Hook
Three dollars per thousand tokens is a rounding error at a macro fund and an existential fact at a two-person startup. On September 10, 2026, DeepSeek shipped V4.1-Flash and moved a number that matters far more than any benchmark row: cache-hit pricing at $0.003 per token during off-peak hours, against $0.022 on the outgoing V4-Pro. A 7.3x compression on the single line item that determines whether a long-running agent loop is a business or a science project.
The second number is colder. KV cache footprint of 890 bytes per token. Down 75% from V4-Flash. One four-hundred-and-thirty-seventh of DeepSeek V1.
I read changelogs before I read press releases. Habit from auditing ICO whitepapers in 2017, when the distance between the deck and the repository was where all the money lived. This changelog is blunt. From September 14, every request for the retiring V4-Pro routes automatically into V4.1-Flash at the lower price. No migration window. No grandfathering. No opt-in. The predecessor is deprecated by routing table — the most aggressive end-of-life mechanism I have seen from a seventy-one-billion-dollar lab.
Then the concurrency line. Five hundred requests to two thousand five hundred. Blended cost down 77-80%. Throughput up fivefold. Those two figures together do not describe a model release. They describe a capacity expansion wearing a version number.
Markets price the headline. The headline is a benchmark table. The body is a memory bus.
Context
DeepSeek is not, and has never been, a model company in the way Western labs are model companies. It is a compression company that ships models as a byproduct. The through-line from V1 to V4.1-Flash is not capability. It is bytes. Every generation, the question is how much state you can drop before the output degrades. That discipline was not chosen for elegance. It was chosen because the binding constraint on a Chinese lab is not talent and not capital — it is memory hardware, and memory hardware is precisely what export controls bite.
Compression is a sanctions workaround that happens to be good engineering. That context matters, because it explains why the price cut is structural rather than promotional. A lab that has to buy HBM on a constrained market has an existential reason to delete three-quarters of its KV footprint. A lab that rents unlimited compute does not.
The rest of the wrapper is the standard 2026 playbook. A one-million-token context window. Native multimodal input. An MIT license, which is the quiet strategic weapon here — no revenue share, no field-of-use restriction, nothing for an enterprise legal team to negotiate. MIT is a distribution decision dressed as an ideological one, and it is how you get onto a regulated balance sheet without a procurement cycle. The valuation attached to all of it: $71 billion.
Also worth isolating is what off-peak pricing is actually doing. It is not a discount. It is demand shaping against a fixed memory fleet. A lab holding large resident-KV capacity has idle residency at 04:00 Stockholm time and saturated residency at 14:00. Pricing that differential at a 7.3x spread converts unused memory into revenue without adding a single accelerator. Anyone reading $0.003 as a permanent price rather than a load-shifted price is modeling the wrong curve.
Now the benchmark table, and this is where I stop skimming. At maximum reasoning effort, V4.1-Flash posts 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 88.1 on CyberGym. Every one of those beats the outgoing V4-Pro. Then the pure reasoning suite: 90.9 on GPQA Diamond, against 93.4 for Opus and 94.1 for GPT-5.6 Sol. Third place.
Read that split carefully, because it is the entire story of 2026.
A model that loses on GPQA and wins on CyberGym is not a worse model. It is a differently shaped one. GPQA measures whether a system can answer a hard question once. CyberGym measures whether a system can hold a threat model across a long horizon, execute tools, and not lose the plot at step four hundred. Those are different skills, and only one of them is billable at scale right now.
And note the parameter count. V4.1-Flash surpasses V4-Pro on every agentic benchmark while running three times fewer active parameters. That is not a marginal tuning win. That is an architectural claim, and architectural claims get tested by the serving bill, not by the leaderboard.
The pricing conflict around all of this has been running for two years and has now changed its unit of account. It used to be dollars per million tokens. It is now, quietly, dollars per sustained agent-hour — and the second metric implies the first plus everything the first ignores: cache residency, SSD spill, retry overhead, and the cost of a loop that fails at step nine hundred.
Core
Now the mechanism, because the mechanism is the price.
The industry's default decoder stack derives its KV cache from every individual decoder layer. Layer by layer, state accumulates, and the memory cost of that accumulation is what caps context length and concurrency. DeepSeek's V4.1-Flash introduces what it calls a Causal Encoder-Decoder architecture, in which the decoder's global KV cache is projected directly from the encoder's hidden states rather than derived per layer. The per-layer derivation path is what gets deleted. I have not seen the architecture paper, and I would not sign off on the claim from a changelog alone — but the observable consequences are consistent with the description, and consistency under observation is the standard I apply.
The consequence that matters is an asymmetric activation pattern. The Mixture-of-Experts backbone carries 552 billion parameters. During prefill it activates 8 billion. During decode it activates 16 billion.
The asymmetry is not a cost trick. It is an error-budget allocation.
Prefill is a single parallel pass over the prompt. There is no reuse pressure, no compounding, no place for a mistake to metastasize. A thin expert slice is sufficient because the pass is one-shot. Decode is autoregressive. One token at a time, each conditioned on the last, and any drift introduced at step fifty is inherited by step five hundred. That is where parameters have to be spent, because that is where error compounds. The design spends compute where mistakes are irreversible and withholds it where they are not. Anyone treating this as a serving optimization has the causality backwards.
Then the memory math, which is the part the market will actually trade.
Eight hundred and ninety bytes of KV per token. At a one-million-token context, that is roughly 890 megabytes of resident cache per live sequence. Compare against the V4-Flash figure implied by a 75% reduction: on the order of 3.5 gigabytes per sequence at the same context length.
Now apply the concurrency number. Two thousand five hundred concurrent requests against a long context is not a flop constraint. It is a working-set constraint. At 3.5 gigabytes per sequence, 2,500 live sequences demand something on the order of nine terabytes of cache residency. At 890 megabytes, the same concurrency demands roughly 2.2 terabytes. A 75% reduction in per-sequence footprint is what buys a fivefold increase in concurrency without quadrupling the fleet. The two headline numbers are not separate announcements. They are the same announcement, printed twice in different units.
And here I will flag the figure I do not trust. The claim that 890 bytes is one four-hundred-and-thirty-seventh of DeepSeek V1 implies roughly 389 kilobytes of KV per token in the original model. Run that forward to a million-token context and you get an absurd quantity. The scalar is a marketing artifact, not a like-for-like comparison — it is almost certainly measuring across different context regimes, layer accounting conventions, or both. The honest ratio is 75% against V4-Flash. Everything beyond that is a rounding artifact wearing a superlative.
Why 890 bytes matters more than the token price: it moves the hot path. The serving cost stack is a hierarchy — HBM, then DRAM, then SSD, then network. Every tier you fall through multiplies latency and cost. When a long-context sequence fits inside a single node's DRAM, you delete a tier from the hot path. That is where gross margin actually comes from. Not the sticker price. The residency. Memory, not compute, sets the price of intelligence.
One piece of the arithmetic is easy to miss. A one-million-token context window and native multimodal input do not merely coexist — they multiply. Vision tokens are expensive in the KV budget, and a long-horizon agent that screenshots a user interface on every step carries an image payload through the entire prefix. Cut the per-token footprint by three-quarters and you have not only made text cheaper. You have made it practical to keep a visual history resident across a thousand-step task. That capability was theoretical at 3.5 kilobytes per token. It is operational at 890 bytes.
I spent three months in 2020 modeling Uniswap v2 and Compound liquidity depth against Ethereum gas spikes. The lesson I carried out of that paper was not about DeFi. It was that aggregate capacity numbers are meaningless without a per-unit residency and a per-unit cost. An AMM with two billion dollars of TVL and no depth at the edge is a two-billion-dollar illusion. A model with a million-token context window and a cache that spills to SSD is the same illusion one layer down, wearing better branding.
So let us price the loop.
A long-horizon agent — one that executes two hundred to five hundred tool calls on a task, each call carrying a large cached prefix — is a token furnace. Two hundred calls against a two-hundred-thousand-token context is 40 million tokens of cache-hit traffic for a single task. At the old $0.022, that task costs on the order of eight hundred and eighty thousand dollars. At $0.003, it costs a hundred and twenty thousand. Neither is cheap. But only one of those numbers belongs in a business plan, and the gap between them is the gap between an agent startup with a gross margin and an agent startup with a fundraising story.
That is the real function of the release. It does not make agents cheap. It makes agents fundable at the unit level rather than at the narrative level.
Which pulls the settlement layer into the conversation, because agents that transact must pay for their own inference, and the arithmetic there is unforgiving. A token of cached thought now costs three-tenths of a cent. If the settlement rail charges half a cent to move a payment, settlement is 167% of the transaction. Per-token on-chain settlement is not a design choice. It is an impossibility at any scale. What survives is batching: pre-funded escrow, channel-based accounting, periodic netting to a cheap execution layer. The agent economy does not need a blockchain per inference call. It needs an accounting layer that treats inference as a metered utility and settles the meter in arrears.
Then the decentralized compute question, where the crypto market has been getting the physics wrong for two years.
DePIN compute networks — the Akash, Render, io.net archetype — have been sold on a per-GPU-hour price. That was never the right pitch, because inference is memory-bandwidth-bound, and consumer hardware has structurally poor memory bandwidth. A node with a consumer accelerator can beat a datacenter on price per hour and lose catastrophically on tokens per second at long context. The binding constraint on decentralized inference was never GPU count. It was bytes of resident state per accelerator. A 75% cut in KV footprint moves that constraint toward the small operator for the first time, and nobody is pricing that yet.
DeepSeek's own framing tells you which metric it thinks the war is fought on. The company points at HBM demand and SSD persistence as the dimensions where competitive survival is decided. That is a hardware-supply-chain sentence emerging from a software release. Read it as a forecast: the next twelve months of model competition get fought on the memory hierarchy, and the winners will be the labs that treat bytes per token as a first-class product metric rather than a serving detail.
Contrarian
The consensus read on V4.1-Flash is already forming, and it has three parts. Cheaper inference reduces total compute demand. Reduced compute demand deflates the AI capex trade. And decentralized compute networks, which sold themselves as the cheap alternative, lose their only pitch.
I think the first two are wrong and the third is right for the wrong reason.
Start with demand. A lab does not quintuple its concurrency ceiling on a model it expects to be used less. The 500-to-2,500 expansion is a capacity-planning statement, and capacity plans get made against expected load. DeepSeek is telling the market, through an infrastructure number rather than a press quote, that it expects the volume curve to keep bending up after the price falls. This is Jevons with a memory bus attached. Every reduction in the cost of a unit of work pulls previously uneconomic work across the viability line, and agentic work has a very long tail of tasks that were uneconomic at $0.022 and are merely marginal at $0.003.
The capex trade is a different question, and I will not pretend the answer is comfortable. If the cost of an inference falls 80% while the capex required to produce it does not, the revenue line servicing that capex compresses. That is real repricing pressure, and it lands on the hardware complex — which is correlated with crypto beta whether or not anyone wants to admit the linkage. Entropy is the only constant in liquid markets, and the inference layer just took a step change in entropy. Every price built on the old cost floor is now a stale quote.
But here is the part the consensus skips. Cheap inference does not eliminate demand for verification. It expands it.
At $0.022 per token, nobody pays for a proof. Adding attestation overhead to an already expensive call is a non-starter. At $0.003, the calculus shifts — but only for work where the consequence is financial. When an autonomous agent signs a transaction, moves collateral, or files a compliance report, the question of which model, which weights, and which version produced that decision stops being a technical detail and becomes an audit artifact. No regulator accepts that the model said so. No institution with a fiduciary duty accepts an unattested inference sitting behind a position.
Verifiability is the premium that survives a price war, because demand for verification scales with consequence, not with cost. TEE attestation, optimistic challenge windows, and the long tail of proof systems are not competing with cheap inference. They are the tax charged on top of it, and the tax base grows as the underlying work gets cheaper.
Fractures in the ledger reveal the truth of value. That line has always meant something narrow to me: the interesting information is never in the aggregate, it is in the seam. The seam here is the moment an inference becomes a financial act. That moment is where an on-chain settlement layer earns its existence — not by paying for tokens, which is uneconomic at any realistic fee level, but by recording which model produced which decision, under which weights, at which time. Provenance, not payment, is the crypto-native contribution to the agent economy.
There is a reflexivity point here that Bitcoin taught us and everyone keeps forgetting. Bitcoin's fee market had no serious bid until inscriptions showed up — a use case nobody modeled, filling blockspace and funding security with demand that had nothing to do with the monetary thesis. The lesson was not that inscriptions are valuable. It was that a fee market only becomes defensible when a second, non-obvious demand source shows up. Inference markets are about to run the same experiment, and the outcome will not be decided by anyone's roadmap.
Now the jurisdictional layer, because capital flows toward the cheapest rulebook. Hong Kong's virtual asset licensing regime has been marketed as innovation-friendly. Read the actual mechanics and it reads as a venue grab — the objective is not to nurture builders but to capture the listing and clearing flow that Singapore has been holding. That contest now has an AI dimension, because compute sovereignty and settlement sovereignty are converging into one argument. A Chinese lab publishing MIT-licensed weights at a $71 billion valuation has already solved its distribution problem through a legal instrument rather than a hardware one. The jurisdictions competing to be the venue for that flow are not making a philosophical choice. They are bidding for the pitch.
Takeaway
We are in a sideways tape, and sideways tapes reward positioning over prediction. So position against the right variable. The mispriced input is not compute. It is memory — bytes of resident state per accelerator, per unit of cost. Everything downstream of that variable just got re-based, and most of the market is still valuing assets against the old floor.
Three things I am watching. Whether the automatic V4-Pro routing on September 14 holds latency at five times the concurrency — if it does not, the price cut is a marketing event and the architecture claim stays unproven. Whether any decentralized compute network publishes an inference price denominated per token rather than per GPU-hour — the first one that does is telling you it believes the memory math. And whether the first settlement layer for attested inference finds real volume, because that is the only place in this stack where a fee is structurally defensible.
If a token of cached thought costs three-tenths of a cent, and the marginal cost of a long-horizon agent task has fallen by three-quarters while its capability rose, then the question is no longer whether machines will transact. The question is what the ledger is for. It is not there to pay for the thinking. It is there to remember who did it.