Add Claude Opus 4.7 support by dsfaccini · Pull Request #5118 · pydantic/pydantic-ai

dsfaccini · 2026-04-16T15:30:15Z

No issue; this is a compatibility update for Anthropic Claude Opus 4.7.

Changes

add claude-opus-4-7 to the Anthropic and gateway known-model registries
bump the optional Anthropic SDK requirement to anthropic>=0.96.0 and refresh uv.lock
map unified thinking to Opus 4.7 adaptive thinking with xhigh effort support
reject provider-specific budget thinking on Opus 4.7 and drop unsupported sampling params
add unit coverage plus a recorded VCR test for the 4.7 request shape
refresh Anthropic thinking docs for Opus 4.7 semantics

Checklist

Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
No breaking changes in accordance with the version policy.
PR title is fit for the release changelog.

dsfaccini · 2026-04-16T15:30:44Z

Claude here: follow-up PR scope for Anthropic task budget support:

add a typed anthropic_task_budget setting rather than overloading anthropic_thinking
send output_config.task_budget only when that setting is present
auto-add the task-budgets-2026-03-13 Anthropic beta when task budgets are used
cover the request shape and validation paths with unit tests first, then add a recorded VCR case if the API behavior is stable enough
document the beta status and the 4.7-specific semantics in the Anthropic/thinking docs

I split this out from #5118 to keep the first PR focused on bare Opus 4.7 compatibility and SDK support, without mixing in a new beta surface area.

dsfaccini · 2026-04-16T16:06:41Z

+    assert exc_info.value.status_code == 400
+    assert exc_info.value.model_name == 'claude-opus-4-6'
+    assert len(vcr.requests) == 1
+    request_body = json.loads(vcr.requests[0].body)
+    assert request_body['model'] == 'claude-opus-4-6'
+    assert request_body['output_config'] == {'effort': 'xhigh'}


this is def not acceptable/relevant/necessary lol, I'd prefer snapshotting the exception

dsfaccini · 2026-04-16T16:12:57Z

+    request_body = json.loads(vcr.requests[0].body)
+    assert request_body['output_config'] == {'effort': 'xhigh'}


this part remains unnecessary, doesn't add any new information (the snapshot shows the effort level sent)

dsfaccini · 2026-04-16T16:13:16Z

+
+
+async def test_anthropic_opus_47_drops_sampling_settings(allow_model_requests: None):
+    from anthropic._types import omit as OMIT


no inline imports, imports go at the top behind an import gate

devin-ai-integration

Devin Review found 2 new potential issues.

View 10 additional findings in Devin Review.

devin-ai-integration · 2026-04-16T16:49:36Z

    supports_json_schema_output = model_name.startswith(models_that_support_json_schema_output)

    # Sonnet 4.6+ and Opus 4.6+ support adaptive thinking; older models use budget-based
-    supports_adaptive = model_name.startswith(('claude-sonnet-4-6', 'claude-opus-4-6'))
+    supports_adaptive = model_name.startswith(('claude-sonnet-4-6', 'claude-opus-4-6', 'claude-opus-4-7'))

    # Opus 4.5+ and Sonnet 4.6+ support the effort parameter in output_config
-    supports_effort = model_name.startswith(('claude-opus-4-5', 'claude-opus-4-6', 'claude-sonnet-4-6'))
+    supports_effort = model_name.startswith(
+        ('claude-opus-4-5', 'claude-opus-4-6', 'claude-opus-4-7', 'claude-sonnet-4-6')
+    )
+    supports_xhigh_effort = model_name.startswith('claude-opus-4-7')
+    disallows_budget_thinking = model_name.startswith('claude-opus-4-7')
+    disallows_sampling_settings = model_name.startswith('claude-opus-4-7')


🚩 claude-mythos-preview gets conservative default profile — may need capability flags

The new claude-mythos-preview model is added to KnownModelName but has no special handling in anthropic_model_profile() (pydantic_ai_slim/pydantic_ai/profiles/anthropic.py:61-87). It won't match any startswith tuple, so it gets: supports_json_schema_output=False, supports_adaptive=False, supports_effort=False, and all disallows_*=False. This is a safe conservative default, but if claude-mythos-preview is actually a frontier model with these capabilities, users would need to manually configure a profile override. Worth verifying what capabilities this model actually has.

(Refers to lines 61-87)

Was this helpful? React with 👍 or 👎 to provide feedback.

valid to point out, but would be in a follow up PR

mythos usage/adoption is gated currently

yeah we don't know anything about it

dsfaccini · 2026-04-16T17:23:57Z

-AnthropicModelName = str | LatestAnthropicModelNames
+LegacyAnthropicModelNames = Literal[
+    'claude-3-5-haiku-20241022',
+    'claude-3-5-haiku-latest',
+    'claude-3-7-sonnet-20250219',
+    'claude-3-7-sonnet-latest',
+    'claude-3-opus-20240229',
+    'claude-3-opus-latest',
+    'claude-4-opus-20250514',
+    'claude-4-sonnet-20250514',
+]
+"""Anthropic model names we continue to support beyond the SDK's latest literal set."""
+
+AnthropicModelName = str | LatestAnthropicModelNames | LegacyAnthropicModelNames


@DouweM want to remove?

DouweM · 2026-04-16T22:10:40Z

+    `{'type': 'enabled', 'budget_tokens': ...}`.
+    """
+
+    anthropic_disallows_sampling_settings: bool = False


Why use _disallows_ here rather than supports as everywhere else?

DouweM · 2026-04-16T22:11:23Z

    supports_json_schema_output = model_name.startswith(models_that_support_json_schema_output)

    # Sonnet 4.6+ and Opus 4.6+ support adaptive thinking; older models use budget-based
-    supports_adaptive = model_name.startswith(('claude-sonnet-4-6', 'claude-opus-4-6'))
+    supports_adaptive = model_name.startswith(('claude-sonnet-4-6', 'claude-opus-4-6', 'claude-opus-4-7'))

    # Opus 4.5+ and Sonnet 4.6+ support the effort parameter in output_config
-    supports_effort = model_name.startswith(('claude-opus-4-5', 'claude-opus-4-6', 'claude-sonnet-4-6'))
+    supports_effort = model_name.startswith(
+        ('claude-opus-4-5', 'claude-opus-4-6', 'claude-opus-4-7', 'claude-sonnet-4-6')
+    )
+    supports_xhigh_effort = model_name.startswith('claude-opus-4-7')
+    disallows_budget_thinking = model_name.startswith('claude-opus-4-7')
+    disallows_sampling_settings = model_name.startswith('claude-opus-4-7')


yeah we don't know anything about it

devin-ai-integration

Devin Review found 1 new potential issue.

View 11 additional findings in Devin Review.

devin-ai-integration · 2026-04-16T22:38:12Z

+    Anthropic's [Claude Opus 4.7 migration guide](https://platform.claude.com/docs/en/about-claude/models/migration-guide) recommends removing `temperature`, `top_p`, and `top_k` from Opus 4.7 requests. Pydantic AI drops those keys automatically for `claude-opus-4-7`, including `extra_body` overrides.
+
+    The same guide also recommends re-evaluating `max_tokens` and any token-count assumptions when migrating from Opus 4.6, since Opus 4.7 uses updated tokenization. If you rely on `count_tokens()` or `count_tokens_before_request`, verify your thresholds against the new model.


🟡 Migration note uses plain backticks for pydantic-ai API elements instead of mkdocs reference links

The new "Claude Opus 4.7 migration" note at docs/models/anthropic.md:45-48 mentions several pydantic-ai API elements (count_tokens(), count_tokens_before_request, max_tokens, temperature, top_p, extra_body) using plain backticks instead of mkdocs reference-style links. This violates the mandatory rule in docs/.cursor/rules.mdc: "Always reference the python code in the docs e.g. ModelSettings should link to [ModelSettings][pydantic_ai.settings.ModelSettings]." At minimum, count_tokens() should link to [count_tokens()][pydantic_ai.models.Model.count_tokens] and count_tokens_before_request should link to [count_tokens_before_request][pydantic_ai.usage.UsageLimits.count_tokens_before_request].

Suggested change

Anthropic's [Claude Opus 4.7 migration guide](https://platform.claude.com/docs/en/about-claude/models/migration-guide) recommends removing `temperature`, `top_p`, and `top_k` from Opus 4.7 requests. Pydantic AI drops those keys automatically for `claude-opus-4-7`, including `extra_body` overrides.

The same guide also recommends re-evaluating `max_tokens` and any token-count assumptions when migrating from Opus 4.6, since Opus 4.7 uses updated tokenization. If you rely on `count_tokens()` or `count_tokens_before_request`, verify your thresholds against the new model.

Anthropic's [Claude Opus 4.7 migration guide](https://platform.claude.com/docs/en/about-claude/models/migration-guide) recommends removing [`temperature`][pydantic_ai.settings.ModelSettings.temperature], [`top_p`][pydantic_ai.settings.ModelSettings.top_p], and `top_k` from Opus 4.7 requests. Pydantic AI drops those keys automatically for `claude-opus-4-7`, including [`extra_body`][pydantic_ai.settings.ModelSettings.extra_body] overrides.

The same guide also recommends re-evaluating [`max_tokens`][pydantic_ai.settings.ModelSettings.max_tokens] and any token-count assumptions when migrating from Opus 4.6, since Opus 4.7 uses updated tokenization. If you rely on [`count_tokens()`][pydantic_ai.models.Model.count_tokens] or [`count_tokens_before_request`][pydantic_ai.usage.UsageLimits.count_tokens_before_request], verify your thresholds against the new model.

Was this helpful? React with 👍 or 👎 to provide feedback.

Add Claude Opus 4.7 Anthropic support

a20e7d9

github-actions Bot added the size: M Medium PR (101-500 weighted lines) label Apr 16, 2026

github-actions Bot added the feature New feature request, or PR implementing a feature (enhancement) label Apr 16, 2026

This comment was marked as resolved.

Sign in to view