OpenAI's New Models: Why Cheaper Doesn't Mean Smarter

The Cost-Performance Reality
OpenAI’s latest release, GPT-6.2 Astra and Sol, aims to dominate the Pareto frontier—the balance where no other model is both cheaper and more capable. According to The Decoder, Astra’s API pricing is 20% lower than its predecessor, charging $8 per million input tokens. Sol maintains a more accessible price point at $2 per million input tokens.
However, the performance gains are uneven. While OpenAI touts massive improvements in scientific workflows and legal documentation, the reality for developers is more nuanced. Independent data from Artificial Analysis indicates that coding performance has flatlined, with the Sol model actually seeing a minor dip in benchmarks compared to the previous version.
Where the Models Fall Short
The most critical takeaway for professional users is the discrepancy between internal benchmarks and real-world utility. Epoch AI, an independent research organization, noted that Astra struggled with simpler research projects, often playing it too safe and omitting key estimates. This suggests that while the model is technically more efficient, it may lack the initiative required for autonomous research tasks.
Furthermore, the hallucination rate remains a significant hurdle. Under maximum effort, Astra exhibits a 59% hallucination rate on unusually hard factual questions. This metric, tracked by Artificial Analysis, measures how often a model guesses rather than admitting it lacks the information.
How to Access the New Models
The models are currently available across several tiers and platforms:
- For End Users: Available for Plus, Pro, Enterprise, and Edu subscribers within ChatGPT Work and Codex.
- For Developers: Accessible via the standard OpenAI API.
- For Power Users: Pro-500 subscribers gain access to the "Ultrafast" mode. Pro-200 subscribers now also receive Ultrafast access, specifically for the GPT-6.2 Sol model.
Verdict: Who Should Use Them?
If your workflow relies on scientific reasoning, 3D design simulation, or high-volume data interpretation, the cost-to-performance ratio makes Astra a compelling choice. The efficiency gains in output tokens can significantly reduce operational costs for specialized tasks.
However, if your primary goal is software development or complex legal research, proceed with caution. The models show signs of stagnation in coding benchmarks and struggle with the nuance required for high-stakes professional research. Test your specific use cases before migrating your entire production pipeline to the new architecture.
Sources
Frequently asked questions
- Are GPT-6.2 Astra and Sol cheaper than previous models?
- Yes. Astra's API price is 20% lower than its predecessor, while Sol maintains the same pricing structure as the previous Sol model.
- Should I use GPT-6.2 for coding projects?
- Independent benchmarks suggest coding performance has flatlined, with some tests even showing a slight decline compared to previous versions.
AI Briefing in 3 Minutes
Get what actually matters in AI & tech delivered free to your inbox. Zero spam, 1-click unsubscribe.
Comments
0 commentsDeixe seu comentário
Be the first to comment.
Continue Reading

NATO Is Consuming 300 Billion AI Tokens a Month — Here's Why
NATO adopts OpenAI tool for cyber defense

OpenAI Sued After AI Agents Hack Companies More Than 80 Times
California sues OpenAI over rogue AI agents

Can Human Defenders Stop AI Hacking Tools Operating in Under 3 Hours?
FBI warns AI hacking tool breaches networks fast