TECHNOSYLVA · AI PLATFORM
INTERACTIVE GUIDE · COMPANY-WIDE

The right model
for every task

How AI cost is calculated, which model to pick for what you're doing, and the habits that keep spend under control. Built for every team — you don't need to know how to code.

01 · BEFORE WE START

Where you see what you're spending today

Each tool shows usage in a different place. Worth a look before the rest of this guide.

CLAUDE (CHAT & APP)

Settings → Usage

Daily spend by model, a running table, and how it moved versus the previous period. This is the same screen for a personal account or a team seat.

Claude Settings, Usage tab: daily spend by model chart and per-model spend table
CLAUDE CODE / DESKTOP

A live status line

You can see the model and the running session cost the whole time you're working, turn by turn.

Tip: set up Claude Code's status line once, and it keeps the model and cost in view without asking.

$ /statusline
Set up Claude Code's status line UI

C:/Projects/AIEficient  |  Sonnet 5  |  effort:med  |  ctx 14% used (86% left)  |  ~$0.420
▶▶ auto mode on (shift+tab to cycle)
CHATGPT

Organization admin panel

If you're on the company account, usage lives in the workspace admin panel, not your personal profile.

These screens change often. Later in this guide: where we're headed is one shared cost dashboard for the whole company.

02 · CONCEPT 1

Everything is measured in tokens

A token is, roughly, three-quarters of a word. Models don't bill per message — they bill per token, both what you send and what comes back.

0words
0approximate tokens

This is a rough estimate. The real count depends on the model, but the ratio holds: a 5,000-line codebase can easily turn into tens of thousands of tokens.

03 · CONCEPT 2

Input and output don't cost the same

Every model has two meters running. The second one is always the expensive one.

INPUT

What you send

The prompt, the conversation history, attached documents, pasted code. Billed every time you send it — even the third time you send the same thing.

OUTPUT

What the model generates

The reply you get back. Usually 5 to 6 times more expensive per token than input. If the model "reasons" before answering, that reasoning also bills as output.

Tip: ask for short, concrete answers. "Three bullet points, no preamble" costs less and reads faster than a five-paragraph reply you'll only skim.
04 · CONCEPT 3

A chat has no memory between messages

Every new message resends the entire conversation so far. That's why a long chat gets expensive without you noticing.

5messages in the chat
15total context units resent

Message number 20 fully reprocesses the 19 before it. Nobody flags it in the moment — you notice it at the end of the month.

05 · PUTTING IT TOGETHER

The formula, in short

cost ≈ (input tokens × input price)
    + (output tokens × output price)

Context that piles up in a long session inflates that first term every single turn. A 150-person team using a coding assistant daily can rack up a monthly bill in the thousands from this alone.

What drives cost without you noticing

Repeated context.

Documents and history are billed every time they're resent.

Reasoning turned on.

Useful for hard tasks, wasteful for easy ones.

Small actions that add up.

Fractions of a cent per action, times thousands of actions, times hundreds of people.

Always reaching for the biggest model.

Because it sounds better — not because the task needs it.

06 · TODAY'S CATALOG

There's a scale of speed, capability and price

The useful equivalence between providers is by task tier, not by exact model name. There's no perfect one-to-one match.

07 · IN PRACTICE

Same task, different tier

Pick your area and see how its everyday tasks split across the three main tiers.

08 · THE CORE IDEA

Important isn't the same as hard

Pick the model for what the task needs, not for how important it feels. A tense reply to an unhappy customer is still just a clear, well-written reply. Figuring out why churn moved is genuinely hard, even with no deadline in sight.

FEELS URGENT, IS STANDARD

"Reply to this angry customer within the hour" → still a clear, empathetic reply. Default tier.

FEELS ROUTINE, IS HARD

"Check why churn ticked up 2% last quarter" → conflicting signals, real reasoning, no single obvious answer. Deep tier.

09 · HOW TO START EVERY TASK

The recommended workflow

STEP 1

Start with the default model

It's the best balance of quality and cost for day-to-day work.

STEP 2

Improve the prompt and the context

If the result isn't convincing, it's almost always the instructions, not the model.

STEP 3

Escalate one tier

Only if, with a good prompt, the output still isn't good enough.

Try it: was the result good enough?

Rule: use the smallest approved model that succeeds reliably.

10 · GOOD PRACTICES

How to manage context

The model pays more attention to what's at the start and the end of your message. In the middle, attention drops.

STARTMIDDLEEND
Where do instructions go?

Put the general framing at the start: role, format, rules. Right before your actual question, repeat the important part in one short line. That's where the model pays the most attention at the moment it answers.

What happens when context gets too long?

The model can lose details from the middle, give more generic answers, and make more mistakes. On top of that, every new message costs more because it drags the whole history along. If a long session starts giving worse answers, that's a signal to start over, not to push harder.

When do I start a new session?

One session per topic. If you switch subjects, open a new chat instead of continuing the same thread. If you need continuity, paste a two or three sentence summary into the new chat instead of dragging the whole history along.

What about memory?

It's automatic. Claude picks up relevant details from your conversations on its own — your role, recurring projects, preferences you've mentioned — and brings them into new chats without you doing anything.

You can also step in. Ask directly to remember something ("remember I lead the QA team") or to forget something ("forget the topic we discussed about X"). You can review or edit what's stored in Settings → Memory.

For anything that has to carry across sessions or days — a decision made, a naming convention, an open question — don't rely on memory alone. Close a session with a short recap ("here's what we decided: X, Y, Z") and paste it into the next chat. Memory is good at knowing who you are and what you generally work on. It's not a substitute for carrying forward the specific facts a task depends on.

11 · GOOD PRACTICES

Prompts, skills and reusing tokens

PRACTICES · META-PROMPTING

Let a cheap model write the expensive prompt

Goal — use a Fast-tier model to draft or critique the prompt you're about to spend Deep-tier tokens on.

Before running a complex task on an expensive model, ask a cheap one to draft the prompt, list what context is missing, or critique your first attempt. The Fast tier is good at exactly this kind of mechanical, well-specified work — so the Deep tier's expensive tokens go toward solving the actual problem, not toward you iterating on phrasing.

Tip: on a complex task, try /model opusplan — it plans with a stronger model, then hands execution to a lighter one. Plan deep, run cheap.

How to write a good prompt

State the goal, the format and the length in your first message. "Summarize this in three bullets, for the team channel" works better than "summarize this." Give an example if the format matters. The more specific you are the first time, the fewer rounds you need afterward.

Ask for concise, clear answers. It's the cheapest optimization there is, and it makes the output easier to act on too.

Reuse skills

A skill stores the prompt, context and format of a task you repeat, so you're not rebuilding it from scratch every time. Synapse already has skills ready for several common workflows, and some are built specifically to cut token use — for example rtk, caveman or graphify, which compress, restructure or pre-filter input before it ever reaches the model. If you're doing something similar every week, check whether a skill already exists before writing the prompt from zero.

TIP · SKILL-CREATOR

Turn a session that worked into a skill

Had a session that nailed a task you know you'll repeat? Call skill-creator right there and let it turn that conversation into a reusable skill — the prompt, the context and the format, packaged so next time it's one step instead of rebuilding it from scratch.

CLI tools vs MCP connectors

An MCP connector is powerful, but every tool definition, every call and every result becomes part of your context — it's billed like any other input. A CLI tool, called directly, doesn't need a full schema loaded into every request, and lets you filter what comes back before it ever reaches the model. For a simple, well-defined job, prefer a CLI over an MCP connector when both are on the table — same result, a fraction of the context. Reach for an MCP when the model genuinely needs to explore or act across a live system, not just fetch one known thing.

Turning connectors on and off

In Claude (chat/app): open the + menu on a message, go to Connectors, and toggle each one on or off. Only turn on what a given task actually needs — every enabled connector's tools are loaded into context whether you use them or not.

Claude + menu open on Connectors, showing Atlassian, Google Calendar, Google Drive, Notion, Slack and Claude in Chrome toggled on

In Claude Code: run /mcp to see which MCP servers are configured and connected, without leaving the terminal.

$ /mcp
MCP Server Status

  ✓ jira-mcp        connected   12 tools
  ✓ slack-mcp       connected    8 tools
  ○ notion-mcp      disabled
  ✓ incident-io-mcp connected    5 tools

Use /mcp <server> to toggle a server on or off
12 · WHERE WE'RE HEADED

What we're building

Model choice won't sit on every individual's shoulders forever. Three pieces are already in motion.

A model router

Automatically decides which model to use for a task, so you don't have to pick it yourself every time.

Hybrid infrastructure

Some work will run on our own or private infrastructure, some on public providers, depending on the case.

A cost dashboard

Visibility into how much is spent, and on what, by team and by use case.

This is in progress, not a finished product announcement. Until it lands, the habits in this guide are the biggest lever you personally have.

13 · KEEP THIS IN MIND

The pocket summary

  • Start with the default model.
  • Give clear instructions, up front and repeated at the end.
  • Open a new session for every new task.
  • Reuse the skills and prompts that already work.
  • Ask for concise, clear answers.
  • Move up a tier only when you genuinely need to.
  • Check your usage every once in a while.

Not sure which model fits your case? Ask in the AI Champions channel. That's what it's there for.