Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 48 additions & 0 deletions dev-notes/claude-opus-5-model-support.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,48 @@
# Claude Opus 5 model support

## What changed

Tau's direct Anthropic catalog now includes `claude-opus-5` with metadata from
Anthropic's July 24, 2026 launch documentation:

- 1,000,000-token context window
- 128,000-token maximum output
- text and image input
- $5 / million input tokens and $25 / million output tokens
- adaptive thinking with `low`, `medium`, `high`, `xhigh`, and `max` effort

Tau retains `claude-sonnet-4-6` as Anthropic's default model. Users opt into
Opus 5 through `/model` or `--provider anthropic -m claude-opus-5`.

## Thinking compatibility

Tau has six UI levels while Opus 5 has five enabled-thinking effort levels plus
a disabled mode. `minimal` is unavailable for this model. Tau maps `xhigh` to
Anthropic's `max` wire value because both represent Tau's top reasoning tier.
`off` now sends `thinking: {"type": "disabled"}` explicitly; omitting the field
would not work because Opus 5 enables adaptive thinking by default.

## Architecture

This remains a catalog and provider-adapter change:

- `tau_coding` owns model discovery, metadata, and thinking-level mapping.
- `tau_ai` serializes the resulting Anthropic Messages API request.
- `tau_agent` remains provider-independent.

## Verification

Run:

```bash
uv run pytest tests/test_provider_catalog.py tests/test_provider_config.py tests/test_tau_ai.py
uv run ruff check .
uv run ruff format --check .
uv run mypy
```

Official sources:

- https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5
- https://platform.claude.com/docs/en/about-claude/models/overview
- https://platform.claude.com/docs/en/release-notes/overview
4 changes: 3 additions & 1 deletion src/tau_ai/anthropic.py
Original file line number Diff line number Diff line change
Expand Up @@ -362,7 +362,9 @@ def _build_messages_payload(
),
"messages": [_anthropic_message(message) for message in messages],
}
if thinking_mode == "adaptive" and thinking_effort is not None:
if thinking_mode == "disabled":
payload["thinking"] = {"type": "disabled"}
elif thinking_mode == "adaptive" and thinking_effort is not None:
payload["thinking"] = {"type": "adaptive", "display": "summarized"}
payload["output_config"] = {"effort": thinking_effort}
elif thinking_budget_tokens is not None:
Expand Down
14 changes: 13 additions & 1 deletion src/tau_coding/data/catalog.toml
Original file line number Diff line number Diff line change
Expand Up @@ -518,7 +518,7 @@ base_url = "https://api.anthropic.com"
api_key_env = "ANTHROPIC_API_KEY"
credential_name = "anthropic"
auth_methods = ["api_key", "oauth"]
models = ["claude-fable-5", "claude-haiku-4-5", "claude-haiku-4-5-20251001", "claude-opus-4-1", "claude-opus-4-1-20250805", "claude-opus-4-5", "claude-opus-4-5-20251101", "claude-opus-4-6", "claude-opus-4-7", "claude-sonnet-4-5", "claude-sonnet-4-5-20250929", "claude-sonnet-4-6", "claude-sonnet-5"]
models = ["claude-fable-5", "claude-haiku-4-5", "claude-haiku-4-5-20251001", "claude-opus-4-1", "claude-opus-4-1-20250805", "claude-opus-4-5", "claude-opus-4-5-20251101", "claude-opus-4-6", "claude-opus-4-7", "claude-opus-5", "claude-sonnet-4-5", "claude-sonnet-4-5-20250929", "claude-sonnet-4-6", "claude-sonnet-5"]
default_model = "claude-sonnet-4-6"
docs_url = "https://docs.anthropic.com"
api = "anthropic-messages"
Expand All @@ -536,6 +536,7 @@ claude-opus-4-5 = 200000
claude-opus-4-5-20251101 = 200000
claude-opus-4-6 = 1000000
claude-opus-4-7 = 1000000
claude-opus-5 = 1000000
claude-sonnet-4-5 = 200000
claude-sonnet-4-5-20250929 = 200000
claude-sonnet-4-6 = 1000000
Expand Down Expand Up @@ -621,6 +622,17 @@ cost = { input = 5, output = 25, cacheRead = 0.5, cacheWrite = 6.25 }
thinking_level_map = { xhigh = "xhigh" }
unsupported_thinking_levels = ["minimal"]

[providers.model_metadata.claude-opus-5]
name = "Claude Opus 5"
reasoning = true
input = ["text", "image"]
context_window = 1000000
max_tokens = 128000
compat = { forceAdaptiveThinking = true }
cost = { input = 5, output = 25, cacheRead = 0.5, cacheWrite = 6.25 }
thinking_level_map = { xhigh = "max" }
unsupported_thinking_levels = ["minimal"]

[providers.model_metadata.claude-sonnet-4-5]
name = "Claude Sonnet 4.5 (latest)"
reasoning = true
Expand Down
13 changes: 11 additions & 2 deletions src/tau_coding/provider_config.py
Original file line number Diff line number Diff line change
Expand Up @@ -1528,7 +1528,9 @@ def anthropic_config_from_provider(
model=selected_model,
thinking_level=thinking_level,
),
thinking_mode=_anthropic_thinking_mode(provider, selected_model),
thinking_mode=_anthropic_thinking_mode(
provider, selected_model, thinking_level=thinking_level
),
)


Expand Down Expand Up @@ -1688,9 +1690,16 @@ def _reasoning_effort_from_anthropic_provider(
return mapped or normalized


def _anthropic_thinking_mode(provider: AnthropicProviderConfig, model: str) -> str:
def _anthropic_thinking_mode(
provider: AnthropicProviderConfig,
model: str,
*,
thinking_level: ThinkingLevel | None = None,
) -> str:
compat = _model_compat(provider, model)
if compat.get("forceAdaptiveThinking") is True:
if thinking_level is not None and normalize_thinking_level(thinking_level) == "off":
return "disabled"
return "adaptive"
return "budget"

Expand Down
11 changes: 11 additions & 0 deletions tests/test_provider_catalog.py
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,7 @@ def test_builtin_catalog_golden_anthropic_entry() -> None:
"claude-opus-4-5-20251101",
"claude-opus-4-6",
"claude-opus-4-7",
"claude-opus-5",
"claude-sonnet-4-5",
"claude-sonnet-4-5-20250929",
"claude-sonnet-4-6",
Expand All @@ -119,11 +120,21 @@ def test_builtin_catalog_golden_anthropic_entry() -> None:
"claude-opus-4-5-20251101": 200_000,
"claude-opus-4-6": 1_000_000,
"claude-opus-4-7": 1_000_000,
"claude-opus-5": 1_000_000,
"claude-sonnet-4-5": 200_000,
"claude-sonnet-4-5-20250929": 200_000,
"claude-sonnet-4-6": 1_000_000,
"claude-sonnet-5": 1_000_000,
}
opus_5 = entry.model_metadata["claude-opus-5"]
assert opus_5.context_window == 1_000_000
assert opus_5.max_tokens == 128_000
assert opus_5.input == ("text", "image")
assert opus_5.cost is not None
assert opus_5.cost["input"] == 5
assert opus_5.cost["output"] == 25
assert opus_5.compat == {"forceAdaptiveThinking": True}
assert opus_5.thinking_level_map == {"minimal": None, "xhigh": "max"}
assert entry.thinking_levels == ("off", "minimal", "low", "medium", "high", "xhigh")
assert entry.thinking_models == ()
assert entry.thinking_default == "medium"
Expand Down
32 changes: 32 additions & 0 deletions tests/test_provider_config.py
Original file line number Diff line number Diff line change
Expand Up @@ -150,6 +150,13 @@ def test_builtin_openai_declares_model_scoped_thinking_capabilities() -> None:
"medium",
"high",
)
assert provider_thinking_levels(anthropic, model="claude-opus-5") == (
"off",
"low",
"medium",
"high",
"xhigh",
)


def test_load_provider_settings_accepts_provider_preferences_with_user_catalog(
Expand Down Expand Up @@ -931,6 +938,31 @@ def test_anthropic_config_from_provider_sets_thinking_budget(
assert high_config.thinking_budget_tokens == 8192


def test_anthropic_config_from_provider_maps_opus_5_adaptive_thinking(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setenv("ANTHROPIC_API_KEY", "test-key")
provider = ProviderSettings().get_provider("anthropic")
assert isinstance(provider, AnthropicProviderConfig)

off_config = anthropic_config_from_provider(
provider, model="claude-opus-5", thinking_level="off"
)
medium_config = anthropic_config_from_provider(
provider, model="claude-opus-5", thinking_level="medium"
)
xhigh_config = anthropic_config_from_provider(
provider, model="claude-opus-5", thinking_level="xhigh"
)

assert off_config.thinking_mode == "disabled"
assert off_config.thinking_effort is None
assert medium_config.thinking_mode == "adaptive"
assert medium_config.thinking_effort == "medium"
assert xhigh_config.thinking_mode == "adaptive"
assert xhigh_config.thinking_effort == "max"


@pytest.mark.parametrize(
("parameter", "expected"),
[
Expand Down
36 changes: 36 additions & 0 deletions tests/test_tau_ai.py
Original file line number Diff line number Diff line change
Expand Up @@ -1599,6 +1599,42 @@ def handler(request: httpx.Request) -> httpx.Response:
assert payload["thinking"] == {"type": "enabled", "budget_tokens": 8192}


@pytest.mark.anyio
async def test_anthropic_provider_explicitly_disables_default_adaptive_thinking() -> None:
requests: list[httpx.Request] = []

def handler(request: httpx.Request) -> httpx.Response:
requests.append(request)
return httpx.Response(
200,
text='data: {"type":"message_stop"}\n\n',
headers={"content-type": "text/event-stream"},
)

async with httpx.AsyncClient(transport=httpx.MockTransport(handler)) as client:
provider = AnthropicProvider(
AnthropicConfig(
api_key="test-key",
base_url="https://api.anthropic.test/v1",
thinking_mode="disabled",
),
client=client,
)

await _collect(
provider.stream_response(
model="claude-opus-5",
system="You are Tau.",
messages=[UserMessage(content="Say hello")],
tools=[],
)
)

payload = loads(requests[0].content)
assert payload["thinking"] == {"type": "disabled"}
assert "output_config" not in payload


@pytest.mark.anyio
async def test_anthropic_provider_streams_thinking_deltas() -> None:
def handler(_request: httpx.Request) -> httpx.Response:
Expand Down
14 changes: 14 additions & 0 deletions website/content/guides/providers-and-models.md
Original file line number Diff line number Diff line change
Expand Up @@ -181,6 +181,20 @@ list before creating or refreshing a runtime provider. This prevents accidental
provider/model mismatches, such as trying to send an API-only OpenAI model to the
separate `openai-codex` subscription provider.

### Claude Opus 5

Tau supports Anthropic's `claude-opus-5` through the direct `anthropic`
provider. The model has a 1M-token context window, accepts text and images,
generates up to 128k tokens, and costs $5 / $25 per million input/output tokens.
Anthropic enables adaptive thinking by default. Tau maps its `low` through
`high` modes directly, maps `xhigh` to Anthropic's maximum effort, and sends an
explicit disabled-thinking request for `off`.

Use `/login anthropic-api` or `/login anthropic-subscription`, then select
**Claude Opus 5** in `/model`. See Anthropic's
[Claude Opus 5 guide](https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5)
for current behavior and availability.

## Adding a custom / local provider

Any OpenAI-compatible endpoint works — including local servers like llama.cpp or
Expand Down