For developers calling Anthropic’s Messages API directly, the mechanics described by MakeUseOf are substantially correct. The API is stateless: the client supplies the previous conversation turns with each request, and Claude generates the next response from that supplied context. A 50-turn exchange can therefore become costly if the application blindly resends every prior message, tool result, attachment, and assistant response at standard input-token rates.
But Claude.ai, Claude Code, and the API are different products with different controls. Treating them as though every word of every old turn is always re-read, billed at the same rate, and retained unchanged leaves out the features that decide whether a long session is actually expensive or merely long.
The API bill is real, but cached context changes the arithmetic
Anthropic’s Messages API documentation explicitly calls the API stateless. If an in-house Windows administration tool sends a user’s entire troubleshooting session on every request—event logs, PowerShell output, deployment notes, old responses, and all—it will keep presenting that material as input context. On a pay-as-you-go API account, that is a direct source of token charges.
MakeUseOf’s strongest point is that repeated clarification has a cumulative cost. A helpdesk workflow that begins with “Why did this Windows 11 update fail?” and then spends 20 turns uncovering the build number, error code, WSUS status, device-management policy, prior remediation steps, and desired output format is paying for a poor intake process. The first request was missing the information needed to do the job.
Yet the article’s simplified “you pay to reprocess the previous 99 messages every time” framing misses Anthropic’s prompt-caching option. Anthropic prices cache reads at a fraction of base input-token pricing after a cached prefix has been written. For API builders, that changes the engineering answer: stable material such as organizational instructions, a PowerShell style guide, a software inventory schema, or a large policy document should sit in a cacheable prefix. The changing task, user question, and fresh tool results should be appended afterward.
That does not make long conversations free. Cache writes have their own cost, caches expire, and any change to early content can invalidate a useful cached prefix. But it does mean that “start a new chat and paste the whole brief again” may cost more than preserving an efficiently cached, well-structured session. The correct optimization is to separate stable context from per-turn state, then cache the former.
For an IT team using Claude through an API, the most expensive pattern is usually not a conversation with follow-up questions. It is an application that continually resends a growing unfiltered transcript, including obsolete tool output and files the model no longer needs.
Claude’s chat products do not preserve every turn unchanged forever
The MakeUseOf article describes context as a pile in which “nothing leaves.” Anthropic’s own support documentation contradicts that as a general description of Claude’s consumer and workplace products.
Claude says that, when code execution is enabled, it automatically manages long conversations by summarizing earlier messages as a chat approaches the context window limit. Claude Code likewise offers
/compactto replace a long exchange with a shorter recap, and it can auto-compact as context fills. The original chat history may remain visible to the user, but a visible transcript is not the same thing as the exact token sequence currently supplied to the model.
This distinction matters. A user can see months of discussion in a Claude interface and reasonably assume Claude has every detail in active working context. It may instead be operating from a generated summary plus selected recent material. Summaries keep a session moving, but they can drop nuance, distort a decision made earlier, or omit a configuration detail that becomes important later.
Anthropic also documents chat search and memory features that use summaries and retrieval to bring earlier information into later work. Projects can use retrieval-augmented generation, or RAG, to load relevant items from a knowledge base rather than force every uploaded document into every prompt. Those systems are designed precisely because past material should not automatically become active context merely because it exists in a chat history or project folder.
The important correction is simple: Claude may retain history, summarize it, search it, or retrieve from it, depending on the product and settings. History storage is not active context. Windows administrators should not mistake a persistent sidebar conversation for a guarantee that a buried KB number, Intune policy name, registry value, or previous test result is still prominent in Claude’s working set.
“Frontload everything” can create the same problem it claims to solve
A detailed first prompt is often better than a vague first prompt. If a user needs Claude to draft a PowerShell remediation script, providing the target Windows versions, execution environment, required logging, rollback expectations, known error output, prohibited actions, and desired output format will usually reduce wasted turns.
That is good task specification. It is not a blanket instruction to attach every file, policy, log bundle, screenshot, and historical discussion “just in case.”
Anthropic’s recent context-engineering guidance warns that more context does not automatically improve recall. As context grows, models can become less effective at finding and using the information that matters. The industry shorthand is context rot: the model has enough room to hold a great deal of material, but its ability to retrieve the decisive detail from that mass declines.
The source article is therefore correct to warn against a chat that has wandered across unrelated tasks for hours. A troubleshooting thread that began with Microsoft 365 licensing, moved to Hyper-V networking, then switched to a Windows 11 driver rollback has accumulated context that may be irrelevant or harmful to the final task. Starting a clean conversation is sensible.
But a giant opening prompt can also bury the goal. Pasting a full 3,000-line
setupact.log, a complete Group Policy export, 40 pages of internal standards, and several unrelated event-log channels into a first message does not give Claude a clean brief. It gives the model a large haystack before it knows which needle to find.
The better sequence is:
- State the objective, environment, hard constraints, and desired deliverable at the outset.
- Include the small set of facts that change the answer, such as Windows build, management platform, error code, reproduction steps, and what has already been tried.
- Attach or retrieve supporting records when the task calls for them, and identify exactly what Claude should inspect in each record.
- Start a fresh task thread when the prior conversation’s decisions and evidence no longer bear on the work.
In other words, frontload the brief, not the entire archive.
Claude Code makes task boundaries a cost control
Anthropic’s Claude Code documentation is unusually direct on this point. It identifies
/clearas the strongest lever for cost and quality when old chat history is no longer useful, while
/compactis intended for a continuing task that needs a concise handoff from prior work.
That maps well to how Windows work actually proceeds. Investigating a failed Microsoft Intune deployment may require a long session: inspect a Win32 app detection rule, compare device logs, review installer behavior, adjust a remediation script, and test the result. Compacting preserves the current working state without dragging every exploratory detour into the next stage.
Once the script is complete, moving to an unrelated task—say, auditing local administrator membership across endpoints—should be a new session. Carrying the Intune investigation forward creates noise and consumes context budget for no operational reason.
Claude Code also separates persistent project guidance from transient chat context through
CLAUDE.mdfiles. That gives teams a better place for durable instructions: naming conventions, approved modules, coding standards, tenant safety rules, test requirements, and change-control expectations. Put the rules that should apply repeatedly into maintained project instructions. Put the current ticket’s evidence and constraints into the current task. Clear the task conversation when the ticket is done.
This is more disciplined than repeatedly pasting the same standards into a new chat, and safer than hoping an ever-growing transcript will remember them accurately.
Subscription users should not confuse usage caps with API invoices
MakeUseOf also blends two distinct costs: literal API billing and subscription usage limits. API customers pay by token according to the model and feature pricing in effect for their account. Claude Pro, Max, Team, and Enterprise users generally encounter usage allowances that reset over time, with consumption affected by model choice, conversation length, tool use, extended thinking, and effort settings.
A long chat can indeed consume more of that allowance. Anthropic says as much. But an individual subscriber cannot calculate a precise dollar saving from one fewer follow-up in the way an API developer can calculate input-token charges. Anthropic does not publish a universal conversion from a Claude.ai message to a fixed dollar cost, because product usage limits are affected by multiple variables and may change by plan and demand.
The more useful subscriber advice is operational: turn off extended thinking for routine work, lower effort when a short answer will do, disable connectors and tools that are irrelevant to the task, keep Project instructions concise, and do not load unused project files. Those are Anthropic’s own current recommendations for preserving both context capacity and usage allowance.
For paid chat users, the “one trick” is therefore less glamorous than MakeUseOf suggests. Write a competent opening brief, keep sessions task-specific, and deliberately compact or restart when a thread has accumulated dead weight. For API owners, add prompt caching and transcript pruning to that list—because those are the controls that change an actual bill.
References
- Primary source: MakeUseOf
Published: August 7, 2026 at 12:00 PM UTC
You're probably overpaying Claude without realizing it — this one trick changes everything
No, it's not a settings toggle
www.makeuseof.com
- Related coverage: platform.claude.com
Context engineering: memory, compaction, and tool clearing | Claude Cookbook
Compare context engineering strategies for long-running agents and learn when each applies, what it costs, and how they compose.platform.claude.com - Related coverage: docs.anthropic.com
Pricing - Claude Platform Docs
Learn about Anthropic's pricing structure for models and featuresdocs.anthropic.com - Related coverage: tomsguide.com
Anthropic looks to beat GPT-5 and Grok 4 with this one major upgrade | Tom's Guide
Anthropic makes major changes to its context windows to allow for even longer promptswww.tomsguide.com - Related coverage: itpro.com
Anthropic reveals Claude Opus 4.6, an enterprise-focused model with 1 million token context window for extended code capabilities | IT Pro
The AI developer highlighted financial and legal tasks, as well as agent tool use, as particular strengths for the new modelwww.itpro.com - Related coverage: support.claude.com
How do usage and length limits work? | Claude Help Center
support.claude.com
- Related coverage: support.claude.com
Get started with Claude | Claude Help Center
- Related coverage: platform.claude.com
Create a Message - Claude API Reference
Send a structured list of input messages with text and/or image content, and the model will generate the next message in the conversation. The Messages…platform.claude.com