feat: add useNativeTokenCount flag to skip token counting API calls

[opieter-aws]

Description

Model providers with native token counting APIs (Bedrock, Anthropic, Gemini, OpenAI Responses, llama.cpp) make an additional API call on every count_tokens() invocation. In scenarios where latency or cost of these calls is undesirable — such as high-frequency proactive compression checks — users need a way to opt out and fall back to the local estimator.

Each provider that overrides count_tokens() with a native API call gains a new optional use_native_token_count field in its config:

# Before: native API always called (with estimator fallback on error)
model = BedrockModel(model_id="us.anthropic.claude-sonnet-4-20250514")

# After: skip native API, always use local estimator
model = BedrockModel(
    model_id="us.anthropic.claude-sonnet-4-20250514",
    use_native_token_count=False,
)

Defaults to True (current behavior). Available on BedrockConfig, AnthropicConfig, GeminiConfig, OpenAIResponsesConfig, and LlamaCppConfig — the five providers that override count_tokens() with native API calls. The base Model class is unaffected since it only has the local estimator implementation.

Related Issues

Python port of strands-agents/sdk-typescript#1009

Documentation PR

NA

Type of Change

New feature

Testing

How have you tested the change? Verify that the changes do not break functionality or introduce warnings in consuming repositories: agents-docs, agents-tools, agents-cli

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

Issue: This PR adds a new configuration parameter to every public model provider constructor (Anthropic, Bedrock, Gemini, LlamaCpp, OpenAIResponses). This is a public API change that affects all customers.

Per the API Bar Raising guidelines, the PR should:

1. Have the needs-api-review label
2. Document expected use cases in the PR description
3. Provide example code snippets showing usage
4. Include complete API signatures

Suggestion: Please update the PR description with:

• The motivation/use case (e.g., reducing API call costs, reducing latency in high-throughput scenarios)
• Example usage: model = BedrockModel(model_id="...", use_native_token_count=False)
• Any alternatives considered (e.g., a class-level method override vs. config flag)

[github-actions]

Issue: The PR checklist items are all unchecked — hatch run prepare hasn't been run, CONTRIBUTING.md hasn't been read, and no tests/documentation confirmation. The PR description sections (Description, Related Issues, Documentation PR, Type of Change) are all left as template placeholders.

Suggestion: Please complete the PR template:

• Fill in the description explaining why this change is needed
• Link to any related issue
• Select the correct change type (this is "New feature")
• Mark checklist items that have been completed

[github-actions]

Review Summary

Assessment: Request Changes

This PR adds a use_native_token_count configuration option to 5 model providers, allowing users to skip native token-counting API calls in favor of the local heuristic estimator. The feature is straightforward and tests are provided for each provider.

Key Areas to Address

• API Bar Raising: This modifies the public constructor API of all major model providers. Per project guidelines, it needs the needs-api-review label and proper documentation in the PR description (use cases, examples, API signatures).
• DRY / Architecture: The same field definition and 2-line check is duplicated across 5 files. This belongs in BaseModelConfig and the base Model class to ensure consistency for current and future providers.
• Documentation Accuracy: The docstring claims "True (default)" but no default is set — the behavior relies on None is not False. This should be clarified.
• Naming: Consider whether use_native_token_count clearly communicates intent (it reads as a noun, not a verb/action).
• PR Description: Template is unfilled — no motivation, use cases, or linked issues.

The implementation is functional and well-tested; the main concerns are around API design process and code organization.