The Week the CFO Walked In
There is a moment in every technology cycle when the finance department stops nodding along and asks what, precisely, this is costing. For enterprise AI, that moment landed this week, certified by a Wall Street Journal piece describing a regime change in how American companies buy intelligence. The framing was almost gentle. Companies, it reported, have come around to a radical idea with the potential to upend the industry powering the global economy: they do not have to blow their budgets on AI. Fed up with ballooning costs, firms large and small are mixing in lower-priced models, including some built in China, and shopping a la carte alongside their OpenAI and Anthropic contracts.
The line everyone quoted came from Mike Saeks at Cursor, who said that using the most powerful and expensive models for mundane work is like driving a Lamborghini to the store for milk when the car was built to be raced around a track. It is a good line. It is also, if you have been shipping production systems rather than reading press releases, roughly eighteen months late. The rest of the coverage marked the same shift with a better word: the end of tokenmaxxing, the era in which consuming tokens was itself the deliverable.
I want to be specific about the scale of what just ended, because the numbers are genuinely absurd, and then I want to show you the configuration change that took my own agentic customer service system from two thousand dollars a day to twenty-five. Then, because this is a newsletter for people who have to keep the thing running on Monday, I am going to spend the back half telling you exactly where this trick fails and how it will bite you if you deploy it the way most people are about to deploy it.
The Receipts
Start with Uber, because Uber is the canonical case. In December it rolled Claude Code out to its engineering organization. By February usage had doubled. By April, four months into the fiscal year, the company had burned through its entire 2026 AI budget. CTO Praveen Neppalli Naga confirmed the exhaustion to The Information. Adoption across roughly five thousand engineers went from thirty-two percent to eighty-four percent in four months, monthly per-engineer costs ran between five hundred and two thousand dollars, and something close to seventy percent of committed code was AI-generated. By every adoption metric anyone had defined in advance, the rollout was a triumph. COO Andrew Macdonald then told investors he could not draw a line from the token spending to anything a customer would notice. That link, he said, is not there yet. Management responded by capping tools at fifteen hundred dollars per employee per month.
Meta ran the same experiment with more theater. In April it launched an internal leaderboard ranking its eighty-five thousand employees by token consumption, handing out badges to the heaviest users. The top employee reportedly burned two hundred eighty-one billion tokens in a single month. Collectively the company chewed through something on the order of sixty trillion tokens in thirty days. The leaderboard came down within days. Amazon dismantled its own employee leaderboard after staff started gaming it, and issued internal guidance in May telling engineers to stop using AI just to use AI. Tesla, which had also gamified consumption, capped employees at two hundred dollars a week starting in July. Walmart put boundaries on its internal assistant.
Microsoft revoked Claude Code licenses in its Experiences and Devices division after per-engineer costs hit roughly two thousand dollars a month, steering developers toward its own tooling instead. Salesforce was reportedly staring down a three hundred million dollar annual Anthropic bill and shopping for model routing. Priceline’s senior director of IT finance, Chris Reed, told TechCrunch that a routine Cursor renewal came back four to five times more expensive, and compared the whole dynamic to an addiction: they let you try it to get you hooked, and now you are beholden to it. He also said he was already seeing gaps between vendor-reported usage and Priceline’s own numbers, which should make the hair on the back of your neck stand up. One unnamed company reportedly ran up a five hundred million dollar Claude bill in a single month after forgetting to set usage limits. I cannot verify that figure and neither can the reporters who printed it, but the fact that it is plausible tells you everything.
J.R. Storment, who runs the FinOps Foundation, described the turn bluntly: in April and May he started hearing from companies who were three times over their entire 2026 token budget four months into the year, and the conversation flipped overnight from going fast to needing guardrails. The Linux Foundation has since announced a Tokenomics Foundation to bring the same cost discipline to tokens that FinOps brought to cloud.
The Paradox Nobody Priced In
Here is the part that should embarrass the entire industry, including the parts of it that write newsletters. Token prices did not go up. They collapsed. Performance equivalent to GPT-4 now runs somewhere around forty cents per million tokens, down from roughly twenty dollars per million in late 2022. That is a ninety-eight percent price reduction in about three and a half years, one of the fastest cost declines in the history of computing. Over the same window, enterprise AI bills rose by an estimated three hundred twenty percent, and token usage grew by something over a thousand percent between January 2025 and April 2026.
Economists have a name for this and it is not a mystery. Jevons paradox: make a resource dramatically cheaper and total consumption rises faster than unit cost falls. Every finance organization in America just relearned it the expensive way. The bills did not explode because models got pricey. They exploded because agentic architectures resend the same payload over and over, and because nobody was measuring the loop.
That is the actual defect. An agent loop pays an invisible tax on every iteration: the system prompt, the tool schemas, the policy text, the retrieved context, the entire accumulated conversation, all of it re-submitted to be re-read from scratch on every single turn. A twelve-step tool-calling sequence with a thirty-thousand-token preamble does not cost you thirty thousand tokens. It costs you three hundred sixty thousand, plus whatever the conversation accreted along the way. Multiply by concurrent users, multiply by retries, and you arrive at Uber’s invoice. Nobody set out to spend that. They just never instrumented the thing they built.
The Corporate Response Is Mostly Wrong
Look at what the response has actually been. Caps. Rationing. Per-seat limits. Leaderboards quietly deleted. Guidance memos telling engineers to use less of the tool the company spent a year insisting was civilizational. Dashboards to shame heavy users.
That is austerity, not engineering. It is the same move as fixing a memory leak by buying less RAM. Every one of those controls reduces spend by reducing capability, which means the ROI problem that started the panic gets marginally worse, not better, because you have now capped the numerator along with the denominator. Uber’s problem was never that engineers used too much AI. It was that seventy percent of committed code came out of a system nobody had connected to a business metric, and that the loop was architecturally wasteful. Capping the invoice at fifteen hundred dollars does not fix either failure. It just makes the invoice smaller and the engineers slower.
There is a real fix, and it is not rationing. It is routing. And the mechanics of it are so mundane that I suspect that is exactly why it took a Wall Street Journal story for anyone to notice.
The Hack
DeepSeek’s API speaks both dialects. That is the whole trick.
Their documentation states it in the first sentence: the API uses a format compatible with OpenAI and Anthropic, and by modifying configuration you can use the OpenAI or Anthropic SDK, or any software compatible with either, to reach DeepSeek models. For OpenAI-format traffic the base URL is api.deepseek.com. For Anthropic-format traffic it is api.deepseek.com slash anthropic. You change the base URL and the model name. That is the entire migration. No client rewrite, no adapter layer, no new SDK, no vendor abstraction library, no six-week integration project with a Gantt chart.
If you are on the .NET stack, as I am, this is even less work than it sounds, because Microsoft already solved the abstraction for you. The Microsoft.Extensions.AI.OpenAI package explicitly provides an IChatClient implementation for OpenAI-compatible endpoints, and the official OpenAI .NET client accepts a custom endpoint through OpenAIClientOptions. You construct an OpenAIClient with your ApiKeyCredential and an OpenAIClientOptions whose Endpoint points at api.deepseek.com, call GetChatClient with the model name, and call AsIChatClient on the result. Microsoft Agent Framework documents the identical pattern for custom service URLs. Everything downstream of that IChatClient registration is untouched, because it was written against the abstraction rather than the vendor. If you built your agents properly, the model provider is a dependency injection registration and an appsettings value, not an architecture decision. Swapping it is a config change and a redeploy.
For the Anthropic-format path there is a detail worth knowing, because it is genuinely useful and slightly cheeky: DeepSeek maps Claude model names onto its own. Anything starting with claude-opus resolves to deepseek-v4-pro, and anything starting with claude-haiku or claude-sonnet resolves to deepseek-v4-flash. Which means tools hardcoded against Claude model identifiers frequently work unmodified. Point the environment variables at the DeepSeek base URL and key, and Claude Code, GitHub Copilot CLI, OpenCode and similar harnesses run on DeepSeek with no code changes at all.
The Numbers That Made Me Do It
Now the arithmetic, which is where this stops being a curiosity and starts being a budget line.
Claude Opus 4.8 lists at five dollars per million input tokens and twenty-five per million output. Sonnet 5 sits at three and fifteen, currently discounted to two and ten through the end of August. Haiku 4.5 is one and five. DeepSeek V4-Flash is fourteen cents per million input on a cache miss and twenty-eight cents per million output. V4-Pro, the 1.6-trillion-parameter flagship, is forty-three and a half cents input and eighty-seven cents output, a price that started as a seventy-five percent promotional discount in May and was then made permanent. Both models carry a one-million-token context window and a maximum output of three hundred eighty-four thousand tokens. Both support tool calling and structured JSON output. The weights are on Hugging Face under an MIT license.
On output tokens, V4-Pro is roughly twenty-nine times cheaper than Opus 4.8. V4-Flash is about eighty-nine times cheaper. That alone gets you past ninety percent without trying.
But the number almost nobody talks about is the cache-hit price, and it is the reason my bill fell as far as it did. A cache hit on V4-Flash costs twenty-eight hundredths of a cent per million input tokens. Not twenty-eight cents. Twenty-eight hundredths of one cent. That is fifty times cheaper than the same model’s cache-miss rate and roughly three orders of magnitude below frontier input pricing. And DeepSeek’s context caching is on by default for every account, with no code changes required, because it is a disk-based prefix cache operating below the API surface. The response body reports prompt_cache_hit_tokens and prompt_cache_miss_tokens in the usage block so you can actually verify what you are getting.
Sit with what that does to an agent loop. The invisible tax I described earlier, the system prompt and tool schemas and policy text resent on every iteration, is a stable prefix. Stable prefixes are exactly what a prefix cache is for. The single most expensive structural property of agentic architecture becomes, on this pricing model, nearly free. That is not a discount. That is the cost curve of the entire pattern changing shape.
This is also why the Anthropic-compatibility layer ignores the cache_control parameter, which looks alarming in the compatibility matrix until you understand why. There is nothing to control. You do not annotate cache breakpoints because the server is already doing prefix detection for you.
My own case was an agentic customer service system: high volume, bounded turns, heavy retrieval context, a large stable system prompt, and a fat tool surface. Textbook profile for everything above. It was running two thousand dollars a day. It now runs twenty-five. That is not ninety percent. That is ninety-eight point seven five percent, and the headline on this piece is deliberately conservative because I would rather under-promise a real number than quote my best day.
Compatibility Is the Whole Strategy
Now step back from your invoice and look at what DeepSeek actually built, because the engineering is unremarkable and the strategy is one of the sharpest competitive moves I have seen in this industry. Implementing somebody else’s wire format is maybe a week of work for a competent team. DeepSeek implemented two of them. That is not a convenience feature. That is a targeting decision, and the targets are named OpenAI and Anthropic.
Be precise about what the American labs did and did not standardize, because the popular version of this story gets it backwards. They did converge, and they converged on Anthropic’s protocol. MCP shipped in late 2024, Sam Altman publicly endorsed a direct competitor’s standard in March 2025, OpenAI deprecated its own Assistants API in favor of it, and in December 2025 Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation with OpenAI as a co-founder and AWS, Google, Microsoft, Cloudflare and Bloomberg signing on as platinum members. Roughly ninety-seven million downloads later, MCP is infrastructure. At the tool layer there is no standards war left to fight. Everybody won.
Now notice which layer they standardized and which one they kept. They gave away the layer that makes developers productive and held onto the only layer that makes developers stuck: the inference wire format itself. OpenAI has Chat Completions and Responses. Anthropic has Messages. Neither has ever implemented the other’s, and that is not an oversight or a roadmap gap. That is the moat. Your MCP servers are portable by design. Your call site is not, and the call site is where the contract renewal lives.
DeepSeek went straight at that layer. Not the model layer, where the benchmark data says it is still behind on the hardest work. The switching-cost layer, where both incumbents are completely exposed. And here is the structural trap they walked into: neither American lab can answer this move. For Anthropic to publish an OpenAI-compatible endpoint would be to hand its own install base a free exit ramp, and the same is true in reverse. Reducing migration cost only ever benefits the challenger. This is the incumbent’s dilemma in its purest textbook form, where the correct competitive response is the one thing both incumbents are structurally forbidden from doing.
The legal ground under this is already settled, and it was settled by Americans. In April 2021 the Supreme Court held six to two in Google versus Oracle that copying the declaring code of an API, so that programmers can put existing skills to work in a new computing environment, is fair use as a matter of law. The Court described it as reimplementation of a user interface. That is a verbatim description of what DeepSeek did. There is no cause of action here for anyone. A Chinese lab is cloning American API surfaces under the protection of American case law, won by Google against Oracle five years before anybody needed it for this purpose. Nobody is going to court, and everybody at both labs knows it.
The model-name mapping is the part that tells you this is deliberate rather than incidental. Anything starting with claude-opus resolves to V4-Pro. Anything starting with claude-sonnet or claude-haiku resolves to V4-Flash. DeepSeek is accepting a competitor’s product taxonomy as valid input, which means they built and shipped a compatibility shim keyed to another company’s SKU names. Their own documentation then states plainly that this lets you get around the model restrictions in Claude Desktop’s developer mode. That is not interoperability in any neutral sense. That is an acquisition funnel with the target’s part numbers stamped on the side.
And the deepest irony is that MCP’s success is what made the migration cheap. A portable tool ecosystem means your integrations survive a model swap, which means the single most successful standards contribution Anthropic ever made also lowered the cost of leaving Anthropic. DeepSeek harvests that for free. Note that its Anthropic-format endpoint ignores the mcp_servers parameter entirely: it takes the client-side ecosystem benefit and declines to implement the hosted server-side half. It free-rides on the standard without carrying any of the weight of maintaining it. That is not a criticism of open standards, which I build on daily. It is an observation that standardization always cuts toward whoever has the least to lose from portability.
We have watched this film before. The S3 API became so dominant that every competing object store implemented it in order to compete with AWS, and in doing so turned durable storage into a commodity bought on price. The PostgreSQL wire protocol did the same thing to databases. The pattern holds with unpleasant reliability: the standard-setter keeps the premium tier and the compatible challenger takes the volume underneath it. That is the bet DeepSeek is placing, and on the current evidence it is a good bet.
But understand what you are agreeing to when you take the deal, because compatibility carries a price that does not appear on the invoice. If you write to the intersection of both APIs in order to stay portable, you have permanently capped yourself at the lowest common denominator. No vision. No hosted MCP. No code execution. No cache breakpoints, no thinking budgets, no server-side anything. Look back at that list of unsupported fields and notice that nearly every item on it is a capability a frontier lab shipped specifically to differentiate itself. Portability is purchased with capability ceiling, and you pay for it in features you will never be able to adopt.
Which is the actual strategy, stated plainly. It was never really about taking customers from both labs. It is about redefining the product as a commodity and then winning the commodity, because once the feature surface is reduced to the portable subset the only remaining axis of competition is price, and price is the one axis where DeepSeek cannot lose. Every engineering team that writes to that portable subset is casting a vote for that outcome. I cast mine, with my eyes open, because at ninety-eight percent the arithmetic makes the decision for you. Just know that you voted.
Now the Part That Will Cost You Money If You Skip It
Everything above is true, and if you deploy it as a blanket swap you will regret it. Here is what the enthusiasm is leaving out.
DeepSeek V4 is at parity on bounded work and it is not at parity on long-horizon agentic work. V4-Pro posts 80.6 percent on SWE-bench Verified, which is within a rounding error of Opus 4.6 and the best open-weight score published. It also posts 93.5 on LiveCodeBench and a Codeforces rating north of thirty-two hundred. Those are real. But on Terminal-Bench it lands around 67.9 against roughly 82.7 for GPT-5.5, a fifteen-point gap, and on the tighter DeepSWE harness it scores anomalously low relative to both its peers and its own SWE-bench Pro result. The audit that surfaced that gap read it as a genuine long-horizon agent capability difference rather than a verifier artifact, because V4-Pro ranks lowest on both independent harnesses. The pattern across the whole V4 scorecard is consistent and it is the single most important thing on it: for discrete coding tasks, open weights have caught up. For long-horizon agentic execution, closed frontier still leads.
The independent government read is harsher than the leaderboards. NIST’s Center for AI Standards and Innovation evaluated V4-Pro in May across nine benchmarks, two of them held-out and non-public, and concluded it trails leading American models by approximately eight months rather than the two to three months implied by public scoreboards. CAISI also called it the most capable model from a Chinese lab it has assessed. Both halves of that sentence matter.
Which brings me to the metric that should govern this entire decision and almost never does. Cost per token is the wrong denominator. The right one is cost per successful task, meaning the cost of running the model divided by the rate at which it actually completes the work correctly. When someone ran frontier models against ninety-one private tasks across four multi-week projects and reported it that way, cost per task varied by more than eight hundred times across the field, from north of thirty-one dollars at the top down to about four cents for a quantized DeepSeek variant that never reached frontier quality. A model that is ninety times cheaper per token and fails a third of the time on your workload is not ninety times cheaper. It is more expensive, once you count the retry, the human who reviews the retry, and the customer who got the wrong answer before either happened. Cheap tokens spent on a failed task are the most expensive tokens in your budget.
There are concrete capability gaps too, and they are documented rather than speculative. The Anthropic-format endpoint does not support image or document content blocks, so anything multimodal is out. Server-side MCP is ignored, which means if you are running MCP the way I do you keep the client-side orchestration and lose the hosted convenience. Code execution blocks, MCP tool-use blocks and container uploads are unsupported. The top_k parameter is ignored, and thinking is supported but budget_tokens is not. And there is a footgun sitting right in the documentation that will eventually cost somebody a very confusing afternoon: if you pass a model name the API does not recognize, it silently maps the request to V4-Flash. It does not error. You can believe you are running your hardest workload on Pro while quietly being served by Flash, and the only evidence will be a quality regression nobody can explain. Assert on the model field in the response.
Operationally, DeepSeek’s own infrastructure sits in China and saturates during global peak demand, which shows up as 503s and latency spikes rather than clean failures. Concurrency on Pro is capped at five hundred, Flash at twenty-five hundred. Plan your degradation path before you need it.
The Compliance Conversation You Cannot Skip
If you work in a regulated shop, the hosted API is very likely off the table, and I would rather you hear that from me than from your general counsel. DeepSeek’s privacy policy stores personal data on servers in the PRC, subject to Chinese law. There is no documented Standard Contractual Clauses mechanism for EU transfers, which means routing EU personal data to deepseek.com is not a gray area, it is an Article 46 violation. Italy’s Garante banned the service within days of reviewing it. New York, Texas and Virginia banned DeepSeek from government devices, Congress introduced a bill to formalize that federally, and agencies including NASA and the Navy restricted it. Australia and Taiwan barred it from government devices, South Korea suspended new downloads. Security firm Wiz found an exposed DeepSeek database leaking over a million log entries including chat history and API keys. In the company’s most recent funding round, a Chinese state fund secured governance rights that other investors did not get, which tells you where decision-making authority formally sits.
Now the part the scare coverage tends to omit: none of that requires you to give up the economics, because the weights are MIT-licensed. V4-Flash and V4-Pro have been available in Microsoft Foundry since May with region selection and EU data residency, which for a Microsoft-stack shop means Entra ID integration and your existing network and security posture rather than a new vendor review. AWS Bedrock and Google Vertex AI carry DeepSeek models in EU regions, with V4-Pro rolling out to US regions first. Together AI, Fireworks and OpenRouter host outside China. Or you self-host the weights on vLLM and no token ever leaves your environment. You pay a premium for this. Together and Fireworks run roughly thirty to fifty cents input and fifty to ninety cents output, and Bedrock lists V3.2 around sixty-two cents and a dollar eighty-five. That is more than the direct API and still five to twenty-five times cheaper than frontier pricing. Two honest caveats: self-hosting fixes the data path but does not by itself satisfy GDPR, and it does not touch the political content behavior, which lives in the weights and travels with them.
What Actually Works
The architecture that survives contact with production is tiered, and the WSJ described it without quite naming it: model mixing. A frontier model orchestrates, plans and handles the long-horizon reasoning where its lead is real and measurable. Cheap open weights execute the bounded, high-volume, well-specified work where parity is also real, which is most of your token volume by an enormous margin. Classification, extraction, summarization, retrieval synthesis, routing, single-turn customer service, structured output generation. None of that needs a Lamborghini, and the benchmark evidence says DeepSeek is not meaningfully worse at any of it.
That is why my customer service agent worked so well and why I would not put a multi-hour autonomous coding loop on the same configuration without evaluating it first. Customer service turns are bounded. They are the good case. Long-horizon autonomous execution is the documented bad case. The split is not ideological, it is empirical, and it is sitting in the benchmark data for anyone willing to read past the headline number.
So build the routing layer. Instrument the loop before you optimize it. Run both models against a golden set drawn from your actual production traffic and score cost per successful task, not cost per token. Assert on the model name in the response so a silent fallback cannot masquerade as a quality mystery. Watch your cache hit ratio like you would watch a database connection pool, because on this pricing model it is the single biggest lever you have. Keep a frontier fallback wired and tested for when Beijing is saturated at four in the afternoon Pacific.
The Real Hack
Here is the uncomfortable conclusion, and it is not about China and it is not about DeepSeek.
The configuration change is twenty minutes of work. Base URL, model name, redeploy. That part is trivial, which is precisely why it cannot be the thing that saved ninety-eight percent. What saved ninety-eight percent is that somebody finally measured. Uber did not have a cost problem, it had an instrumentation problem, and so did Meta and Salesforce and Microsoft’s Experiences and Devices division and every company that built a token leaderboard and called it a strategy. If you do not know your cache hit ratio, your tokens per successful task, and your cost per resolved ticket, you do not have an AI budget. You have an AI subscription with a variable price and no meter, and you found out what it cost the same way everyone else did, in April, from accounting.
The one detail I keep turning over is that Microsoft, which pulled Claude Code licenses from its own engineers over roughly two thousand dollars a month per head, is reportedly evaluating DeepSeek V4 to power a cheaper Copilot tier. The company that rationed the expensive model is shopping the cheap one. That is not hypocrisy. That is the correct engineering response arriving about two quarters after the austerity memo, which is roughly the industry’s standard latency between panicking and thinking.
Do not read this as an argument about geopolitics. Read it as an argument about portability, with both edges of that word intact. The moment your model provider became a base URL and a string in appsettings rather than an architectural commitment, you acquired real leverage over the most volatile line item in your budget, and you also accepted a ceiling on the features you can ever use. That is a trade, not a free lunch, and it is a trade worth making for the ninety percent of your token volume that is bounded, high-volume and unglamorous. Most teams are holding that leverage and not using it, because they wrote their agents against a vendor SDK instead of an abstraction, and now the swap that should take twenty minutes takes a quarter. Fix that first. The savings follow, and they follow no matter which lab wins next year, which is the entire point of building it that way.
I am the .NET AI guy that ships. Ship the routing layer.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs, and Chief Architect at Prana Entertainment, a Las Vegas-based enterprise software and AI consultancy. He has spent nearly thirty years shipping production software on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy and Allegiant Air. He was employee number six at Break.com and a Los Angeles Business Journal CTO of the Year finalist.
His current work is production agentic AI: MCP servers, the Microsoft Agent Framework, RAG pipelines, SQL Server 2025 vector search, and the PLogger observability framework, which exists largely because too many teams are running agent loops they cannot see. He also runs ShopSnap, a zero-transaction-fee e-commerce platform. He is the author of the novel The Archive of Lost Suns.
Outside of work he trains Brazilian jiu-jitsu under Sergio Penha in Las Vegas, rides motorcycles, and co-hosts Edge Grip Podcast. He writes about what actually happens when AI systems meet production, which is rarely what the launch post promised.