Why I went back to Claude's 200K context window
Почему я вернулся на контекстное окно в 200К у Клода
B · Useful, limited Russian practitioner case on Habr with specific token, cache and file-size figures from the author's own multi-agent setup; useful practice, but a single self-reported project without controlled measurement.
Summary based on the full text·Original: about 9 minutes to read
In short
A Habr case post (Russian, 9-minute read, about 21K reach) by Egor Vasilyev (@bissquit), October 2026. Running a pet project on Claude Code, he went from a $100 to a $200 subscription and still hit limits, then traced the cost to convenience habits rather than task difficulty: long sessions, overnight batches and bloated instruction files that defeat prompt caching. His conclusion is to work again in smaller contexts, near the old 200K window, compact often and cut what agents read at startup.
Key numbers
- ~8xcost of the same actions in an 800K-token session versus a 100K-token session
- 150-200Ktokens a subagent could consume reading CLAUDE.md files before starting work
- 140 KBsize the top-level CLAUDE.md reached after months, from a 20-30 KB start
- ~90%compression of instruction files after keeping only information whose absence would break the project
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- By August the weekly limit ran out in about 3 days; friends hold top subscriptions to Claude and GPT at once.
- He sent batches of 5-15 tasks overnight; the lead agent finished on 500-800K tokens of its own context, subagents averaged 400-600K.
- Lead-agent cache lives 1 hour, subagent cache 5 minutes between tool calls; a subagent's TTL can be raised to 1 hour (subagentPromptCacheTtl) at higher cache price.
- A single bug fix takes 20-50 tool calls, each rereading the cache; an 800K session costs about 8 times a 100K one.
- Subagents start in the main session's directory, so they read both the top-level and repository CLAUDE.md: 150-200K tokens before work begins, with 20-50 such subagents a night.
- The top-level CLAUDE.md grew from 20-30 KB to 140 KB as each subagent logged its work; a "keep only what would break the project" rewrite cut about 90%.
- Agent-written comments drifted from code and formed a "second codebase" that later agents trusted over the code; he banned comments by default.
- Model routing: Sonnet for cheap fixes, Haiku for search, Opus or Fable for analysis and review.
How they got it
Personal operation of a multi-repository pet project with an analyst-developer-reviewer-QA subagent setup since Opus 4.5. Figures come from his own usage and cache behaviour; the 8x ratio is his inference that cache-read cost scales with context size.
Limits
- One project, one user, subscription-limit observations rather than metered bills.
- Model names and cache settings are tied to Anthropic's October 2026 product and will change.
- The claim that quality degrades with more context is stated, not measured.
Why it matters for companies
Agent cost is driven by context hygiene as much as by task volume: long sessions and large instruction files multiply spend on every tool call. Teams adopting coding agents need owners for shared instruction files, which agents will otherwise inflate. Stale generated comments are a maintenance risk that review processes should check. Budgeting by seat price alone misses these multipliers.
Key passages in translation
Russian original
Главная причина нашлась не в сложности выполняемых задач и объеме вычитываемого кода, а в банальном повседневном удобстве.
The main cause turned out to be neither the complexity of the tasks nor the volume of code read, but plain everyday convenience.
Russian original
Работа на сессии с 800К токенов примерно в 8 раз дороже, чем те же действия на 100К токенов.
Working in an 800K-token session is roughly 8 times more expensive than the same actions at 100K tokens.
Russian original
субагент на старте еще до начала работы мог съесть 150–200К токенов, просто чтобы понять где находится.
At startup, before doing any work, a subagent could eat 150-200K tokens just to figure out where it was.
Russian original
если не приводить в порядок комментарии, то их совокупность формирует вторую кодовую базу приложения
If you do not keep comments in order, together they form a second codebase of the application.
Russian original
комментария нет по умолчанию, он должен доказать право на существование.
No comment by default; a comment has to prove its right to exist.
Russian original
деградация качества работы самих агентов, которая тем сильнее, чем больше контекста израсходовано.
The degradation in the agents' own work quality, which grows stronger the more context has been used.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com
- Tags
- AI enablement