship stopped being worth serving at flagship prices.

The reversal created a split record online. Several outlets published explainers based on the original notice and still describe September 14 as a forced migration, sometimes in the present tense. DeepSeek's own documentation and the independent trackers that re-checked it, including the LLM Data Hub, have updated to the continuation notice. The accurate current position is the one on DeepSeek's pricing page: V4 Pro remains available, billed exactly as before, with the company adding that it "will provide further notice should there be any changes."

Where it wins, and where it trails

Independent evaluation is still thin, because the model is only days old, but the first results are consistent.

The Vals Index, run by the evaluation firm Vals AI, puts V4.1-Flash at 57.86% overall, which the firm says makes it the number-one open-weight model and 15th of 56 models overall. It edges out Kimi K3 (57.81%) while costing $0.30 per test against $6.47, and finishing tasks in less than half the time. Vals also ranks it first among open-weight models on its Code Migration benchmark and first of 34 on SkillsBench.

Kilo Code, which runs coding-agent evaluations, reports Terminal-Bench 2.0 at 75.3% for V4.1-Flash, at an average cost of $2.58 per benchmark attempt.

DeepSeek's own numbers, which should be read as vendor-reported, put Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, GPQA Diamond at 90.9, and HLE at 36.8 to 39.1 depending on the subset. Those headline figures carry a caveat the technical report makes explicit: results depend on the scaffold around the model. The same model scored 90.6 on Terminal-Bench 2.1 under DeepSeek's minimal harness and 88.0 under Claude Code.

The picture is not uniform superiority. V4 Pro still leads on GPQA Diamond (92.4 to 90.9) and on the text-only HLE subset (42.7 to 39.1). Against the strongest closed models, the gap widens further. On Terminal-Bench 4.0, an independent summary of the published tables puts V4.1-Flash at 31.2 and Claude Opus 5 at 51.8. On HLE, Opus 5 is reported at 56.3. V4.1-Flash is a cheap, fast agent model, not a frontier reasoner, and DeepSeek's decision to keep V4 Pro around acknowledges that the tradeoff is not the same for every workload.

Open weights, and the hardware people are actually using

The weights are open under MIT, and the community moved quickly. Hugging Face lists dozens of community quantizations, and at least one group reports serving the full 552B model on four NVIDIA DGX Spark machines in tensor parallel, at roughly 77 tokens per second single-stream and more than 200 tokens per second aggregate across six streams in one benchmark writeup. That is enthusiast and small-team territory rather than a production reference. The separate 203 GB Engram n-gram memory has to be read from disk instead of held in GPU memory, which is what lets the checkpoint fit across four machines with 128 GB of unified memory each.

Self-hosting still has rough edges. The release does not ship a Jinja chat template, and DeepSeek instead provides a reference encoding implementation and recommends a separate Rust library for production prompt formatting. The parameter counts also invite confusion: DeepSeek describes a 552B backbone, the technical report mentions a 196B Engram component, and Hugging Face's page lists 763B parameters. The company has not reconciled the three figures. The active-parameter numbers, 8B and 16B, describe compute per token, not the memory required to store and serve the model.

Caveats and what to watch

A few cautions are warranted before treating V4.1-Flash as a default.

• Benchmarks are young and partly vendor-supplied. Vals and Kilo are credible independent evaluators, but their runs are days old, and Vals measures a single default configuration. Your workload is the only benchmark that transfers.
• Peak pricing can double a bill. The cheap off-peak numbers assume you can schedule around weekday mornings in UTC. Real-time agents cannot always wait until the window closes.
• Cache hits are not automatic. DeepSeek's caching is best-effort and requires an exact prefix match. A model that looks nearly free in a demo can cost far more when prompt order or tenant-specific prefixes change on every call.
• The roadmap is unsettled. The withdrawn V4 Pro retirement shows DeepSeek is still deciding how its tiers relate. Anyone who depends on a specific model ID should treat it as a moving target.

The larger point is what V4.1-Flash says about where model competition is heading. For two years the pitch was that better models cost more. A flash tier that beats its own vendor's pro tier, at a fraction of the price, on the repetitive work that fills most agent pipelines, argues that the fight over capability is increasingly a fight over cost per completed task. That is good news for anyone paying the bill. It is also a reminder that the bill depends as much on how you architect prompts and cache as on which model name you call.

Cover: DeepSeek logo artwork by RoadMaster19 (CC BY 4.0) via Wikimedia Commons.