GPT models

OpenAI’s new ChatGPT-6 Astra model claims significant advances in autonomous task completion and multi-step reasoning, marking a claimed milestone toward artificial general intelligence. The model will reach ChatGPT Plus subscribers at $20 per month, with enterprise versions rolling out immediately.

• Astra achieved 41.4% accuracy on multi-step task completion, compared to 18.1% for GPT-5.6 Sol
• Standard Astra included in existing $20-per-month ChatGPT Plus plan starting rollout September 3, 2026
• OpenAI president Greg Brockman suggested people may later identify this model as the arrival of AGI

• 41.4% Multi-step task completion rate, versus 18.1% for prior GPT-5.6 Sol
• 97.6% Extremely difficult math problem accuracy, versus 83.0% for GPT-5.6 Sol
• $20 Monthly cost for ChatGPT Plus with standard Astra access
• Sept 3, 2026 Official release date for enterprise and selected business accounts

OpenAI announced ChatGPT-6 Astra on September 3, 2026, describing it as a model capable of performing complex cognitive work at human level across multiple domains simultaneously. The system can browse webpages, complete forms, work through spreadsheets, and execute multi-step tasks with reduced user intervention, according to OpenAI’s demonstration video and benchmark data. Astra represents a significant shift in AI capability focus, moving from models that primarily answer questions toward systems designed to execute actions directly on computers and software platforms.

This transition reflects a broader evolution in the artificial intelligence industry. Earlier generations of large language models excelled at knowledge retrieval and text generation but struggled with sustained task execution requiring multiple decisions and environmental interactions. Astra’s architecture appears designed to maintain context across complex workflows while making sequential decisions that depend on previous actions and outcomes.

Astra’s Performance Gains in Autonomous Task Execution

OpenAI’s internal benchmarks highlight Astra’s largest improvements in areas where prior AI systems have underperformed: autonomous action and reasoning across extended workflows. On multi-step task completion, Astra scored 41.4% compared to 18.1% for GPT-5.6 Sol, a gap of more than 23 percentage points. The model also achieved 97.6% accuracy on extremely difficult mathematics problems, up from 83.0% in the previous version.

Advanced coding tasks reached 74.1% accuracy versus 70.8% for GPT-5.6 Sol, while scientific computing work jumped to 64.6% from 22.4%, a gain of 42 percentage points. Image-to-3D design conversion reached 95.9% accuracy compared to 83.3% in the prior model. OpenAI notes that Astra became the first company model to meet its “critical” cybersecurity threshold, though the company has not specified what that threshold entails.

These performance improvements suggest Astra has better learned how to decompose complex problems into sub-steps and track state changes across extended interactions. The gains in scientific computing and multi-step reasoning indicate the model may better understand domain-specific logic and requirements than predecessors, potentially valuable for research, engineering, and professional applications.

Notably, these benchmarks come from OpenAI’s own testing and vendor-reported evaluations, not independent third-party verification. The industry standard for evaluating major AI claims involves peer review and external testing, which has not yet occurred with Astra.

Pricing and Rollout Strategy Across User Tiers

Standard Astra access rolls out to existing ChatGPT Plus subscribers at no additional cost beyond the $20-per-month subscription fee. OpenAI is deploying the model first to enterprise and cybersecurity customers, with availability expanding to Plus, Pro, Business, and Enterprise account holders over the following days. Free tier users will not receive access to Astra, according to OpenAI’s announcement.

A higher-tier variant called Astra Pro remains restricted to Pro, Business, and Enterprise customers, positioning the $20 Plus plan as the entry point for individual users seeking Astra capabilities. OpenAI has not announced pricing for the Pro version or detailed feature differences between standard and Pro tiers. This tiered approach reflects OpenAI’s broader strategy of using subscription levels to segment access to advanced capabilities.

The inclusion of Astra in the existing Plus tier without a price increase represents a competitive positioning move against other AI companies developing similar autonomous reasoning systems. Industry observers note this pricing strategy may pressure competitors offering comparable capabilities at higher price points.

The AGI Question and Unresolved Technical Caveats

OpenAI has deliberately avoided formally declaring Astra as true artificial general intelligence, despite the framing in public messaging. President Greg Brockman characterized AGI as a “gray, fuzzy thing” while stating his personal belief that people may eventually view this model as the moment AGI arrived.

This is the AGI era, and Astra might be the first model there.

Greg Brockman, OpenAI President

The distinction matters because artificial general intelligence, as typically defined by researchers, refers to AI systems capable of understanding, learning, and applying knowledge across any intellectual domain without domain-specific training. No widely accepted consensus exists on whether Astra meets such criteria or represents incremental improvements in narrow task domains.

A significant methodological limitation underlies Astra’s most striking results. The model’s impressive performance on complex benchmarks was achieved using an OpenAI agent setup equipped with memory and external tools surrounding the base model. This architecture makes it difficult to isolate Astra’s raw reasoning capability from the broader system’s performance, raising questions about what precise component drove the benchmark gains.

Critics and researchers note that such “scaffolding” systems combine multiple components whose individual contributions become hard to measure. When an AI system has access to external tools, memory systems, and iterative correction capabilities, performance improvements may come from these architectural additions rather than core model advances. This distinction becomes crucial when evaluating claims about progress toward AGI, which would require the base model itself to demonstrate general reasoning ability rather than relying on supporting infrastructure.

Independent testing of Astra’s capabilities remains pending, and OpenAI’s own evaluation methodology will likely face scrutiny from researchers and competitors seeking to verify whether the model’s autonomous task performance genuinely represents a step toward AGI or reflects architectural scaffolding that inflates measured capability beyond the underlying model’s reasoning power.