The SDK estimates how much of the model’s context window is in use so it can trigger compression before overflow. Three pieces work together: the context window limit, token counting, and utilization estimation.

## Context window limit

The threshold check requires the model’s context window size. The SDK auto-populates `context_window_limit``contextWindowLimit` from built-in lookup tables ([Python](https://github.com/strands-agents/harness-sdk/blob/main/strands-py/src/strands/models/_defaults.py), [TypeScript](https://github.com/strands-agents/harness-sdk/blob/main/strands-ts/src/models/defaults.ts)) for known models. Override it manually for models not in the lookup table:

(( tab "Python" ))
```python
model = BedrockModel(
    model_id="my-custom-model",
    context_window_limit=128_000,
)
```
(( /tab "Python" ))

(( tab "TypeScript" ))
```typescript
import { BedrockModel } from '@strands-agents/sdk'

const model = new BedrockModel({
  modelId: 'my-custom-model',
  contextWindowLimit: 128_000,
})
```
(( /tab "TypeScript" ))

Inaccurate compression with default fallback

If

`context_window_limit``contextWindowLimit`

is not set and the model ID is not in the built-in lookup table, the SDK falls back to a default of 200,000 tokens. When the actual context window is significantly different, compression thresholds will not behave correctly.

## Token estimation

The agent estimates input tokens using the following strategy:

1.  **Known baseline**: reads `input_tokens + output_tokens``inputTokens + outputTokens` from the last assistant message’s usage metadata
2.  **Delta estimation**: estimates tokens for new messages added since that baseline using the model’s `count_tokens()``countTokens()` method
3.  **Cold start fallback**: when no prior usage metadata exists (first call or after session restore), estimates all messages via `count_tokens()``countTokens()`

The `count_tokens()``countTokens()` method uses a character-based heuristic by default (characters / 4 for text, characters / 2 for JSON). Some model providers support native token counting APIs for exact counts. See the Token Counting section on each provider’s page:

-   [Amazon Bedrock](/docs/user-guide/concepts/model-providers/amazon-bedrock/index.md#token-counting)
-   [Anthropic](/docs/user-guide/concepts/model-providers/anthropic/index.md#token-counting)
-   [Google Gemini](/docs/user-guide/concepts/model-providers/google/index.md#token-counting)
-   [OpenAI Responses](/docs/user-guide/concepts/model-providers/openai-responses/index.md#token-counting)
-   [llama.cpp](/docs/user-guide/concepts/model-providers/llamacpp/index.md#token-counting)

## Utilization estimation

The Model base class provides an `estimate_utilization(input_tokens)``estimateUtilization(inputTokens)` method that computes the fraction of the context window consumed by a given input token count:

(( tab "Python" ))
```python
ratio = model.estimate_utilization(
    input_tokens=projected_tokens,
)
# ratio is 0-1+ (above 1.0 means overflow)
```
(( /tab "Python" ))

(( tab "TypeScript" ))
```typescript
// Section "utilization_imports" not found in user-guide/concepts/context-management/context-estimation_imports.ts

const ratio = model.estimateUtilization(projectedTokens)
// ratio is 0-1+ (above 1.0 means overflow)
```
(( /tab "TypeScript" ))

The method divides `input_tokens``inputTokens` by the model’s `context_window_limit``contextWindowLimit`, falling back to the 200,000 default with a warning when not configured. A return value above 1.0 means overflow.

Both the built-in [context management modes](/docs/user-guide/concepts/context-management/built-in-modes/index.md) and the [conversation managers](/docs/user-guide/concepts/agents/conversation-management/index.md) use this internally for compression decisions. You can also call it directly when building custom logic.

## Related pages

- [Built-in Modes](/docs/user-guide/concepts/context-management/built-in-modes/index.md) (2 shared tags)
- [Context Management](/docs/user-guide/concepts/context-management/index.md) (2 shared tags)
- [Custom Strategies](/docs/user-guide/concepts/context-management/custom-strategies/index.md) (2 shared tags)
- [Strategy Presets](/docs/user-guide/concepts/context-management/presets/index.md) (2 shared tags)
- [Context Offloader](/docs/user-guide/concepts/plugins/context-offloader/index.md) (2 shared tags)
- [Conversation Management](/docs/user-guide/concepts/agents/conversation-management/index.md) (2 shared tags)
- [Context Injector](/docs/user-guide/concepts/plugins/context-injector/index.md) (1 shared tag)
- [Coherence Evaluator](/docs/user-guide/evals-sdk/evaluators/coherence_evaluator/index.md) (1 shared tag)
- [Conciseness Evaluator](/docs/user-guide/evals-sdk/evaluators/conciseness_evaluator/index.md) (1 shared tag)
- [Goal Success Rate Evaluator](/docs/user-guide/evals-sdk/evaluators/goal_success_rate_evaluator/index.md) (1 shared tag)


## Implementation

### Python

- [harness-sdk/strands-py/src/strands/models/model.py](https://github.com/strands-agents/harness-sdk/blob/main/strands-py/src/strands/models/model.py)

### TypeScript

- [harness-sdk/strands-ts/src/models/model.ts](https://github.com/strands-agents/harness-sdk/blob/main/strands-ts/src/models/model.ts)