Token economics.
From prompt to production.

1# Every prompt is a series of tokens
2import tiktoken
3
4enc = tiktoken.encoding_for_model("gpt-5.2")
5
6system = "You are a helpful assistant."
7user = "Fix the auth bug."
8
9system_tokens = len(enc.encode(system))
10user_tokens = len(enc.encode(user))
11
12print(f"System: {system_tokens} tokens")
13print(f"User: {user_tokens} tokens")
14print(f"Total: {system_tokens + user_tokens} tokens")

1# Context Window = Your Token Budget
2import tiktoken
3
4enc = tiktoken.encoding_for_model("gpt-5.2")
5WINDOW = 256_000  # max tokens
6
7system = "You are a helpful assistant."
8system_tokens = len(enc.encode(system))
9
10remaining = WINDOW - system_tokens
11print(f"Budget remaining: {remaining:,} tokens")

1# Trimming — Last-N Strategy
2def trim_history(messages, n=3):
3    """Keep only the last N turns."""
4    system = [m for m in messages if m['role'] == 'system']
5    turns = [m for m in messages if m['role'] != 'system']
6    return system + turns[-n * 2:]
7
8# 20 messages → 6 (last 3 turns)
9messages = trim_history(conversation, n=3)
10print(f"Kept {len(messages)} messages")

1# Summarisation — Token Economy
2import anthropic
3
4def summarise_history(messages, client):
5    """Condense conversation to save tokens."""
6    history_text = "\n".join(
7        f"{m['role']}: {m['content']}" for m in messages
8    )
9    response = client.messages.create(
10        model="claude-sonnet-4-20250514",
11        max_tokens=500,
12        messages=[{
13            "role": "user",
14            "content": f"Summarise this conversation:\n{history_text}"
15        }]
16    )
17    return response.content[0].text

1# Token Budget Manager
2class TokenBudget:
3    def __init__(self, limit=128_000):
4        self.limit = limit
5        self.allocations = {}
6
7    def allocate(self, name, tokens):
8        self.allocations[name] = tokens
9
10    @property
11    def remaining(self):
12        used = sum(self.allocations.values())
13        return self.limit - used
14
15    @property
16    def utilization(self):
17        return sum(self.allocations.values()) / self.limit

1# Context Pollution Detection
2def detect_pollution(messages):
3    """Flag stale or redundant content."""
4    stale = []
5    for i, msg in enumerate(messages):
6        if msg.get('tool_result'):
7            age = len(messages) - i
8            if age > 10:  # older than 10 turns
9                stale.append(i)
10    print(f"Found {len(stale)} stale entries")
11    return stale

1# Anthropic Automatic Context Compaction
2import anthropic
3
4client = anthropic.Anthropic()
5
6runner = client.beta.messages.tool_runner(
7    model="claude-sonnet-4-5",
8    max_tokens=4096,
9    tools=tools,
10    messages=messages,
11    compaction_control={
12        "enabled": True,
13        "context_token_threshold": 5000,
14    },
15)
16
17for message in runner:
18    total_input += message.usage.input_tokens
19    total_output += message.usage.output_tokens

1# Context Editing Example
2def edit_context(messages, client):
3    """Remove stale info from context."""
4    prompt = "Review this conversation. Remove"
5    prompt += " outdated or redundant information."
6    prompt += " Keep decisions and current state."
7
8    response = client.messages.create(
9        model="claude-sonnet-4-20250514",
10        max_tokens=4000,
11        messages=[{
12            "role": "user",
13            "content": f"{prompt}\n\n{format_msgs(messages)}"
14        }]
15    )
16    return parse_edited_messages(response.content[0].text)

1# Custom summary prompt for domain needs
2compaction_control={
3    "enabled": True,
4    "context_token_threshold": 5000,
5    "summary_prompt": (
6        "Preserve: ticket IDs, categories, "
7        "priorities, teams, outcomes. "
8        "Discard: full KB articles, draft text."
9    ),
10}
11
12# Use cheaper model for summaries
13compaction_control={
14    "enabled": True,
15    "model": "claude-haiku-4-5",
16}