DeepSeek says its newest coding model finishes a DeepSWE task for an estimated just under 20 06 cents, according to a cost chart. DeepSeek says its previous Flash model needed about 50 07 cents. That is a 60% 25 cut in one release.
MiniMax went further on October 19. The company says its M3.1 Flash cuts agent task costs by 70% 08, and it lists input at $0.15 12 per million tokens. DeepSeek lists the same $0.15 13 rate for off-peak cache-miss input on V4.1 Flash. Two rivals, one price.
DeepSeek posted V4.1-Flash-1017 to Hugging Face on Saturday, October 17. On Monday, Politico reports, the Department of Homeland Security filed an order removing DeepSeek products from federal civilian agency IT systems. The order bars new purchases and reaches third-party services built on DeepSeek models. DHS wrote that the models "present multiple threat vectors," according to Politico.
We should admit something up front: the timing is less dramatic than it looks. BBC reports the order was issued in July and only made public Monday. For a builder selling into government, that detail changes nothing. The cheapest model in the race became unsellable to an entire class of buyers in two days.
Matching price tags push the decision upstream
Once two vendors post the same $0.15 13 input rate, the token price works like a floor. Nobody wins on it. The choice moves to questions a price sheet can't answer, and those questions come in a fixed order.
What one task costs, and what one merged change costs
The first is a gate: are you allowed to buy this model at all? The DHS order is exactly that kind of gate. For a civilian agency, DeepSeek's price now carries zero weight, because the agency cannot sign the contract. Approved vendor lists and data residency rules work the same way.
The second is a meter: what does one accepted change cost, all in? Token price is only one input. Agent overhead such as retries and tool calls makes up much of the rest. Human review adds more on top. Artificial Analysis figures reported by The New Stack put MiniMax M3.1 Flash at about $0.22 10 per Intelligence Index task against $0.51 for MiniMax M3, a nearly 57% cut, while MiniMax claims 70%.
The third is the tag itself. MiniMax lists $0.60 per million output tokens and $0.015 27 per million cached input tokens, a 90% discount on repeated prompts. Those numbers still matter. They just get read last.
Sorted this way, the week's news gets a lot simpler. The DHS order belongs at the gate. The 13-point 28 gap between MiniMax's 70% 10 claim and The New Stack's 57% measurement belongs on the meter. The matched $0.15 13 rate is just the tag.
Tracing the model behind every agent
Think of a coding agent as a contractor holding a master key. It edits files in your repositories and runs shell commands on your machines. When two contractors bid the same day rate, you stop comparing rates and start asking whether building security will let them past the lobby.
The DHS order checks the badge of every contractor's employer too. The order applies to U.S. federal civilian agencies. Politico reports it covers third-party offerings developed using DeepSeek implementations. A team selling a coding assistant to a civilian agency now has to know which model runs underneath. Ground News quotes DHS warning that use "may result in the transfer of sensitive data to the PRC."
Knowing what runs underneath is harder than it sounds. Tech Insider reports V4.1-Flash-1017 kept the same deepseek-flash API name as its predecessor, so existing code picks it up with zero changes. Good for shipping speed. It also means a vendor can swap checkpoints without a line in your changelog.
The real money hides in how agents behave. The New Stack reports repeated runs of the same agentic task can burn as much as 30 11 times more tokens. At that multiple, DeepSeek's 20-cent 06 task becomes a $6 29 task, and the sticker price never moved.
So the number to track is cost per accepted change. Take everything you spent on a batch, then divide by the changes your team actually merged. Here is a made-up batch of 100 30 tasks with round inputs you can swap for your own.
Model A costs $0.20 30 a task and lands 90 of 100. Model B costs $1.00 31 a task and lands 98. Say each failed patch eats 30 minutes 32 of an engineer billed at $100 an hour, so $50.
- Model A: $20 33 in inference plus 10 failures at $50 is $520. Divide by 90 34 merged changes and you get $5.78 each.
- Model B: $100 35 in inference plus 2 failures at $50 is $200. Divide by 98 36 and you get $2.04 each.
In that batch, the cheap model costs more per merged change. Failure recovery ate the savings. Your own pass rates will decide which side of that line you land on.
Quality signals here are thin. MiniMax's published comparison scores DeepSeek V4.1-Flash-1017 at 77 21 on DeepSWE and M3.1 Flash at 73.7 22. The same chart lists Claude Opus 5 at 74 24, while DeepSeek's own comparison table reports GPT-5.6 Sol at 73. VentureBeat's comparison called MiniMax competitive but flagged a high hallucination rate.
It's unclear whether a four-point benchmark spread predicts anything on your codebase. We think public benchmarks earn a model a spot on the shortlist. Your own backlog should pick the winner.
Where the cheap model gets expensive
The ban moved from phones to the supply chain.
In about 21 months 38, restrictions went from DeepSeek's app on government devices to any third-party product built on its models. A vendor selling to a civilian agency now has to know which model runs underneath.
Vendor savings claims outrun measured savings.
MiniMax claims a 70% 08 cut in agent costs. Artificial Analysis figures reported by The New Stack show about 57% 10, from $0.51 to $0.22 per task.
Repeated runs can multiply the real bill.
The New Stack reports that repeated runs of the same agentic task can burn up to 30 11 times more tokens. At that multiple, a 20-cent 06 task becomes a $6 29 task while the sticker price stays the same.
Model provenance checks spread by 2028
The restrictions have widened on a steady clock. DeepSeek's R1 app became the most downloaded in the US in late January 2025, Inside Government Contracts reports. On February 7, 2025 15, Representatives Gottheimer and LaHood introduced the No DeepSeek on Government Devices Act. Ground News reports the Navy had already banned DeepSeek on government devices, Commerce Department bureaus reportedly prohibited it as well, and several states added limits.
October 19, 2026 19 moved the line from devices to the supply chain. In about 21 months 38, the scope went from one app on government phones to any product built on the model. Inside Government Contracts compared the early response to the FAR rule against Huawei telecom gear and services. Telecom buyers learned that once a supplier is named, procurement teams start checking components.
Price runs on a faster clock. DeepSeek's own figures show task cost falling about 60% 25 in one Flash generation. If the next generation repeats that, a 20-cent 06 task drops to 8 cents. At 10,000 40 tasks a month, the bill goes from $2,000 to $800, a saving of $1,200 a month.
A contractor that loses civilian agency eligibility doesn't lose $1,200 40 a month. It loses the contract, plus the engineering time to rip a model out of shipped software. The upside of chasing the cheapest model shrinks with every price cut. The downside of shipping an ineligible model has no ceiling.
The ban does have limits. It applies to civilian agencies, and reports describe military systems separately. DeepSeek ships the weights under an MIT license, according to its Hugging Face release, so a commercial team can self-host and keep code on its own network. For startups and low-risk batch work, a 60% 39 cut still decides budgets.
If the scope keeps widening at the same pace, another 21 months 41 lands in mid-2028. Our read is that regulated buyers will ask for model provenance in standard contracts by then. Provenance here means the documented origin of the model: who trained it and who serves it.
Map your models, then price merged changes
Step one: list every coding tool your team uses. Next to each, write the model and the provider serving it. Ask vendors in writing when the docs are silent. Every blank cell is a finding, and you will probably find a few.
Step two: tag each repository green or red. Green is public, internal tooling or throwaway code. Red holds customer data or touches a government contract. Red repos only get models your contracts allow.
Step three: pull 20 06 closed tickets from your green repos. Run each one through OpenHands twice, once per model you are cleared to use. OpenHands 1.28.1 is an open-source coding agent that edits, debugs and maintains repositories, and it asks for secrets inline during a run. Log tokens, retries, pass or fail, and the minutes a human spent reviewing each patch.
Some runs will stall or loop. Log those too. The New Stack's 30x 11 token figure comes from exactly that kind of loop, and your failed runs are the data that matters most.
Step four: compute cost per accepted change. Add inference, tool calls, review time and failure recovery, then divide by merged patches. At the just-under-20-cent task cost DeepSeek reports, 40 42 runs cost about $8 in inference. The review minutes will be the bigger line.
Step five: write a routing rule, one line per tag. Green goes to the cheapest model that clears your pass rate on the 20 06 tickets. Red goes only to models on your approved list. Commit the rule to the repo README.
Rerun the 20 06 tickets every time a vendor ships a checkpoint. DeepSeek's deepseek-flash endpoint changed models under the same name on October 17, according to Tech Insider. Your scorecard should catch the next silent swap before a customer does.
Map your models, then price merged changes
- Build a model provenance table. List every coding tool your team uses, with the model and the provider serving it. Ask vendors in writing when the docs are silent, and treat every blank cell as a finding.
- Run a 20-ticket 06 bake-off. Pull 20 closed tickets from your green repos and run each through OpenHands once per cleared model. Log tokens, retries, pass or fail, and human review minutes, including stalled or looping runs.
- Commit a routing rule. Send green repos to the cheapest model that clears your pass rate, and send red repos only to approved models. Put the rule in the repo README and rerun the 20 tickets whenever a vendor ships a new checkpoint.
Check the badge first, then price the merged change.
With two rivals at the same $0.15 13 input rate, the token price no longer separates them. The questions that do come in order: can you buy the model at all, and what does one accepted change cost once retries and review are counted. Our worked example shows a cheap model losing on cost per merged change, and the DHS order shows a cheap model becoming unsellable overnight. Teams that track provenance and their own pass rates will catch the next swap before a customer does.