Tech • AI • Robotics • Game
It's scary! I tested GPT 6.1 and it's crazy!
9.4/10TL;DR
GPT-6.1 Sol emerges as the strongest all-around AI model on cost-adjusted performance, approaching top-tier capability while undercutting rivals on price and token efficiency.
KEY POINTS
A new pricing ladder
OpenAI now fields three main tiers: GPT-6 Astra at $10 per million input tokens and $50 output, GPT-6.1 Sol at $2 input and $10 output, and GPT-6 Luna at $0.10 input and $0.50 output. The shift signals a broader market trend: higher intelligence is becoming cheaper, with mid-tier models increasingly matching older flagship systems on practical work.
Why GPT-6.1 Sol stands out
GPT-6.1 Sol reportedly lifts capability from an internal level of 36 to 48, a gain of 12 points while cutting production cost. In cost-to-intelligence terms, it becomes especially attractive at medium and high reasoning levels, where it is presented as more profitable than GPT-6 Astra despite Astra’s premium positioning.
Competition with Anthropic models
Against Opus 5.5, GPT-6.1 Sol is described as roughly on par in intelligence, trailing or leading by about a point depending on the benchmark, but with a major efficiency advantage. For equivalent work, GPT-6.1 is said to use 7 to 8 times fewer tokens than Opus 5.5, sharply reducing operating cost. Claude Fable 5.1 is positioned closer to Astra in pricing dynamics, but with heavier token use.
Gap with Chinese challengers
The comparison places Kimi K3 and GLM 5.3 notably behind the leading American models. Their capability is framed around 34 to 35 points, versus 50-plus for top U.S. systems. Kimi was initially perceived as a cheaper alternative, but its token consumption and lower intelligence level, cited at 44, weaken that advantage in real workloads.
Benchmarks matter more than headline scores
The most useful reading of an AI benchmark centers on three questions: the task, the reliability of the result, and the cost of reaching it. That framing is increasingly important as vendors advertise high scores that may hide expensive inference or weak performance on real tool use, coding, and constrained workflows.
Software autonomy improves sharply
On a software-engineering benchmark described as Deep SWE, GPT-6.1 Sol gains more than 10 points in autonomy over GPT-6 Sol in less than a week of model evolution. It is said to come close to GPT-6 Astra while costing about five times less, making it one of the strongest options for coding, automation, and no-code project work.
The profitability ceiling
There is a reported break-even point beyond level XI, where profitability drops across models because extra reasoning starts to overthink tasks. For many difficult jobs, XI is the best compromise, but for everyday use the High setting may deliver better returns by avoiding unnecessary token burn.
Real business use: autonomous invoice handling
A detailed business example shows an agent using GPT-6.1 Sol to process invoices, check Stripe records, verify customer identity, determine VAT treatment inside or outside the European Union, detect installment payments, generate deposit invoices, and archive compliant documents in Google Drive. A task that previously took 30 to 40 minutes per invoice is reduced to an automated workflow, with one email preparation-and-send sequence taking about eight minutes.
Human review remains essential
Even with improved autonomy, high-responsibility workflows still require pause-and-check stages before execution. The recommended pattern is to split analysis and preparation from final sending or filing, allowing a person to validate outputs before legal, financial, or tax actions are completed.
PDF understanding is still weak
One of the most important limitations remains document comprehension. GPT-6 Sol is cited at about 24% confidence on PDF relationship understanding, while GPT-6.1 Sol rises to roughly 30%. GPT-6 Astra reaches about 40%, and Opus 5.5 around 32%, but at much higher cost. Raw PDF ingestion therefore remains unreliable for complex professional tasks unless document structure and relationships are explicitly engineered.
Security and restrictions improve
GPT-6.1 Sol is presented as substantially safer than GPT-6 Sol for tool use and system boundaries, reportedly respecting restrictions three times more often and being twice as secure in everyday use. By contrast, GPT-6 Luna is flagged as unsuitable for high-responsibility tasks, with a failure rate described as near 30%.
CONCLUSION
For most users and businesses, GPT-6.1 Sol currently offers the best balance of price, capability, and operational safety. The larger shift is not just toward smarter models, but toward affordable AI agents that can execute real business workflows with supervision rather than simple prompting.
Ask a question