Getting More Out of Your Usage Limits.
How to do more, not less, with the new Fair Usage Policy.
Since April 2026, Langdock no longer counts messages but costs. That sounds like a restriction, but it is the opposite: those who have model selection, chat hygiene, and workflows under control get noticeably more done with the new system. Increased limits run through May 17. These weeks are the best opportunity to calibrate your workflow properly before the standard limit kicks in.
What Has Changed
Since April 2026, Langdock no longer counts messages but costs. Each chat turn consumes a different amount of your limit depending on the model: a Haiku call draws little, an Opus call with Reasoning considerably more.
The background: new models like Opus 4.7 or Sonnet 4.6 write entire documents, build presentations, conduct multi-step research. A single message today often triggers 10 to 20 intermediate steps where 1 to 3 used to be enough. A message is therefore no longer a meaningful unit of measurement. Hence the switch to cost-based limits.
What this means for you: instead of a fixed number of messages, you now have two usage windows.
| Window | How It Works |
|---|---|
| Session Window | Starts with your first message. Resets automatically after 5 hours. |
| Weekly Window | Fixed weekly limit. Resets every Monday. |
What Happens When You Reach the Limit
The most important message first: reaching the limit does not lock you out. You are automatically rerouted to GPT-5.2 which was the top model just a few months ago. You keep working, just with a different model. Once the window resets, all models are available again.
With Claude directly or ChatGPT, reaching the limit locks you out completely until the reset. With Langdock, you keep working throughout.
If You Work with Your Own API Keys (BYOK)
Workspaces with BYOK are not directly affected by the new Fair Usage Policy. Model costs go directly to your provider, not through Langdock. Instead of session and weekly limits, BYOK works with monthly spending limits at the workspace, plan, and user level, plus optional per-model message limits. When a limit is reached, requests are rerouted to the backbone model and the workspace stays active.
The tips in this newsletter also apply to BYOK: with BYOK, every token costs real money to your provider. Those who have model selection and chat hygiene under control directly lower the monthly bill.
Lever 1: Model Selection with the Three-Tier Model
The biggest lever is model choice. Not every task needs the most powerful model. A simple three-tier model to keep in mind:
| Tier | For What | Recommendation |
|---|---|---|
| Routine | Summarizing, drafting emails, classifying | Haiku 4.5 or GPT-5.4 Mini |
| Day-to-Day Work | Standard chats, research, writing | Sonnet 4.6 (Default) |
| High Quality | Strategy, nuance, code architecture, deep analysis | Opus 4.7 |
Two Additional Notes
- Reasoning Toggle: In the model selector, only enable this when the task genuinely requires reasoning. It produces longer responses and higher token consumption.
- Web Search: Here, GPT-5.4 is often more cost-efficient and comparable in quality to Opus.
Lever 2: Chat Hygiene
The model processes the complete chat history on every turn. Long chats therefore not only become cluttered but also consume more per response.
Four Routines That Make a Big Difference
New Chats for New Topics
Start a new chat after 10 to 15 turns at the latest, or when switching topics. The simplest lever of all.
Use Edit Instead of Following Up
If the first response is not right, sharpen the original prompt using the edit icon. Saves a full turn and is usually faster.
Consciously Limit Output Length
"Maximum 5 bullet points", "in 100 words", or "one paragraph, not a novel". Saves tokens and usually makes the response better too.
Use Attachments Selectively
Instead of dropping a 50 MB PDF into the chat, copy out the relevant section or place the file in a folder and have it searched semantically.
Lever 3: Skills, Agents, and Workflows as a Multiplier
What many overlook: Skills, Agents, and Workflows are not just convenience features. They actively save tokens.
Skills
Prevent you from typing the same context into a chat over and over. Defined once, automatically active, no more copy-paste. These are tokens that never get consumed in every new chat in the first place.
Specialized Agents
With small, focused knowledge, they are more efficient than an all-purpose bot that knows everything. Smaller context, lower consumption, better answers. Five focused agents beat one agent with the entire knowledge base.
Workflows
The underrated lever. What runs deterministically enough (reporting, data extraction, standard classification) belongs in a workflow, not in an agent chat.
Workflows do not count toward the session or weekly limit.
Lever 4: Auto Mode and Cost Indicator
Langdock is launching two features in the coming days that will make all of this easier.
Auto Mode
A new model option that automatically selects the best model for each task. Trivial requests go to a cost-efficient model; complex tasks go to the frontier.
Cost Indicator per Model
Each model card gets a visible indicator showing how much it draws from the limit. This makes selection more intuitive without requiring you to memorize pricing tiers.
Real-World Example: A Day at Quellberg GmbH
To make this tangible, let us look at a typical workday for Sandra. Sandra works in marketing at Quellberg GmbH, the fictional Munich-based beverage company we covered last month.
Morning Emails
Sandra drafts three reply emails to sales partners. She has a Skill "Quellberg Emails" that automatically sets tone and signature. Model: Haiku, because the briefing is all it takes. Consumption: minimal.
Weekly Newsletter
Sandra writes the customer newsletter. She activates the workspace-mandated Skill "Quellberg Communication Standard", drops three bullet points into the chat, and limits the response to "400 words, three sections". Corrections via the edit icon instead of starting a new turn.
Strategy Memo for Management
Quellberg is considering switching sales to D2C. Sandra needs an honest pro/con analysis with market data. Here she uses Opus 4.7 with Reasoning enabled. The task clearly justifies the extra consumption.
Reporting
The weekly performance report is not created in chat but in a workflow. Trigger is every Friday at 3:00 PM; output lands as a Google Doc in the marketing folder. Does not count toward the limit.
End of day: Sandra has had a productive day, is far from her weekly limit, and still used Opus where it mattered.
Common Mistakes
Three patterns that regularly lead to limit problems.
The classic mistake. If Opus with Reasoning is the default, you hit the weekly limit before you have really started. Sonnet is the right choice in 80% of cases.
Long chats consume exponentially more tokens per response because the full history is sent along. Start a fresh chat every 10 to 15 turns.
Long custom instructions are sent as a system prompt on every turn. Plus: they often worsen response quality because the model has to juggle too many instructions at once. Focus on the essentials.
Limitations
Three points we want to address honestly.
Auto Mode Is New
It is coming in the next few days and will need time to settle. Initially, it is worth checking occasionally which model was actually selected, and switch manually if needed.
Power Users Will Still Hit the Limit
Even with all these tips, 5 to 15 % of your team will regularly hit the limit. For them, Business Max (5x limits) or Extra Usage (pay-as-you-go) is the right lever. Talk to your admin.
No Automatic Tracking per Skill or Agent
Currently you only see total consumption. There is no breakdown per skill or agent yet. Those who need that can export the usage data and analyze it manually.
Our Tip
And: these two weeks with increased limits are the best opportunity to observe where your usage is concentrated. Who uses what for what? Who really needs Opus? Who gets by with Sonnet? This data will then help you decide cleanly whether a Business Max seat makes sense or whether workflow optimization is the better lever.
Until next month,
Timo