No. 122
2026-11-18
the AI pulse — brewed, not raw
 AI CRITIQUE

OpenAI introduces GPT-6.2 Astra and Sol models at low cost

Read the article    English (automatic translation)

Summary

OpenAI launched its flagship GPT-6.2 Astra and the everyday-use GPT-6.2 Sol on November 17, 2026. The company claims these models deliver superior performance at roughly one-third the cost per task compared to previous generations, while API prices have generally dropped between 20% and 60%.

Why it matters

This aggressive price-performance balance is putting pressure on the entire frontier model market. It forces competitors like Anthropic to justify their premium pricing. On the other hand, the modest 5-point increase in the intelligence index and flat or negative coding scores indicate that the real story is cost engineering rather than a leap in capability. This is reshaping how companies plan their budgets for agentic workloads.

Highlights

  • Agents' Last Exam: Success rate on GPT-6 Astra improved from 73.1% at a cost of over $8.67 per task to 78.3% at $2.89. This represents an approximate 3x cost reduction in the same comparison.
  • Intelligence Index: GPT-6.2 Astra scored 62 points, trailing Claude Fable 5.2's 65 points. GPT-6.2 Sol tied the previous Astra version at 57 points.
  • Coding Agent Index: GPT-6.2 Astra is at the same level as the previous generation. GPT-6.2 Sol is 3 points lower than GPT-6.1 Sol.

Critical take

The 'one-third cost' headline highlights a single agentic benchmark. However, on the broader intelligence index, GPT-6.2 Astra still lags behind Anthropic's Fable 5.2. Coding performance is stagnant or has declined. This serves as a reminder that cheaper tokens do not automatically translate to better engineering output. Furthermore, the claim of 'the strongest security measures in history' creates a tension that needs monitoring, especially as third-party security testing is restricted and these models are rolled out to enterprise customers.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue