Why isn't the industry freaking out about DeepSeek 4.1 Flash?

What it claims

One developer reports a month of heavy use of DeepSeek 4.1 Flash across about a dozen projects. Their verdict is that mid-session they cannot tell it apart from Opus 5.5 in conversation, output quality or speed. They use it for complex planning and research as well as execution. Two concrete cost figures:

• Typical spend: sessions rarely exceed $1 in expected cost, even ones that run most of a day.
• Trivial jobs: a task like reorganising desktop files costs about $0.003 on DeepSeek against roughly $1 on a frontier model.

The author attributes this to DeepSeek's KV-cache reduction of roughly 437x against its V1 model. They also assert, without a source, that Opus 5.5 got its own efficiency gain from similar cache techniques.

How to use it

The access route and workflow the author describes:

• Access: an OpenCode Go subscription at $10/month, which the author says makes DeepSeek 4.1 Flash "basically unlimited" for their usage.
• Default model: DeepSeek 4.1 Flash does everything, including planning, research, and cheap throwaway work such as exploratory UI monkey testing that the author would not spend frontier tokens on.
• Critical tasks: 1. DeepSeek does the work. 2. Opus 5.5 does a final code review, which "will catch a few edge cases." 3. The fixes go back to DeepSeek to implement.
• Why a second model: the author says pulling in Opus or GLM is "less about quality and capabilities and more about getting new eyes on a problem."

Self-hosting: the author says it is not worth it if saving money is the goal, because hosted pricing means you will never recoup the hardware cost. The model is "technically self-hostable, even if not practically so." For privacy-driven self-hosters, the advice is to wait for the cache optimisations to reach local inference. This is a prediction, not a current option.

Limits

• Anecdotal evidence: the assessment is subjective and comes from one developer. The page links to "more complete benchmarks" but reproduces no numbers, baselines or task types. Check those before relying on the parity claim.
• No access terms stated: OpenCode Go's rate limits and what "basically unlimited" means are not given.
• Dollar figures are not measured: the $1/session and $0.003 figures are "expected costs" from the author's usage, not metered breakdowns.
• Workload unknown: the projects and languages behind the experience are not described.
• Unsupported claims: the environmental claim (less water and electricity than Claude) is inferred from caching, not measured. The claims about distillation from Claude and about Opus 5.5's cache gains are also unsupported on the page.
• Possible bias in comparison: the author also has employer-provided Claude, Cursor and other subscriptions, so the frontier models are free to them. That may affect how they weigh the two.