diff --git a/docs/features/model-temperature.md b/docs/features/model-temperature.md index 82b14e70..3b6a2597 100644 --- a/docs/features/model-temperature.md +++ b/docs/features/model-temperature.md @@ -36,11 +36,11 @@ Temperature controls output randomness, not code quality or accuracy directly. K ## Default Values in Zoo Code -Zoo Code uses a default temperature of 0.0 for most models, optimizing for maximum determinism and precision in code generation. This applies to OpenAI models, Anthropic models (non-thinking variants), LM Studio models, and most other providers. +Zoo Code uses a default temperature of 0.0 for many models, optimizing for maximum determinism and precision in code generation. The exact default and whether temperature is accepted depend on the provider and selected model. Some models use higher default temperatures - DeepSeek R1 models and certain reasoning-focused models default to 0.3, providing a balance between determinism and creative exploration. -Models with thinking capabilities (where the AI shows its reasoning process) require a fixed temperature of 1.0 which cannot be changed, as this setting ensures optimal performance of the thinking mechanism. This applies to any model with the ":thinking" flag enabled. +Some thinking modes require a fixed temperature, while other reasoning models reject temperature entirely. For example, GPT-6 Astra does not accept custom `temperature` or `top_p` values. Some specialized models don't support temperature adjustments at all, in which case Zoo Code respects these limitations automatically. @@ -93,9 +93,7 @@ Zoo Code implements temperature handling with these considerations: * User-defined settings take priority over defaults * Provider-specific behaviors are respected -* Model-specific limitations are enforced: - * Thinking-enabled models require a fixed temperature of 1.0 - * Some models don't support temperature adjustments +* Model-specific limitations are enforced, including fixed-temperature modes and models that do not support temperature adjustments. --- @@ -111,7 +109,7 @@ Experimenting with different temperature settings is the most effective way to d 4. **Document results** - Note which values produce the best outcomes for specific types of tasks 5. **Create profiles** - Save effective settings as [API configuration profiles](/features/api-configuration-profiles) for quick access -Remember that different models may respond differently to the same temperature values, and thinking-enabled models always use a fixed temperature of 1.0 regardless of your settings. +Remember that different models may respond differently to the same temperature values, and some reasoning models do not expose a temperature control at all. --- diff --git a/docs/providers/litellm.md b/docs/providers/litellm.md index 78279ed7..4b368b1e 100644 --- a/docs/providers/litellm.md +++ b/docs/providers/litellm.md @@ -57,6 +57,12 @@ To use LiteLLM with Zoo Code, you first need to set up and run a LiteLLM server. litellm_params: model: openai/gpt-model-id api_key: os.environ/OPENAI_API_KEY + + # GPT-6 Astra requires Responses for Zoo Code's tool calls + - model_name: gpt-6-astra + litellm_params: + model: openai/responses/gpt-6-astra + api_key: os.environ/OPENAI_API_KEY # Configure Azure OpenAI - model_name: azure-model @@ -106,6 +112,14 @@ Once your LiteLLM server is running, you have two options for configuring it in * Use the refresh button to update the model list if you've added new models to your LiteLLM server. * If no model is selected, Zoo Code will use a default model. Ensure you have configured at least one model on your LiteLLM server. +### GPT-6 Astra + +LiteLLM added `gpt-6-astra` metadata and pricing on September 3, 2026. Use a LiteLLM release that includes that support and configure the underlying route as `openai/responses/gpt-6-astra`, as shown above. Zoo Code sends a supported reasoning effort with tool requests, which activates LiteLLM's documented Chat Completions-to-Responses bridge. This is required because OpenAI does not support Astra tool calling through Chat Completions. + +Zoo Code omits temperature, uses `max_completion_tokens`, and limits reasoning effort to `low`, `medium`, `high`, `xhigh`, or `max` when LiteLLM reports the exact underlying route `openai/responses/gpt-6-astra`. The model still requires OpenAI account access; configuring LiteLLM does not bypass OpenAI's staged rollout. + +See [LiteLLM's OpenAI provider documentation](https://docs.litellm.ai/docs/providers/openai) for the Responses bridge and [OpenAI's Astra model page](https://developers.openai.com/api/docs/models/gpt-6-astra) for the upstream capability contract. + ### Option 2: Using OpenAI Compatible Provider Alternatively, you can configure LiteLLM using the "OpenAI Compatible" provider: @@ -132,6 +146,7 @@ When you configure the LiteLLM provider, Zoo Code interacts with your LiteLLM se * `supportsImages`: Determined from `model_info.supports_vision` provided by LiteLLM. * `supportsPromptCache`: Determined from `model_info.supports_prompt_caching` provided by LiteLLM. * `inputPrice` / `outputPrice`: Calculated from `model_info.input_cost_per_token` and `model_info.output_cost_per_token` from LiteLLM. + * GPT-6 Astra aliases receive its reasoning and sampling rules only when `litellm_params.model` is the verified `openai/responses/gpt-6-astra` route. * `supportsComputerUse`: This flag is set to `true` if the underlying model identifier matches one of the Anthropic models predefined in Zoo Code as suitable for "computer use" (see `COMPUTER_USE_MODELS` in technical details). Zoo Code uses default values for some of these properties if they are not explicitly provided by your LiteLLM server's `/model/info` endpoint for a given model. The defaults are: diff --git a/docs/providers/openai-chatgpt-plus-pro.md b/docs/providers/openai-chatgpt-plus-pro.md index 8571a093..b0b9cbf4 100644 --- a/docs/providers/openai-chatgpt-plus-pro.md +++ b/docs/providers/openai-chatgpt-plus-pro.md @@ -40,6 +40,13 @@ keywords: 5. Back in Zoo Code settings, pick a model from the dropdown. 6. Save. +## GPT-6 Astra + +GPT-6 Astra is available through this provider for eligible ChatGPT Pro, Enterprise, and Business Premium accounts. Select `gpt-6-astra` after signing in; availability remains controlled by OpenAI for the signed-in account. + +Zoo Code uses Codex's Responses Lite request format for Astra, including subscription-based prompt caching and tool calls. The Codex model exposes an 872,000-token maximum context window, image input, and `low`, `medium`, `high`, `xhigh`, and `max` reasoning levels. Zoo Code defaults Astra to `low` reasoning and does not send unsupported disabled or `none` reasoning values. + +For the upstream rollout and model contract, see [OpenAI's Codex model catalog](https://github.com/openai/codex/blob/main/codex-rs/models-manager/models.json) and [GPT-6 Astra model page](https://developers.openai.com/api/docs/models/gpt-6-astra). ## Tips and Notes diff --git a/docs/providers/openai-compatible.md b/docs/providers/openai-compatible.md index 2bac4722..e255235a 100644 --- a/docs/providers/openai-compatible.md +++ b/docs/providers/openai-compatible.md @@ -85,6 +85,8 @@ At a high level: For native tool calling to work, the model you select **must** support OpenAI-compatible tool calling. If a model does not support native tool calling, it cannot be used with Zoo Code. +GPT-6 Astra is not supported through this provider when tools are enabled. OpenAI allows Astra tool calling only through the Responses API, while Zoo Code's general OpenAI Compatible provider uses Chat Completions. Use the native [OpenAI provider](/providers/openai), or the dedicated [LiteLLM provider](/providers/litellm) with a documented Responses-backed route. + ### Example: simple native tool flow Here's a simplified example of how a file-reading tool might be exposed when using an OpenAI-native endpoint: diff --git a/docs/providers/openai.md b/docs/providers/openai.md index 99ec80e3..35ba6840 100644 --- a/docs/providers/openai.md +++ b/docs/providers/openai.md @@ -15,7 +15,7 @@ keywords: # Using OpenAI With Zoo Code -Zoo Code supports accessing models directly through the official OpenAI API, including the latest GPT-5 family with advanced features like reasoning effort control and verbosity settings. +Zoo Code supports a curated set of models through the official OpenAI API, including GPT-6 Astra and GPT-5 models with model-specific reasoning controls. :::info Want to use a ChatGPT Plus/Pro subscription instead? Use the **OpenAI – ChatGPT Plus/Pro** provider to sign in via OAuth (no API key): [OpenAI – ChatGPT Plus/Pro](/providers/openai-chatgpt-plus-pro). @@ -36,9 +36,15 @@ Use the **OpenAI – ChatGPT Plus/Pro** provider to sign in via OAuth (no API ke ## Available Models -Zoo Code supports all models available through OpenAI's API. +Zoo Code's model picker lists models whose capabilities and pricing have been verified for the native integration. It does not automatically list every model returned by OpenAI. -For the complete, up-to-date model list and capabilities, see [OpenAI's models documentation](https://platform.openai.com/docs/models). +GPT-6 Astra uses the model ID `gpt-6-astra`. OpenAI released it in the API and made it available in ChatGPT Work and Codex for Pro, Enterprise, and Business Premium users. You can use the API model here or connect an eligible ChatGPT subscription through the [ChatGPT Plus/Pro provider](/providers/openai-chatgpt-plus-pro). + +GPT-6 Astra supports text and image input, a 1,050,000-token context window, up to 922,000 input tokens, and up to 128,000 output tokens. It produces text output. Zoo Code uses OpenAI's Responses API because Astra tool calling is not supported through Chat Completions. + +For current capabilities and rollout status, see [OpenAI's GPT-6 Astra model page](https://developers.openai.com/api/docs/models/gpt-6-astra) and [API changelog](https://developers.openai.com/api/docs/changelog). + +OpenRouter and NanoGPT also list `openai/gpt-6-astra` and `openai/gpt-6-astra-pro` in their live model catalogs, while Vercel AI Gateway lists `openai/gpt-6-astra` and `openai/gpt-6-astra-fast`. Zoo Code discovers these models dynamically and applies Astra's required reasoning and sampling constraints. OpenCode Go does not currently list Astra, and Zoo Code's Amazon Bedrock provider does not yet use the OpenAI-compatible Responses transport required by Bedrock's Astra offering. --- @@ -58,13 +64,19 @@ For the complete, up-to-date model list and capabilities, see [OpenAI's models d ### Reasoning Effort Control -For models that support reasoning (GPT-5, o1, o3, o4 families), you can control how deeply the model thinks: +For models that support reasoning, Zoo Code shows only the levels accepted by the selected model. + +**GPT-6 Astra:** +- `low` +- `medium` (Zoo Code default) +- `high` +- `xhigh` +- `max` + +GPT-6 Astra does not accept `none` or `minimal`. Zoo Code falls back to its `medium` default if an imported profile contains an unsupported effort. **GPT-5 Models:** -- `minimal` - Fastest responses with basic reasoning -- `low` - Quick responses with light reasoning -- `medium` (default) - Balanced reasoning and response time -- `high` - Deep reasoning for complex problems +- Supported levels vary by model and can include `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, and `max`. **o1/o3/o4 Models:** - `low` - Minimal thinking time @@ -75,7 +87,7 @@ Some models have preset reasoning levels that cannot be changed. ### Verbosity Control -Available for GPT-5 models and select others, verbosity controls the detail level of responses: +Available for models that explicitly advertise support in Zoo Code, verbosity controls the detail level of responses: - `low` - Concise, direct responses - `medium` (default) - Balanced detail @@ -83,20 +95,16 @@ Available for GPT-5 models and select others, verbosity controls the detail leve ### Temperature Settings -Temperature controls output randomness (0.0 to 2.0): - -- **GPT-5 models:** Default 1.0 for balanced creativity -- **Other models:** Default 0.0 for deterministic output -- **Note:** Not available for o1/o3 reasoning models +Temperature controls output randomness where the selected model supports it. GPT-6 Astra does not accept custom `temperature` or `top_p` values, so Zoo Code omits them. The same is true for several reasoning models. -### Conversation Continuity (GPT-5) +### Conversation Continuity -GPT-5 models maintain conversation context efficiently through response IDs, reducing token usage while preserving context. This happens automatically - no configuration needed. +The native provider uses the Responses API and preserves compatible response and reasoning items across tool turns. GPT-6 Astra also supports prompt caching. Cache writes are billed separately, cached reads use a lower rate, and requests above 272,000 input tokens use long-context pricing for the entire request. See [OpenAI's prompt caching guide](https://developers.openai.com/api/docs/guides/prompt-caching). --- ## Tips and Notes -* **Pricing:** Refer to the [OpenAI Pricing](https://openai.com/pricing) page for current model costs and discounts, including prompt caching. +* **Pricing:** Refer to [OpenAI API pricing](https://developers.openai.com/api/docs/pricing) for current standard, Flex, Fast, Batch, cache, and long-context rates. * **Azure OpenAI Service:** If you'd like to use the Azure OpenAI service, please see our section on [OpenAI-compatible](/providers/openai-compatible) providers. * **Context Optimization:** For GPT-5-Codex, leverage prompt caching by maintaining consistent context across requests to reduce costs significantly.