One developer's all-day AI coding sessions now cost under $1

The Claim: Frontier Behavior at Pocket-Change Prices

In a post published October 7, a developer says he has used DeepSeek 4.1 Flash heavily for about a month across a dozen projects. His claim: mid-session, he often could not tell whether he was talking to DeepSeek or to Anthropic's Opus. His costs, by his own account, rarely exceed $1 per session, including sessions that run most of a day. He pays $10 a month for an OpenCode Go subscription, which he describes as 'basically unlimited' for this model.

This is one person's subjective experience, not a controlled evaluation. He points readers to separate benchmarks for a fuller comparison, and he says he still keeps frontier subscriptions through his employer. Treat it as a field report worth testing, not a verdict.

Why Long Sessions Get Expensive, and What Changed

When you work with a coding assistant for hours, the model has to keep the whole conversation available so it can refer back to earlier code and decisions. Internally, it stores a compressed record of everything it has already read, called the KV cache (key-value cache). That record lives in GPU memory, which is scarce and costly, and it grows with every turn. For long agentic sessions, holding it is one of the biggest line items.

The author says DeepSeek shrank this cache by roughly 437x compared with its V1 model, and credits that for his sub-dollar days. That figure is his; the source does not show the underlying measurement, so verify it before building a budget on it. He also says Opus 5.5 got its own efficiency boost from the same kind of caching work, though he offers no evidence for that.

The Playbook: What a Sensible Operator Does This Month

The author's own workflow is the most concrete thing in the source, and it suggests a pattern you can copy and test:

• Route by task, not by prestige. He uses the cheap model for planning, research, exploratory UI testing and chores like reorganizing files, which he says cost about $0.003 instead of $1.
• Buy a second opinion, not a default. For occasional critical work, he brings in Opus 5.5 for a final code review, which he says catches a few edge cases, then has DeepSeek apply the fixes.
• Keep sessions tight. Even at low prices, he tries to limit session scope.

The mistake would be the opposite of what he describes: either ignoring cheap models because the expensive ones feel safer, or swapping everything over on the strength of one blog post. Run your own workload through both and compare the output you would actually ship.

On self-hosting, his view is that the economics of 4.1 Flash mean you will never recoup the hardware costs if saving money is the goal. If privacy is the driver, he expects the cache optimizations to reach local setups eventually. He notes the model is technically self-hostable now, though not practically.

Questions You Should Be Asking

• Does a month of one developer's feel for 'no difference' hold up on our codebase, with blind comparisons of outputs rather than impressions?
• Where does the cheap model fail silently? The author still pulls in Opus 5.5 for edge cases, so what is our equivalent review step, and who owns it?
• What happens to our costs and workflow if a $10 subscription's 'basically unlimited' terms change?
• The author waves off the dispute over how these models were trained. Can our legal and procurement teams do the same, given that he says they were distilled from Claude?
• Is the 437x cache reduction independently verified, or is it a vendor-adjacent number we are repeating?

What To Watch Next

The signal that matters is whether independent benchmarks and other heavy users reproduce the 'can't tell the difference' experience on unattended, multi-hour tasks, and whether cache-shrinking techniques show up in locally runnable models, as the author expects. If both happen, the argument for paying frontier prices for routine work gets much harder to make.