Source: raw/Why_Claude_Keeps_Hitting_Usage_Limits_how_to_fix_it.md (primary). Supporting: raw/Claude_Haiku_vs_Sonnet_vs_Opus_vs_Fable_Clearly_Explained_use_cases.md, raw/New_Claude_OPUS_4.7_Upgrade_Explained_for_non-technical_peeps.md, raw/Anthropic_Just_Dropped_Claude_Motion_Dashboards_+_More_AI_News.md (monthly API credit, October 2026). First-party confirmation of the credit: ai-research/anthropic-claude-haiku-5-5-announcement-2026-10-27.md (Haiku 5.5 launch post, fetched 2026-10-27). Fable 5.1-era additions from Ben AI: raw/Paste_This_Into_Claude_Never_Hit_a_Token_Limit_Again-Jr-jyTL2MYI.md, checked against ai-research/claude-docs-prompting-claude-opus-5-2026-10-28.md, ai-research/claude-docs-effort-2026-10-28.md and ai-research/claude-docs-prompting-best-practices-2026-10-28.md.
Creator: Eliot Prince URL: https://www.youtube.com/watch?v=vnocKrxasg4 Duration: 26:12 Platform: YouTube Published: 2026-04-03
Eliot Prince’s checklist for Claude.ai and Cowork users who keep hitting the five-hour session cap. He made it after Anthropic announced that peak-hour usage would drain sessions faster. He separates the two limits people confuse, shows how to see what is using up a chat’s context, and gives about fourteen fixes in four tiers: settings, chat habits, workspace setup, and timing. It is the non-coder counterpart to Claude Code Token Optimization. The video is from April 2026, when Opus 4.6 and Sonnet 4.6 were current and chat and Cowork were separate. The “What has changed since April” section below lists the parts that are out of date.
Key Takeaways
- There are two limits, and they behave differently. The usage limit is your plan’s allowance: a five-hour session that starts with your first message, plus a weekly cap. Both show under Settings → Usage. The length limit is the context window, meaning how much one chat can hold before Claude compacts it and loses detail. Most fixes reduce what each message carries, and that helps both.
- Every message re-sends everything before it. System prompt, custom instructions, connector and tool definitions, skills’ descriptions and the whole chat history are counted on each turn. Eliot measured about 30,000 of a 200,000-token window used before he had typed anything.
- Most of the cost is overhead. The biggest savings in his list are switching off tools you don’t use, keeping instructions short, starting fresh chats, and converting heavy files before upload, more than any single prompt trick.
- Peak hours drain sessions faster. Anthropic’s post (as Eliot reads it): during weekday peak hours, 5am–11am Pacific (1pm–7pm UK), you move through the five-hour session limit faster. Weekly limits are unchanged. This matches the 8am–2pm Eastern window recorded in the Claude Code article.
- Extra usage lets you finish a task instead of waiting. Adding a little credit lets a task that hits the session cap finish rather than stop halfway. Eliot reports spending about 8 GBP in a month this way.
The fixes, tier by tier
Tier 1 — Settings
- Watch Settings → Usage. Check the current-session percentage and its reset time, and the weekly percentage. Plan heavy work around both.
- Pick the model on purpose. Sonnet is his default for everyday work. Opus is for intensive tasks when you have allowance to spare. Haiku is not automatically the cheapest overall: on a big job, many small Haiku round-trips can cost as much as one Sonnet pass (his analogy is moving a crowd by motorbike instead of by train). Eliot also says Opus 4.6 “burns tokens five times faster” than Sonnet. Per-token list prices don’t show that gap; see the note at the end of this section.
- Run
/contextto see what is using your context. He shows a breakdown with system prompt about 20k tokens, MCP tools up to about 45k (partly deferred), plus memory files and skills. He then asks Claude “what is burning the most tokens in my chats?” - Switch off extended thinking when you don’t need it. He calls a forgotten extended-thinking toggle “the real killer,” because thinking adds steps and output on every turn. Also switch off web search and research when the task doesn’t need them.
Tier 2 — Chat habits
- One chat per task. Moving from spreadsheet analysis to marketing copy in the same thread drags the earlier context into every new message. After 20–30 messages the thread is messy and each message costs much more.
- Be specific from the first prompt. “Summarise the financial risks in section three of this PDF” reads less than “tell me about this document” followed by narrowing it down.
- Batch edits into one message. “Fix the typo, shorten the intro, add a CTA linking to the vault” in one prompt pays the per-message overhead once instead of three times.
- Disable idle connectors and duplicate tools. He found two Chrome-control tools running at once (Claude in Chrome and an older custom Chrome MCP), both counted on every message, and switched the old one off. Review the connector list under the + button. See Connectors.
Tier 3 — Workspace setup
- Convert PDFs to Markdown or plain text before heavy work. He says that for people who drop in many PDFs, extraction can take “up to 80% of a session.” He does the conversion in a different tool (Perplexity) so it doesn’t use his Claude allowance, then brings the Markdown in.
- Put reference files in a Project. In chat, Projects use retrieval, so only the relevant parts of the files enter the context instead of every file. Project instructions and project memory also save you re-explaining yourself in each chat. See Projects.
- Keep custom and project instructions under ~500 words. They load on every task, so a long instruction file adds overhead on every message.
- Remove files you don’t need from Projects and Cowork folders. Even if Claude doesn’t read them each time, they are clutter it may wander into.
- Turn repeated jobs into skills. Only a skill’s short description sits in context until the skill runs. The full steps load on demand, and Claude stops spending tokens working out a process it has run before. Keep descriptions short. See Agent Skills.
Tier 4 — Timing
- Move heavy work and scheduled tasks off-peak. For him in the UK that means doing heavy Claude work in the morning, before 1pm.
- Use the session-start trick. The five-hour window starts at your first prompt, not when you open the app. Send a short “let’s get started” message a few hours before a heavy session. When you come back, the window is close to resetting and a fresh one follows, so one sitting can use two windows. This is a creator tactic. It follows from how he describes the window, not from any Anthropic guidance.
On the "five times" figure
Eliot’s claim that Opus 4.6 uses allowance five times faster than Sonnet is not backed by the list prices in this wiki. Opus models from 4.7 to 5 were listed at 25 per million tokens. Today Opus 5.5 (20) costs exactly twice Sonnet 5.5 (10) per token. The real gap on a plan also depends on how many tokens each model uses per task and on how Anthropic weights each model against subscription limits, which it does not publish. Use 2× per token as the known floor and measure your own use.
What has changed since April 2026
- Chat and Cowork merged on 2026-09-16. There is no longer a mode to choose; Claude decides what a task needs (see Claude Cowork). The habits above apply to both kinds of work.
- Limits went up. With Opus 5.5 (2026-09-22), Anthropic raised five-hour limits on Pro, Max and Team and added a rate-limit reset you can save and use when you choose. Keeping that saved reset for a known heavy session does the same job as the session-start trick.
- Context windows grew. Eliot’s “most plans have 200,000 tokens per chat” was true in April. Sonnet 5 shipped with a native 1M-token window in June, and Opus 5.5 runs at 1M in Claude Code. A bigger window doesn’t make overhead cheaper: every message still re-sends the history.
- Thinking can’t be switched off on Opus 5.5. Tip 4 now applies only to models that still have a thinking toggle. On Opus 5.5, lower the effort setting instead (see Model vs. Effort).
- Model choice is still the biggest single setting. His later model guide tells the same story: his fiancee used up a Pro allowance quickly because Opus 4.8 was selected for invoices and emails. See Which Claude Model for Which Job.
- Max and Team plans now include a monthly API credit (creator-relayed, October 2026). Per a creator’s weekly news video, you get “200 a month on Max 20X, and up to $500 on Team.” You have to claim it under Settings → Billing, it “doesn’t roll over,” and Pro gets none (
raw/Anthropic_Just_Dropped_Claude_Motion_Dashboards_+_More_AI_News.md). Not yet checked against Anthropic’s billing docs. If it holds, it gives Max users a separate pool for API or SDK work that would otherwise use plan allowance. ^[inferred]- Confirmed 2026-10-27 by Anthropic’s Haiku 5.5 launch post: “a new monthly API credit to all Max and Team subscribers for use on the Claude Platform”. Max 5x: 200. Team: up to $500, pooled across users. It “can be used on any of our models” and is meant for “building tools, apps, and agents that call our API”, so it pays for Claude Platform API calls, not for chat, Cowork or Claude Code usage under your plan. ^[inferred] Rollout was the launch week (2026-10-07). Full terms: Help Center article 17154008 (
support.claude.com/en/articles/17154008), not fetched for this wiki. The claim-under-Billing, no-rollover and no-Pro-credit details are still creator-only.
- Confirmed 2026-10-27 by Anthropic’s Haiku 5.5 launch post: “a new monthly API credit to all Max and Team subscribers for use on the Claude Platform”. Max 5x: 200. Team: up to $500, pooled across users. It “can be used on any of our models” and is meant for “building tools, apps, and agents that call our API”, so it pays for Claude Platform API calls, not for chat, Cowork or Claude Code usage under your plan. ^[inferred] Rollout was the launch week (2026-10-07). Full terms: Help Center article 17154008 (
- Tokenizer changes can raise usage silently. In his Opus 4.7 video, Eliot reads the Claude API docs as saying the new tokenizer “might use 35% more tokens for the same fixed text.” So a model upgrade can increase usage even when the per-token price doesn’t change.
Fable 5.1-era fixes (Ben AI, September 2026)
Source: raw/Paste_This_Into_Claude_Never_Hit_a_Token_Limit_Again-Jr-jyTL2MYI.md · Creator: Ben AI · URL: https://www.youtube.com/watch?v=Jr-jyTL2MYI · Duration: 16:46 · Platform: YouTube · Uploaded: 2026-09-14. Added 2026-10-28. Recorded on Fable 5.1, a week before Opus 5.5.
Ben’s list overlaps Eliot’s (fresh chats, fewer connectors, one chat per task) and adds the parts that matter on a model that delegates heavily.
- Put Fable’s subagents on Sonnet. His biggest saving: one report that ran 67 Fable 5.1 subagents cost about 44 with an “almost identical” result. Details and a correction to his sourcing in Fable 5.1.
- Why not drop subagents altogether. Without them the main chat would read every transcript itself, fill its context window and work in what he calls the “dumb zone,” where answers get worse the more tokens a chat holds. Subagents return summaries, so the main chat does the analysis with room to spare.
- Overhead before your first prompt. His Claude Code tab showed about 84,000 tokens used before typing, against Eliot’s 30,000. Check yours with
/contextin the Code tab; he says it does not work in Cowork. - Disable, don’t delete. In Customize, switch off skills, plugins and connectors you don’t use daily. A plugin loads every skill inside it, so disable whole plugins too.
- One connector instead of many. He routes his apps through Composio, a single meta-connector, so Claude does not load each app’s tool definitions. His claim, not measured here; see Connectors.
- Plan the prompt so medium effort is enough. His four parts: the whole job, why it is needed, guardrails (what could go wrong, plus the subagent line), and what done looks like. A well-planned first prompt gets an 80% result that needs one or two edits, where a quick prompt plus many iterations re-reads the growing chat each time. He uses an “interview me” skill or a voice brain dump cleaned up by a “prompt master” skill when the task is unclear. The framework maps to Anthropic’s guidance in New Rules of Context Engineering.
- Drop to low effort for the final tweaks. Changing effort mid-chat keeps the prompt cache on Fable 5.1 (and, per the Effort docs, on Opus 5.5, Opus 5, Sonnet 5.5 and Haiku 5.5). See Which Effort Level to Use.
- Switching model mid-chat is a judgement call. He agrees a model switch re-reads the whole chat, but says it “could still be worth it” because Opus and Sonnet use far fewer tokens than Fable. Jay at Robo Nuggets relays the opposite advice for an Opus-to-Haiku switch at 100,000 tokens (see Which Claude Model for Which Job). ^[ambiguous] Measure it on your own chats before making it a habit.
- Start a new chat around 300,000 to 500,000 tokens, even mid-task, because output quality drops as the chat grows. (The captions say “three to four to 500 tokens”; thousands is the likely meaning.) ^[inferred]
/usageshows the spend;/compactsummarises in Claude Code but, he says, keeps too little and is not available in Cowork. His own/refreshskill writes a handoff prompt for a new chat. - Three prompt habits to drop:
- “Double-check your work” or “verify twice.” Confirmed: Anthropic’s Opus 5 guide says the model checks its own work and that such instructions “add cost without improving results.”
- Heavy emphasis (“be maximally thorough”, “YOU MUST”). Consistent with Anthropic’s guidance since Opus 4.5 that aggressive wording causes over-triggering.
- Fixed step-by-step processes, numbered lists and “think step-by-step.” He says Anthropic found dropping these cut tokens 15% and raised accuracy 5%. That figure is not in any Anthropic source saved here, and Anthropic’s general best-practices page still recommends numbered steps “when the order or completeness of steps matters.” ^[ambiguous] Drop “think step-by-step”; keep numbered steps where order matters.
Try It
- Take a
/contextsnapshot in a fresh chat with nothing typed. Write down the system-prompt, tool and skill figures. Switch off every connector you haven’t used this week and take the snapshot again. - Cut your custom instructions to 500 words or fewer. Move rarely used rules into a skill or a project file.
- Convert this week’s reference PDFs to Markdown before giving them to Claude. Compare session usage with last week’s equivalent task.
- Schedule recurring tasks outside 5am–11am Pacific on weekdays.
- Add the Sonnet-subagent line to your next big Fable prompt and compare the usage with a similar past task.
- Keep the saved rate-limit reset for the day you know will be heaviest, instead of using it on an ordinary day.
Open Questions
- Is the peak-hour drain still in force? The April 2026 announcement is the latest statement on it in this wiki. Whether the September limit increases kept, changed or removed the peak window has not been confirmed.
- How each model is weighted against plan limits. Anthropic publishes API prices, not how much plan allowance each model uses. The “5×” figure is a creator reading.
- Monthly API credit terms. The 200 / up-to-$500 figures for Max 5×, Max 20× and Team come from one creator’s video. Which products the credit can be spent on (API, Claude Code
--print, SDK) is not stated. Partly answered 2026-10-27: the Haiku 5.5 launch post confirms the amounts and says the credit is for the Claude Platform API, on any model. Whether a Claude Code session authenticated with an API key draws on it, whether it rolls over, and how to claim it are in Help Center article 17154008, which has not been fetched. /contexton claude.ai. The video shows it in the desktop app. This wiki hasn’t confirmed which surfaces support it after the merge.
Related
- Which Effort Level to Use — low effort for tweaks, high only for checks.
- Claude Fable 5.1 — where subagent cost is highest.
- Claude Code Token Optimization — the 18-technique version for Claude Code users.
- Which Claude Model for Which Job — the routing guide that does most of the saving.
- Model vs. Effort — what the effort setting costs and buys.
- Projects — retrieval over project files and project memory.
- Connectors — auditing what is connected.
- Agent Skills — why skills cost little until they are used.
- Kahn v. Anthropic — the class action over Max-plan usage limits.
- Image-as-Context Token Arbitrage — another route to cheaper context.
- Claude 5.5 Mistakes and Fixes — 23 habits that waste usage on the 5.5 models, each tagged by evidence.