fix: set use_native_token_count default to false

[opieter-aws]

Description

Changes the default value of use_native_token_count from True to False across all model providers (Bedrock, Anthropic, Gemini, LlamaCpp, OpenAI Responses).

Since v1.38.0, the event loop calls count_tokens() before every model invocation. When use_native_token_count=True (the previous default), this makes an additional network round-trip that processes the full message payload -- including images -- before the actual inference call. For multimodal/image workloads, this adds 25-50% per-invocation latency with no indication of why.

This change makes native token counting opt-in rather than opt-out, so users must explicitly set use_native_token_count=True to enable the native API calls.

Related Issues

Fixes #2277

Documentation PR

N/A

Type of Change

Bug fix

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]

Assessment: Approve

Clean, well-scoped fix that correctly changes use_native_token_count from opt-out to opt-in across all 5 model providers with native token counting support. The logic change from is False to is not True properly handles the None (unset) case, and the approach aligns with tenet #4 ("The obvious path is the happy path") — users shouldn't silently pay a 25-50% latency penalty by default.

Review Details

• Correctness: The is not True check correctly treats both None (key absent) and False as "skip native counting," which is the right behavior for making this opt-in.
• Consistency: All 5 providers with count_tokens overrides are updated uniformly. The other providers (OpenAI chat, LiteLLM, etc.) don't override count_tokens and are unaffected.
• Testing: New test_skip_native_api_by_default tests verify the default behavior for each provider, and existing tests are properly updated to pass use_native_token_count=True explicitly.

Nice fix addressing a real user-reported latency regression.