I’m using Grok Bot on a Pro+ plan. I noticed that when using custom bots for discussion/consultation sessions (text-only prompt engineering, no codebase editing or agent file creation), the token accumulation scales up extremely fast. A single prompt run can register 249K+ tokens behind the scenes.
I understand that the conversation context builds up with every turn, but for a text-based consultation, this burns through the usage pool very quickly.
Is there a recommended workflow to compact or prune the context window during a long session with a custom bot without losing its core prompt instructions? I’d love to know how other users manage their token consumption during theoretical design meetings with their bots.
Hey @mohsendev! This is expected: every reply re-reads the bot’s whole conversation, so each usage row is roughly the conversation size (mostly cached tokens), not your message. It only auto-summarizes near the model’s limit, and there’s no manual compact yet (it’s a request we’re tracking).
To reset cost mid-session: ask the bot for a short handoff summary, then right-click it in the sidebar → Duplicate. The copy keeps its name, instructions, routines and avatar but starts with an empty conversation. Paste the summary as its first message and keep core instructions in Edit Profile (not chat) so they always carry over.
Thank you, @mohitjain , for the detailed breakdown and the brilliant “Duplicate” workaround!
I must be the luckiest developer alive for capping my limit at $15. If I hadn’t, my bots would have probably charged me enough to fund the next SpaceX rocket launch just for saying “hello”!
Honestly, I absolutely love Grok Bot—it is a truly amazing and fantastic concept. However, if the pricing continues to be this expensive, it is just going to be unsustainable.
While your duplicate trick is an awesome lifesaver, Mohit, paying nearly $14 for a brief text consultation still feels like buying a digital cup of coffee at airport VIP prices. I’ll definitely adopt your routine to survive this token gold rush, and I really hope the team addresses this aggressive pricing logic soon. Appreciate your help!
@mohitjain I decided to pull my actual usage logs and run some data analysis to see what kind of numbers we are dealing with behind the scenes.
The results are both fascinating and absolutely terrifying!
I mapped out how the context window blots recursively per turn. Here is the visual proof of the digital VIP airport coffee machine in action:
Context Window Bloat (The Actual Snowball Effect)
As you mentioned, every single line reads the whole history. Look at how my tokens scaled vertically from a cute 10.1K to a heavy 254.6K in just 7 quick turns, even though my later prompts were basically text modifications or short check-ins!
The Hyper-Scary 15-Turn Trajectory Forecast
Out of pure mathematical curiosity, I simulated what would happen if I had blindly continued the discussion for 15 turns on the exact same extended topic.
Spoiler alert: By turn 13, it shatters the 1.1 Million token ceiling (which is like making the engine read a 400-page book for a 1-word answer). By turn 15, we land at a monstrous 1.8 Million tokens!
What this means financially:
If I hadn’t manually capped my budget at $15, that single 15-turn session would have easily cost me around $154. If I had actively continued this loop for a full month (30 days) of uninterrupted consults without a hard reset, the cloud computing bill would have easily cleared $4,500+ (enough to buy a premium used car or 4 MacBook Pros).
I’m definitely going to use your Duplicate + Handoff summary trick to cheat this gold rush. Since my weekly limit is currently maxed out, I will wait for the upcoming weekly reset to run my tests and make sure I don’t run into this pricing trap again. I really hope the team introduces a manual ‘Prune/Compact Context’ button soon. Paying surge-pricing rates just because the bot remembers everything we said 5 minutes ago is wild!
Hi everyone, and special thanks to @mohitjain for the workaround recommendation.
I’ve spent the last few hours running rigorous, isolated benchmark tests on my Pro+ account (starting from a clean 0% weekly reset) to track exactly where our token quotas are disappearing. The results are eye-opening and fully expose the architectural flaws of the current system. Here is the data-driven breakdown of what is actually happening behind the scenes:
1. The Multi-Bot “Fleet Tax” is Exponential (and Destructive)
In group chats (Multi-Bot rooms), even in a brand new, empty room with zero message history, typing a single short prompt like “Where did we leave off?” immediately injected 149.4K tokens on the first turn! Within 5 minutes of sitting idle, background loops/routines between the bots caused a massive, silent text dump, cycling from 130K to 273K tokens repeatedly without a single keystroke from my end. The system forces a massive “context tax” by blindly injecting the identity blueprints and skill catalogs of ALL bots in the room simultaneously, hitting what I am highly confident is a hard 300K limit/cap (though further official verification is needed) almost instantly.
2. Testing the @mohitjain Workaround (Isolated 1:1 Tests)
I isolated my quality bot to a clean 1:1 conversation to test the Duplicate + Handoff Summary workflow:
The Original Room Baseline: Stood heavily at 158.8K input tokens due to the bot being continuously hooked to read a large 16,000-line repository.
The Duplicate Room (Turn 1): After dumping the comprehensive text summary into the freshly cloned room, it successfully smashed the historical bloat, plunging the baseline down to just 35.7K tokens! A massive 75% token reduction. The workaround absolutely works to clear operational debt.
3. The Catch: The “Lazy Loading” Background Loop
However, the most fascinating discovery came on Turn 2 in the newly duplicated room. The moment I drilled down into specific code implementation details (asking for Jenkinsfile and package.json modifications), a massive background injection loop triggered automatically. Over the span of two minutes, the active input size crawled vertically: 38K → 43K → 49K → 57K → 61K → 67K tokens.
This proves the engine uses a Lazy-Loading Context Injection mechanism. Even if you wipe the chat history clean via duplication, the very instant your prompts look for deep code assistance, the backend starts aggressively pulling and parsing secondary project files under the hood to construct the response, rapidly rebuilding the snowball effect.
Conclusion & Strategic Workflow:
This massive overhead means we are burning through our token quotas at an unsustainable rate for simple text consultations, primarily because the platform severely lacks an on-demand skills architecture or indexed file-filtering.
Until a native Compact / New Session button is officially shipped to prune text history dynamically, the only sustainable way to keep our quotas alive is The 4-Turn Duplication Rule:
Completely abandon multi-bot group chats for long code reasoning.
Conduct deep work in 1:1 rooms, allowing the bot to lazy-load what it needs for 3 to 4 turns.
Once the active token count starts creeping past 120K+, pull a concise 5-line Handoff summary, duplicate the bot, and start fresh.
I hope the dev team looks closely at these logs to optimize how project codebases are injected during deep chat sessions. Looking forward to a permanent fix!
Hey @mohitjain , this may be expected, but it makes the platform effectively unusable for any serious work. Unless you can find a way to optimise token spend and/or use smaller models, this product will be dead in the water. I love Grok Bot and find it really polished, but I’ve had to start using OpenMausBot instead despite being on the USD$300 SuperGrok Heavy plan because I literally watch the usage counter creep up in real time just doing a single bot task.
@mohitjain Please could you let me know how destructive it is to duplicate each bot’s chat and delete the original? In theory there’s nothing in their transcripts that I haven’t told already told the bots to store in either their individual memory or the shared memory.
I have 14 bots. What’s the quickest way to duplicate all of them, send them all their summaries, and delete their originals please? Can my Chief of Staff take care of this in 1 prompt somehow?
I am slightly cautious to duplicate my bots because I am worried about its potential for destructiveness… but I am hopeful that this could be a good solution to the weekly usage burn rate. I will await your reply before I try it.
@DannyB before you delete anything: a duplicated bot keeps its name, instructions, routines and avatar, but starts with an empty conversation and empty memory. Memory does not carry over, so the copy won’t know what the original knew, even facts you had it save. Deleting is permanent and removes the transcript and memory, so hide the original instead of deleting until you’re sure. Per bot: ask it for a short handoff summary, Duplicate it, paste that summary as the clone’s first message and re-teach any key facts, then hide (not delete) the original. There’s no one-prompt way to do all 14. Duplicate, hide and delete are manual sidebar actions and bots can’t duplicate or delete other bots, so your Chief of Staff can draft the summaries but the duplicating and hiding you do yourself.
@mohsendev@StevePritchard on the burn rate: today every turn re-reads the whole conversation and it only auto-summarizes near the model’s limit, so long sessions get costly. There’s no manual compact yet. Reducing how much context each turn costs is something we’re tracking. No timeline to share, but I’ll post here when there’s an update. Until then the Duplicate-and-handoff flow above is the way to keep a session’s cost down.
Have we lost the ability to duplicate bots in the latest Windows release? It’s no longer in the right-click menu for a bot. If it’s been deprecated, what should we do now?
