Anthropic Discounted the Cache Line. DeepSeek Engineered It Away. V4.1 Flash Prices Cache Hits at $0.003.

DeepSeek released V4.1 Flash this morning. New API prices took effect at 04:00 UTC. The weights went up on Hugging Face under an MIT license. Off-peak cache-hit input is priced at $0.003 per million tokens.

Nine days ago I wrote that Anthropic had quietly become the most interesting company in AI pricing by holding the Fable sticker at $10 and $50 while cutting cache reads from $1.00 to $0.25 per million, a 75 percent cut on the line that actually dominates an agent invoice. That looked like the frontier floor. It was the frontier floor for nine days.

$0.003 is roughly one eighty-third of $0.25. But the number is not the story, because anyone can print a number. The story is that DeepSeek did not get there by accepting less margin. It got there by changing what a cache read costs to serve.

What Actually Shipped

V4.1 Flash is a 552 billion parameter mixture-of-experts model. Eight billion parameters are active during prefill and 16 billion during decode. Context window is one million tokens. It is multimodal, and it is the first model in a new DeepSeek architecture family the company calls Causal Encoder-Decoder, or CED.

CED is 40 Transformer layers arranged as a 20-layer causal encoder followed by a 20-layer decoder. The load-bearing change is where the KV cache comes from. In a conventional decoder-only stack, every decoder layer derives its own keys and values, and the cache you have to hold in memory grows with layer count. CED projects the decoder's global KV cache directly out of the encoder hidden states instead, so the per-layer derivation goes away.

The number that falls out of that: 890 bytes of KV cache per token. DeepSeek says that is a 75 percent improvement over V4 Flash and roughly one four-hundred-and-thirty-seventh of what the original V1 required per token. HBM demand drops to a quarter of V4 Flash. SSD demand drops to an eighth.

Do the arithmetic on the cache figure, because it is the whole argument. A one million token context at 890 bytes per token is about 890 megabytes of KV state for a single fully loaded request. That fits inside a fraction of one accelerator. The previous generation of this model needed four times the high bandwidth memory to hold the same conversation, and the generation before that was in a different universe.

The Cache Line, Across the Market

I keep coming back to cache reads because agents do not work like chat products. An agent loop resends the same system prompt, tool definitions, repository context and a growing transcript on every single step. Cached input is not a footnote on that invoice. It is usually the largest line on it.

Here is where the frontier sits this morning, all figures per million tokens.

Two caveats on that table before anyone screenshots it. DeepSeek runs time-of-day pricing, with peak hours at 01:00 to 04:00 and 06:00 to 10:00 UTC Monday through Friday and off-peak rates at half the peak rate, so I have shown both rows rather than quoting the friendlier one. And I have marked the Opus 5 cache line as not published rather than guessing at it.

The ratio column is the part I find genuinely interesting. Anthropic and DeepSeek have landed within half a percentage point of each other on cache read as a share of input price, 2.5 percent versus 2.0 percent, from completely different directions and completely different cost bases. When two labs that share no silicon, no cloud and no incentive structure converge on the same ratio, that ratio is probably telling you something true about what a cache read costs to serve on 2026 hardware. GPT-5.6 Sol sitting at 10 percent now looks less like a pricing strategy and more like a line nobody has revisited.

A Discount and an Architecture Are Not the Same Move

This is the distinction I want to draw hard, because the coverage today is going to collapse it into one number and miss it.

Anthropic's cut was a repricing. The company looked at what a cache read costs on Trillium and Blackwell class hardware, concluded the old $1.00 was well above cost, and moved the line to $0.25 while keeping the sticker intact. Smart, well-timed, and entirely reversible. It is a margin decision, and margin decisions can be unmade in a quarter.

DeepSeek's cut is a consequence. If your KV cache is 890 bytes per token and your HBM footprint is a quarter of what it was, then the thing you are billing for genuinely costs you a small fraction of what it cost your previous model, and $0.003 is not a loss leader, it is a margin on a cheaper operation. DeepSeek also raised the concurrency ceiling from 500 to 2,500 requests at the same time, which is not what a company does when it is eating losses to buy attention. It is what a company does when a serving unit got four times lighter.

You do not have to take that on faith, which is the part that matters. The weights are MIT on Hugging Face. Anybody with a couple of accelerators can load V4.1 Flash, measure the resident KV state per token, and check whether 890 bytes is real. Anthropic's $0.25 rests on an invoice. DeepSeek's $0.003 rests on a file you can download. Whatever else is true about this company, that is the first cache-economics claim at this tier that any outsider can independently falsify.

The Benchmarks, With Salt

The scores DeepSeek is leading with put V4.1 Flash ahead of or near models that cost thirty to eighty times more per token. On CyberGym, a security-focused evaluation, V4.1 Flash reports 88.1 against 84.5 for both GPT-5.6 Sol and GLM 5.3, and 80.0 for Kimi K3. Across several agentic and coding evaluations the model lands ahead of or within striking distance of GPT-5.6 Sol and Claude Opus 5.

Standard disclaimer applies and I want to apply it loudly. These are lab-reported figures on the release-day post, and release-day figures from every lab in this industry have a well-documented habit of compressing under third-party reproduction. The CyberGym result in particular deserves a hard look, because a security benchmark is exactly the sort of evaluation where contamination in the training mix produces a flattering number that does not survive contact with a novel target. We will track the reproductions on our benchmarks page over the next two weeks, and the reproductions are what count.

One structural note that is not a benchmark. V4 Pro is being retired, with its traffic rerouting to V4.1 Flash at V4.1 Flash pricing on September 14. DeepSeek is not positioning Flash as the cheap tier under a flagship. It is positioning Flash as the replacement for the flagship. That is a real statement of confidence and also a real risk, because everybody currently on Pro gets moved onto a four-day-old architecture whether they benchmarked it or not.

The Thing Nobody Wants to Put in the Same Paragraph

This model shipped two days after the NSA, CISA and FBI published a joint advisory naming DeepSeek first among six Chinese labs accused of industrial-scale distillation against US frontier models. Kira Nolan covered that advisory yesterday, including the detail that DeepSeek opened the V4.1 Flash public test endpoint on September 8, the same day the advisory landed.

I am not going to relitigate the training-data question here. But I do want to separate two claims that are getting welded together this week, because keeping them apart is the only way to read today's release accurately.

The advisory's argument about cost is that DeepSeek's widely cited $5.6 million training figure is misleading because it excludes the value of data acquired through distillation. Grant that argument entirely, for the sake of the point I want to make. It says something about how the model learned. It says nothing whatsoever about the Causal Encoder-Decoder.

You cannot distill an architecture out of an API. Prompting a frontier model several billion times gets you its outputs. It does not get you a scheme for projecting a global KV cache from encoder hidden states, because that scheme is not expressed in any token any model has ever emitted. Whatever provenance argument you accept about the training corpus, the memory layout is original engineering, it is published, and it is licensed permissively enough that American labs can read it this afternoon and copy the good parts.

Which produces the genuinely awkward situation of the week. A US enterprise reading the September 8 advisory has a real reason to keep its production traffic off a DeepSeek endpoint. That same enterprise's inference team has a real reason to read the CED paper carefully, because if the 890 byte figure holds it is the most important serving result published this quarter and it arrived under a license with no strings on it.

Our Take

The frontier labs spent 2026 competing on the cache line by discounting it. DeepSeek just demonstrated that the line is a memory-architecture problem, not a pricing problem, and that the ceiling on how far it can fall is set by bytes per token rather than by anyone's willingness to compress margin.

That reframes the last nine days. Anthropic's 75 percent cut was the right move and I still think so. But it was a move inside the old frame, where cache reads cost roughly what they cost and the only variable is how much of that you mark up. If CED holds up, the variable was never the markup. It was the byte count, and nobody at the frontier had attacked it directly.

Practical read for anyone building agent-shaped products. Nothing forces you onto a DeepSeek endpoint, and if you operate under US federal procurement or a security review that cites the September 8 advisory, today changes nothing about where your traffic goes. What today changes is the number you should expect from your existing provider twelve months out. A cache-hit price at two percent of input is now demonstrated at the substrate level, under a license anyone can read, and your provider's inference team read it this morning too. If you are modeling 2027 cost of goods sold on a cache line that holds at $0.25, model a second scenario. You can size both against your own traffic on our cost calculator.

Three things I am watching over the next 90 days. Whether an independent group reproduces the 890 bytes per token figure from the public weights, which is the direct test of whether this is an architecture result or a marketing result. Whether any US or European lab ships an encoder-decoder cache scheme in a production serving stack, which is the direct test of whether CED is a DeepSeek trick or the next default. And whether GPT-5.6 Sol's cache read moves off 10 percent of input, because Sol is now the only frontier tier still pricing that line the way the whole industry priced it a year ago, and that gap is visible to every buyer who runs an agent.

One release does not settle this. But the pricing floor argument changed shape today. It stopped being about who is willing to charge less and started being about who has to.