Nobody is panicking about DeepSeek Flash because nobody benchmarks their own work

The article asks why the frontier labs are not freaking out about DeepSeek V4.1 Flash. I went looking for the reason and found it in DeepSeek’s own launch table. On Terminal-Bench 4.0, the hardest agentic coding benchmark either company publishes, DeepSeek reports 31.2 for Flash. Anthropic reports 66.4% for Claude Opus 5.5. Artificial Analysis, running its own harness, got Opus down to 59.6%. Take the independent number and Flash still lands at roughly half.

I started out on the author’s side. He says that mid-session he cannot tell Flash from Opus, and I believe him. But his list of work is file reorganising, exploratory UI monkey testing, planning and research. “Cannot tell” is a measurement of those tasks, taken by one person, with no control. On that kind of work a model scoring 31.2 on Terminal-Bench 4.0 and one scoring 60 may well be indistinguishable. On the tasks that benchmark was built to capture, I think they plainly are not.

DeepSeek has also chosen its showcase numbers well. Flash scores 90.6 on Terminal-Bench 2.1, an older version that the field has largely saturated, and 90.9 on GPQA Diamond. Those figures make it look like a frontier model on the tests the frontier has moved past. The 4.0 score is in the same table, to DeepSeek’s credit, but nobody quoting the launch leads with it.

The price gap is real and enormous. OpenRouter lists Flash at $0.0184 per million input tokens and $1 per million output tokens. Opus 5.5 costs $4 and $20. That makes Flash 20 times cheaper on output and cheaper still by a wide margin on input, and long agent sessions spend most of their tokens rereading context, so input is the column that matters. The engineering behind it is impressive: a 552B-parameter mixture-of-experts that activates 16B parameters per token in decode and 8B in prefill, with a global KV cache of 890 bytes per token, a quarter of V4-Flash’s. The author’s “437x” cache figure is a comparison with DeepSeek’s own V1, so his leap to Flash using less water and electricity than Claude rests on a number that never mentions Claude.

So here is what I think the labs’ calm amounts to. Anthropic’s revenue depends on customers sending most of their work to its most expensive model by default. Anthropic already sells Haiku 5.5 at $0.10 input and $0.50 output per million tokens, which is cheaper than Flash on output. It has had a cheap tier all along. The threat from DeepSeek is that it might teach developers to route: to sort their tasks by difficulty and pay frontier prices only for the hard ones. The labs are betting that almost nobody will build the evaluation needed to do that sorting, because writing an eval suite for your own codebase is tedious work that ships no features.

The author’s own workflow looks like a rebuttal. He runs Flash for nearly everything on a $10 monthly subscription and calls in Opus for a final review on critical tasks, where it catches a few edge cases that Flash then fixes. That is routing, and it already takes money out of Anthropic’s pocket. But look at how he decides what counts as critical: gut feel, informed by a month of use across a dozen projects. The Opus review catching edge cases is exactly the gap Terminal-Bench 4.0 measures, showing up in his own sessions. He has not measured how often Flash would have shipped those bugs on the tasks he did not consider critical. Neither has anyone reading his post and switching tomorrow.

The labs also have time on their side in one direction and not the other. Opus 5.5 is billed as 40% cheaper than Opus 5 on typical workloads, so the frontier is getting cheaper. Flash, meanwhile, ships as open weights under an MIT licence and is served by 27 providers on OpenRouter, so its price will be set by competition among hosts rather than by DeepSeek. I expect the cheap end to keep falling faster than the expensive end, and Anthropic’s bet to hold only for as long as its customers keep guessing which tasks need Opus.

You can stop guessing this week. Pull a few dozen recent agent tasks your team completed with Opus or another frontier model, replay each prompt and starting commit through Flash, and run your existing test suite against the output. Count the passes. At $1 per million output tokens, the whole replay will cost less than a single heavy Opus session, and the pass count tells you what share of your frontier bill is paying for work a model scoring 31.2 can already do.

Prompted by Why Isn't The Industry Freaking Out About DeepSeek 4.1 Flash?, dgt.is.