openai / gpt-6-astra

Overview

GPT 6 Astra is OpenAI’s flagship reasoning model for hard, end-to-end work. It accepts text and images, then produces high-accuracy text responses, including structured JSON when you ask for it. It’s trained to think through multistep tasks and coordinate tool calls, which makes it useful for agent workflows that must plan, act, and verify.

It supports very large prompts with a 1,050,000-token context window. It can generate up to 128,000 output tokens in one response. Its published knowledge cutoff is April 30, 2026.

What you can build

• Visual Q&A for screenshots, diagrams, UI mocks, and product photos
• Document and slide review workflows where images need precise critique and edits
• Coding agents that plan changes, call tools, and produce review-ready patches
• Research assistants that browse the web, summarize findings, and cite sources you provide
• Cybersecurity triage helpers for logs, attack-path reasoning, and remediation checklists
• Structured extraction that turns messy notes into JSON for downstream systems

Inputs

• A required instruction in plain text that describes the task and success criteria
• Optional image uploads (you can send multiple images in one request, up to 10)
• Optional chat-style conversation history as a JSON array of role-based messages; use this instead of a single prompt when you need multi-turn context
• Optional system-level instructions that set global behavior, tone, and boundaries for the run
• Optional identifiers for the user and session when you want the platform to group turns into a single chat history
• Optional reasoning depth setting with 5 levels (Low, Medium, High, Extra High, Max) to trade speed for deeper deliberation
• Optional verbosity control (Low, Medium, High) to shape how concise or detailed the answer is
• Optional cap on the maximum length of the generated answer, up to 128,000 tokens
• Optional tool definitions in JSON Schema form when you want the model to call your functions
• Optional tool routing preferences, including forcing a specific tool, allowing no tools, or letting the model decide
• Optional setting to allow multiple tool calls in the same turn
• Optional continuation inputs for tool-using https://wiro.ai a prior task reference plus the tool results from earlier calls

Outputs

The model returns a structured response object that can include:

• One or more assistant messages with generated text
• Structured JSON output when you request a schema-bound format
• Tool call requests (function name plus arguments) when you provide tool definitions and the task needs actions
• Metadata that reflects the selected reasoning depth and the token usage for the run

Text is the primary output modality. Images are supported as input only.

Recommended settings

• Start with Medium reasoning for most multi-step tasks and professional work.
• Use Low reasoning for short answers, classification, and quick edits.
• Move to High or Extra High when the task needs planning, proof-style reasoning, or careful verification.
• Use Max only for the hardest agent workflows, deep debugging, or long decision trees.

GPT 6 Astra doesn’t support a “no reasoning” mode. If you previously relied on that, begin with Low.

Limitations

• Audio input and output aren’t supported.
• Video input and output aren’t supported.
• Image generation is a separate capability from image understanding. This model treats images as input for analysis.
• Fine-tuning isn’t supported for GPT 6 Astra.
• Low-quality images can break analysis. Blurry photos, tiny text, heavy compression, and cropped UI states can cause mistakes.
• Messy context can degrade results. Conflicting instructions, long unstructured chat logs, and mixed task goals can reduce reliability.

Safety & compliance

OpenAI positions GPT 6 Astra as a major step up in cybersecurity capability. It ships with stronger protections against harmful cyber actions and stronger jailbreak resistance than earlier models.

Follow your organization’s policies for sensitive data. Don’t submit private credentials, secrets, or regulated personal data unless you’re authorized to do so.

Use the model for legitimate security and engineering work. Don’t request instructions for wrongdoing, exploitation, or evasion of safeguards.