Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: add GPT-6 Astra to the fallback catalog and reasoning controls (#10475)

Michael Han06263ed22cd2
diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index fac96539108..2a690dc8118 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -28,6 +28,7 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
"gpt-5.6-luna",
"gpt-5.6-sol",
"gpt-5.6-terra",
+ "gpt-6-astra",
],
"model_capabilities": {
"gpt-5.4": {"vision": True, "studio_tools": True},
@@ -36,6 +37,7 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
"gpt-5.6-luna": {"vision": True, "studio_tools": True},
"gpt-5.6-sol": {"vision": True, "studio_tools": True},
"gpt-5.6-terra": {"vision": True, "studio_tools": True},
+ "gpt-6-astra": {"vision": True, "studio_tools": True},
},
"supports_streaming": True,
"supports_vision": True,
diff --git a/studio/backend/tests/test_openai_codex_subscription.py b/studio/backend/tests/test_openai_codex_subscription.py
index 1629f326e40..6cbee8717af 100644
--- a/studio/backend/tests/test_openai_codex_subscription.py
+++ b/studio/backend/tests/test_openai_codex_subscription.py
@@ -70,6 +70,7 @@ def test_protocol_constants_and_curated_provider_contract():
"gpt-5.6-luna",
"gpt-5.6-sol",
"gpt-5.6-terra",
+ "gpt-6-astra",
]
assert OPENAI_CODEX_DEVICE_REDIRECT_URI == ("https://auth.openai.com/deviceauth/callback")
row = next(
@@ -262,7 +263,14 @@ def test_successful_callback_does_not_wait_for_its_own_connection(monkeypatch):
assert len(persisted) == 1
-def test_fixed_host_transport_sends_subscription_headers_and_normalizes_sse():
+ ("model", "reasoning_effort"),
+ [("gpt-5.4", "none")]
+ + [("gpt-6-astra", effort) for effort in ("low", "medium", "high", "xhigh", "max")],
+)
+def test_fixed_host_transport_sends_subscription_headers_and_normalizes_sse(
+ model, reasoning_effort
+):
captured = {}
class FakeResponse:
@@ -298,9 +306,9 @@ def test_fixed_host_transport_sends_subscription_headers_and_normalizes_sse():
provider_id = "provider-1",
thread_id = "thread-1",
messages = [{"role": "user", "content": "hello"}],
- model = "gpt-5.4",
+ model = model,
max_tokens = 100,
- reasoning_effort = "none",
+ reasoning_effort = reasoning_effort,
tools = None,
tool_choice = None,
)
@@ -317,7 +325,8 @@ def test_fixed_host_transport_sends_subscription_headers_and_normalizes_sse():
assert captured["json"]["store"] is False
assert "max_output_tokens" not in captured["json"]
- assert captured["json"]["reasoning"] == {"effort": "none", "summary": "auto"}
+ assert captured["json"]["model"] == model
+ assert captured["json"]["reasoning"] == {"effort": reasoning_effort, "summary": "auto"}
assert captured["json"]["include"] == ["reasoning.encrypted_content"]
assert any("hello" in line for line in lines)
assert not any("secret-token" in line for line in lines)
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index c20a7f5fdb3..84ceaa9ce8d 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -646,6 +646,11 @@ function resolveAnthropicReasoningEffortCapabilities(modelId: string): Reasoning
}
const OPENAI_REASONING_MODELS = [
+ {
+ prefixes: ["gpt-6-astra"],
+ supportsOff: false,
+ levels: ["low", "medium", "high", "xhigh", "max"],
+ },
{
prefixes: ["gpt-5.5-pro", "gpt-5.4-pro"],
supportsOff: false,
diff --git a/studio/frontend/tests/provider-capabilities-current-models.test.ts b/studio/frontend/tests/provider-capabilities-current-models.test.ts
index deccfe09f63..c67ba606902 100644
--- a/studio/frontend/tests/provider-capabilities-current-models.test.ts
+++ b/studio/frontend/tests/provider-capabilities-current-models.test.ts
@@ -9,6 +9,7 @@ import { registerBundlerResolver } from "./helpers/kit.ts";
registerBundlerResolver();
const {
+ clampReasoningEffortToLevels,
getExternalMaxOutputTokens,
getExternalReasoningCapabilities,
providerSupportsBuiltinCodeExecution,
@@ -76,6 +77,27 @@ test("the gpt-5.6 family gets the gpt-5.5 reasoning ladder", () => {
}
});
+test("Astra exposes mandatory reasoning with its full effort ladder", () => {
+ const caps = getExternalReasoningCapabilities("openai_codex", "gpt-6-astra");
+ assert.equal(caps.supportsReasoning, true);
+ assert.equal(caps.reasoningStyle, "reasoning_effort");
+ assert.equal(caps.supportsReasoningOff, false);
+ assert.deepEqual(
+ [...caps.reasoningEffortLevels],
+ ["low", "medium", "high", "xhigh", "max"],
+ );
+ for (const effort of ["none", "minimal"] as const) {
+ assert.equal(clampReasoningEffortToLevels(effort, caps.reasoningEffortLevels), "low");
+ }
+ assert.equal(clampReasoningEffortToLevels("max", caps.reasoningEffortLevels), "max");
+});
+
+test("Astra reasoning does not enable unrelated model families", () => {
+ for (const model of ["gpt-6-other", "gpt-60-astra"]) {
+ assert.equal(getExternalReasoningCapabilities("openai_codex", model).supportsReasoning, false);
+ }
+});
+
test("the gpt-5.1 and gpt-5.2 ladders drop minimal for none", () => {
// "minimal" was replaced by "none" from 5.1 on, and offering it fails the
// turn with "does not support 'minimal' with this model".
@@ -149,8 +171,9 @@ test("ChatGPT subscription models expose Unsloth-owned search and code tools", (
"gpt-5.3-codex-spark": { vision: false, studio_tools: true },
"gpt-5.4": { vision: true, studio_tools: true },
"gpt-5.6-sol": { vision: true, studio_tools: true },
+ "gpt-6-astra": { vision: true, studio_tools: true },
});
- for (const model of ["gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.6-sol"]) {
+ for (const model of ["gpt-5.3-codex-spark", "gpt-5.4", "gpt-5.6-sol", "gpt-6-astra"]) {
const caps = getExternalReasoningCapabilities("openai_codex", model);
assert.equal(caps.supportsReasoning, true, model);
assert.equal(caps.reasoningStyle, "reasoning_effort", model);