On September 10, DeepSeek released V4.1 Flash under an MIT license: a million-token context window, 211 output tokens per second, and a price list — $0.15 per million input tokens, $0.60 per million output — that makes comparison shopping look like a rounding exercise. Blended against Grok-4.1, it lands roughly 18 times cheaper. Hours after the weights went up, someone had posted dealignai/DeepSeek-V4.1-Flash-UNCENSORED-FP8 to HuggingFace, a build with the refusals stripped out, no signature from the lab. It has since cleared 2,254 downloads. And the industry, as this week’s 494-point thread demanded to know, did not freak out.

That’s the wrong question. Freaking out is something you do about a competitor. DeepSeek’s September was mostly a story about DeepSeek.

The September 14 Changelog Is the Whole Story

Four days after the launch, DeepSeek pushed a changelog line that should have been the loudest thing in the room. From noon Beijing time on September 14, requests to deepseek-v4-pro would be routed to V4.1 Flash and billed at V4.1 Flash prices — “until V4.1 Pro is released in the future.” In business terms: the lab informed paying customers that its flagship is now its mid-tier, and that they’d be charged the cheaper rate for the demotion. By DeepSeek’s own testing, Flash beats Pro on cost, speed, and total completion time.

This is the same company that warned developers on August 6 that a significant API price increase was coming, without disclosing the rate or the date. Five weeks later, the increase arrived — and it was pointed at DeepSeek’s own product line. That’s the event. Not a rival’s benchmark score, not another headline about Chinese AI leapfrogging anyone: a vendor repricing its own catalog, downward, on four days’ notice.

Set that next to the numbers that actually got measured. AutomationBench-AA at 68.9%, Terminal-Bench 4.0 at 26.8%, Humanity’s Last Exam at 39.2%. Very good. Not epochal.

$0.15 Per Million Tokens, Re-Keyed Twice in Five Weeks

Cheap tokens are wonderful. Cheap tokens from a supplier that repriced itself twice in five weeks are a different instrument than cheap tokens from a supplier that didn’t, and the procurement departments are only now working out that they were sold the first while budgeting for the second.

A platform engineer at a mid-size logistics firm, in a Slack message to her own team once the changelog went up: “We spent three days benchmarking Flash and about three minutes deciding we’d never be able to forecast it.”

She’s right, and the reason has nothing to do with where the lab is headquartered. A company willing to cannibalize its own flagship inside a week is telling you something about its unit economics, and the something is not “efficiency.” The number that belongs in the board deck is no longer dollars per million tokens. It’s variance. Anyone who built a three-year AI budget on the assumption that frontier prices fall tenfold a year got the price right and the counterparty risk wrong — a memo that will get written in a lot of CFO offices this quarter, and not one of them will cite Hacker News.

The Freak-Out Was a Product, and It Got Repriced Too

Here’s the part that should make the usual crowd uncomfortable. The freak-out was never a spontaneous reaction to a model. It was a genre with a business model — a schedulable Sputnik moment, useful for fundraising decks, procurement urgency, and a certain kind of op-ed. It worked because every release arrived with a story attached: somebody, somewhere, has been blindsided.

This one wasn’t aimed at us. It was aimed at DeepSeek’s own price list. When the thing that lands threatens nobody you can name, the ritual has nothing to grip. The absence of panic isn’t complacency or cope — it’s that a changelog outperformed a crisis narrative, and the people who write budgets noticed before the people who write threads did.

Note what else went unremarked. The MIT license meant the abliterated builds could appear the same day — 2,254 downloads of a version the lab never signed, alongside a control-vector build from a security researcher that strips refusals at inference time rather than in the weights. Nobody freaked out about that either. Everyone in the room understands that once you choose MIT, the safety work is a courtesy rather than a control, and they have understood it for two years. That is the one silence in this story that deserves noise.

So no, the industry isn’t freaking out about DeepSeek 4.1 Flash. It’s doing something less flattering and more adult: reading the changelog, marking down the vendor’s durability right next to its price, and waiting to see what the next one says.

Sources