How to stretch your AI plan limits

Same job, with and without the tips

One Claude Pro session, one 40-page brief, seven questions. Left: the brief is re-attached with every question and the five-hour session runs out before the last one. Right: the brief goes into a Project once and the questions go in one batched message, following Anthropic's advice to combine related questions and refer back instead of repeating.

Anthropic's own list for the five-hour session and the weekly cap. Every message re-reads the whole conversation, so the article is mostly about sending fewer, fuller messages.

1. 1.Batch related questions into one message instead of a back-and-forth.

   “Can you combine multiple related questions into a single message?”

2. 2.Be specific up front so Claude does not need clarifying rounds.

   “Provide clear, detailed instructions or questions in each message.”

3. 3.Refer back to earlier context instead of pasting it again.

   “Refer back to previous information instead of repeating it.”

4. 4.Put documents you will reuse into a Project; cached content counts less when reused.

   “When you upload documents to a project, they're cached for future use.”

5. 5.After a long break the cache has expired, so the first message counts the project content in full again.

   “Caches expire after a period of inactivity.”

6. 6.For code reviews, send the complete snippet and your environment in one message.

   “Include entire relevant code snippets in one message for reviews or debugging.”

7. 7.For editing, send the whole text at once rather than in pieces.

   “Send entire texts for editing in one message rather than breaking them up.”

8. 8.Watch the two bars under Settings > Usage: current five-hour session and weekly.

   “Settings > Usage”

Source: Usage limit best practices (Claude Help Center), checked October 31, 2026. Not on that page: The article lists message length, attachment size, conversation length, tool use, model choice, effort level and artifact creation as factors, but gives no numbers for any of them.

Claude Code sends the full conversation with every request, so a one-line question in a session that has been open all day still draws usage for the whole history. The docs' advice is about keeping that context small.

1. 1.Clear between unrelated tasks with /clear; stale context costs tokens on every later message.

   “Use /clear to start fresh when switching to unrelated work. Stale context wastes tokens on every subsequent message.”

2. 2.Tell /compact what to keep instead of letting it summarise everything.

   “/compact Focus on code samples and API usage”

3. 3.Use Sonnet for most work and keep Opus for architecture or multi-step reasoning; switch with /model.

   “Sonnet handles most coding tasks well and costs less than Opus.”

4. 4.Write specific prompts; vague ones trigger broad scanning of the codebase.

   “Vague requests like "improve this codebase" trigger broad scanning.”

5. 5.Keep CLAUDE.md short (the docs suggest under 200 lines) and move workflow instructions into skills that load on demand.

   “Aim to keep CLAUDE.md under 200 lines by including only essentials.”

6. 6.Disable MCP servers you are not using (/mcp) and prefer CLI tools such as gh or aws.

   “Run /mcp to see configured servers and disable any you're not actively using.”

7. 7.Lower the effort level (/effort) for simple tasks; thinking tokens are billed as output.

   “For simpler tasks where deep reasoning isn't needed, you can reduce costs by lowering the effort level”

8. 8.Use plan mode (Shift+Tab) before big changes and stop early with Escape when the direction is wrong.

   “preventing expensive re-work when the initial direction is wrong”

9. 9.Send verbose jobs (tests, logs, docs fetches) to subagents so only a summary returns to your main context.

   “Delegate these to subagents so the verbose output stays in the subagent's context while only a summary returns”

10. 10.Run /usage: on Pro, Max, Ultra and Team it shows what is eating your plan limits and recommends fixes.

    “Its breakdown recommends a change for behaviors behind 10% or more of your recent usage.”

Source: Manage costs effectively (Claude Code docs), checked October 31, 2026. Not on that page: The page's dollar figures ($13 per developer per active day on average) are for API billing, not for subscription plans.

OpenAI publishes message ranges per five hours, not fixed counts, and says what makes a message cost more. The levers are mode, model, task size and where the task runs.

1. 1.Standard mode is the cheapest: Fast uses 2.5x the included allowance and Ultrafast 8x.

   “Fast and Ultrafast modes use included subscription limits and paid credits at different rates”

2. 2.Cloud tasks can use more of the allowance than local messages, and both share the same pool.

   “Cloud tasks may use more of your allowance than local messages.”

3. 3.Smaller, routine tasks cost a fraction of a large multi-file job.

   “Small scripts or routine functions may consume only a fraction of your allowance”

4. 4.Pick the model for the task; the published Plus ranges per five hours run from 5-45 messages on GPT-6 Astra to 350-3,000 on GPT-6 Luna.

   “The number of messages you can send depends on the model used”

5. 5.Image generation draws on the same limits, 3-5x faster on average.

   “Image generations use included limits 3-5x faster on average”

6. 6.Check remaining limits in the usage dashboard, or with /status inside Codex CLI.

   “If you want to see your remaining limits during an active Codex CLI session, you can use /status.”

Source: Codex pricing (OpenAI developer docs), checked October 31, 2026. Not on that page: The page does not describe how the five-hour or weekly windows reset; it points to the usage dashboard. OpenAI's 29 Oct 2026 announcement about doubled Pro limits and a shared ChatGPT+Codex allowance from December is not yet reflected on this page.

On Cursor the lever is the model picker: included usage is one pool for Cursor models, and third-party models or Max Mode bill differently.

1. 1.Model choice decides how fast the included pool drains.

   “your model selection affects how quickly your included usage is consumed.”

2. 2.Stay on models that draw from the included pool (the page names Grok 4.7, 4.6, 4.5 and Composer 2.5); third-party models bill from prepaid credit.

   “Third-party models aren't included.”

3. 3.Max Mode is not part of current plans; on legacy request plans it bills at the model's API rate plus 20%.

   “Max Mode is billed at the model's API rate plus 20%.”

4. 4.Composer 2.5 requests carry no Cursor Token Rate on any plan.

   “Composer 2.5 requests carry no Cursor Token Rate on any plan.”

5. 5.Unused included usage does not roll over, so spread heavy work across the week rather than saving it.

   “Unused included usage does not roll over on either kind of plan.”

6. 6.Watch the Usage page: it shows the weekly pool used, next reset and prepaid credit.

   “It shows how much of your weekly pool you've used, the next reset date, and your prepaid credit.”

Source: Usage limits (Cursor Help), checked October 31, 2026. Not on that page: The page does not say how many requests or tokens a given task consumes.

Copilot meters premium requests monthly; inline completions are unlimited on paid plans. GitHub's own optimisation list is short.

1. 1.Choose the model for the task; each model has a premium request multiplier.

   “If you're using a premium request, you can strategically choose which model you use to get the best result from Copilot.”

2. 2.Do not resubmit the same long prompt; rephrase or simplify instead.

   “Submitting the same long or complex prompt multiple times may use more premium requests.”

3. 3.Set a budget and get alerts at 75%, 90% and 100%.

   “Set a budget to track your overages and receive alerts when you reach 75%, 90%, or 100% of your budget.”

4. 4.In agentic features only the prompts you send count; tool calls do not.

   “only the prompts you send count as premium requests”

5. 5.Code review is the most expensive feature listed: 13 premium requests per review.

   “13 premium requests are consumed.”

6. 6.Counters reset on the 1st of each month at 00:00 UTC; check usage from the Copilot icon in your IDE or github.com/settings/billing.

   “Premium request counters reset on the 1st of each month at 00:00:00 UTC.”

Source: Monitoring your Copilot usage and entitlements (GitHub Docs), checked October 31, 2026. Not on that page: The model multiplier table lives on a separate page and is not copied here.

Common questions

Why does one short message use so much of a Claude or Codex session?

Both vendors say usage depends on the whole conversation, not the new message alone. Anthropic lists conversation length and attachment size among the factors that affect limits, and the Claude Code docs say the full conversation is sent with every request. Long threads and re-attached files are the usual cause.

Does uploading a document to a Claude Project save usage?

Anthropic says documents uploaded to a project are cached for future use and that caches expire after a period of inactivity, so the saving applies while you keep working and disappears after a long break.

Which Codex mode uses the least allowance?

Standard. OpenAI states that Fast mode uses 2.5x the included allowance and Ultrafast 8x on GPT-6 Astra and GPT-6.1 Sol, and that image generation draws on the same limits 3-5x faster on average.

Are these tips measured by StackVersus?

No. Every tip quotes the vendor’s own page as of the check date. The animation on this page is an illustration of that advice; the bar movement is schematic and not a measurement.