Where Is GPT-6’s Progress Bar?
OpenAI’s next-generation flagship model GPT-6 has completed pre-training at the Stargate data center and officially entered the safety alignment phase. According to OpenAI’s release cadence, this means the official version may arrive within weeks to months.
Meanwhile, some capability data for GPT-6 has been made public:
| Metric | GPT-6 | GPT-4o | Improvement |
|---|---|---|---|
| Math reasoning | 92.5% | ~75% | +17.5pp |
| Code generation pass rate | 96.8% | ~85% | +11.8pp |
| Professional tasks at human expert level | 83% | ~60% | +23pp |
| Parameter scale | 5-6 trillion | 1.8 trillion | ~3x |
Symphony Architecture: Three-in-One Super Application
GPT-6 is not just a model upgrade — it’s a restructuring of product architecture.
OpenAI introduced the Symphony architecture, integrating three major products — ChatGPT, Codex, and Atlas — into a single entry point. This means:
- Previously, you needed to switch between three tools.
- Now, one entry point handles everything: conversation, code execution, physical world perception.
This isn’t a simple feature merger — it’s deep integration of underlying capabilities. The Symphony architecture allows GPT-6 to understand the user’s complete intent, freely switching between conversation, coding, and reasoning without the user manually specifying “I’m now using Codex mode.”
Pricing Signal: The AI Ceiling Is Far Away
GPT-6’s API pricing has been leaked:
- Input: $2.5/million tokens
- About 10x more expensive than GPT-4o
This pricing strategy sends a clear signal: OpenAI doesn’t plan to fight a price war — it’s using capability premium to define the high-end market. Enterprises and individuals willing to pay for top-tier AI capabilities are GPT-6’s target customers.
”AGI Deployment Division”: OpenAI’s Internal Signal
More noteworthy is an internal change at OpenAI: the product department has been renamed to “AGI Deployment Division.”
Whether you believe in AGI or not, OpenAI is all in. This renaming isn’t a PR move — it’s a strategic adjustment at the organizational level, meaning OpenAI believes AGI has moved from a “research goal” to a “deployable product.”
Benchmarking Against Competitors
GPT-6’s entry into the safety alignment phase coincides with multiple competitors’ release windows:
| Model | Status | Benchmarking GPT-6 |
|---|---|---|
| Claude Mythos Preview | Released | Anthropic still leads in May benchmarks |
| GPT-5.5 | Released (Cyber + Ultra variants) | OpenAI’s transition model |
| Gemini 3.5 Pro | Teased | Google’s response |
| Sonnet 4.8 | Coming soon | Anthropic’s mid-tier product |
| DeepSeek V4 | Open-source available | Costs only 1/3, capability gap narrowing |
| MiniMax M3 | Coming soon | Chinese model’s office scenario differentiation |
Market Outlook
The significance of GPT-6 lies not only in its own capabilities but also in the industry pace it defines:
- The parameter race is not over yet: 5-6 trillion parameters means Scaling Law is still valid, but marginal returns are diminishing.
- Safety alignment is the new bottleneck: The stronger the model, the harder the alignment. The safety alignment phase may take longer than pre-training.
- AGI deployment moving from concept to practice: OpenAI’s organizational adjustment shows that industry leaders are already preparing for AGI-scale deployment.
Action Recommendations
- API users: Watch GPT-6’s pricing and API specification changes — the $2.5/M token pricing may push enterprises to re-evaluate AI cost structures.
- Developers: Research the multimodal capability integration of the Symphony architecture — this may be the standard model for future AI applications.
- Enterprise decision-makers: Evaluate the impact of AGI deployment on existing business processes, especially in high-value scenarios like customer service, R&D, and data analysis.
GPT-6’s safety alignment phase is a key observation window — it will tell us how hard AI safety really is when model capabilities approach human expert levels.