OpenAI dropped GPT-6.2 Sol on November 17, 2026, positioning it as the go-to engine for everyday coding, writing, and agent workflows. The model is now live in ChatGPT Work, Codex, and the API under the identifier gpt-6.2-sol. While it sits in the shadow of the flagship Astra, Sol achieves a 57.3 score on the Artificial Analysis Intelligence Index at max effort, ranking #5 globally. The real story isn't just the score; it's the price tag of $2 per million input tokens and $10 per million output tokens, which is a quarter of Astra's cost.
The Performance-Price Arbitrage
For developers watching token burn rates, this release is a direct challenge to the assumption that high reasoning requires high spend. Sol maintains the exact API pricing of its predecessor, GPT-6.1 Sol, despite the intelligence jump. OpenAI claims Sol lands within 0.5 points of GPT-6.2 Astra on Cognition's FrontierCode 2.0 benchmark, which tests code merge readiness, while costing roughly one-fifth as much in that specific evaluation. With a 1,050,000-token context windowβallowing up to 922,000 input and 128,000 output tokensβSol offers massive headroom for long-context tasks without the premium tier surcharge, provided you stay under the 272,000 input token threshold where pricing escalates.
Reasoning Efforts and Hidden Latency
The model introduces a granular reasoning effort system that forces users to trade speed for accuracy. At 'Max' effort, Sol hits 57.3 on the Intelligence Index but takes 58 seconds to produce the first token and averages 85 tokens per second thereafter. Dropping to 'Medium' effort reduces the score to 53 but cuts first-token latency to 2.7 seconds. Notably, Sol supports an 'Ultra' effort level exclusively within Codex CLI, the Codex app, and ChatGPT Work, enabling parallel subagent delegation for complex tasks. However, the API rejects 'none' or 'minimal' reasoning settings, signaling OpenAI's push to standardize on low-effort baselines for routine queries.
Hallucination Rates and Agent Metrics
Don't let the headline score blind you to the caveats. Artificial Analysis data reveals a 55% hallucination rate at max effort on the AA-Omniscience benchmark, which is concerning for autonomous agents. Furthermore, Sol scores 3 points below GPT-6.1 Sol on the Coding Agent Index, suggesting that while raw reasoning improved, specific agentic coding reliability may have regressed slightly or plateaued. The model is text-and-image input, text output only, and requires the Responses API for tool calling, a structural shift that might break legacy integrations built on older completion endpoints.
Key Takeaways
- GPT-6.2 Sol ranks #5 on the Artificial Analysis Intelligence Index with a score of 57.3 at max effort.
- Pricing remains at $2/$10 per million tokens, offering flagship-class performance at a fraction of GPT-6.2 Astra's cost.
- The model features a 1.05M token context window with higher pricing triggered above 272,000 input tokens.
- A 55% hallucination rate at max effort on AA-Omniscience warrants caution for high-stakes autonomous tasks.
- 'Ultra' reasoning effort is restricted to Codex and ChatGPT Work, not available via the standard API.
The Bottom Line
GPT-6.2 Sol is a pragmatic workhorse that delivers flagship-class scores at a fraction of the price, but the regression in coding agent metrics and high hallucination rates mean it should not be trusted blindly for autonomous loops. Developers should benchmark their specific workflows to see if the latency trade-offs at max effort are worth the cost savings, or if the 'Medium' effort tier offers a better balance for routine tasks.