Improving GPT‑5.6 Sol in ChatGPT—and expanding access to GPT-5.6 Luna for free users

Source: OpenAI Date: 2026-08-06 URL: https://openai.com/index/improving-gpt-5-6-sol-in-chatgpt

Summary

OpenAI retuned GPT-5.6 Sol for Plus/Pro users toward more direct, less-hedging answers (claimed ~68% fewer factual errors than GPT-5.5 Instant) and added a Reasoning Slider so users pick how much compute a response gets before it’s generated. Separately, GPT-5.6 Luna becomes the default model for Free and Go users with unlimited text chats (file uploads/images/tools stay capped), plus a Think button rolling out next week for harder questions under abuse guardrails.

Implications

Squarely on the cost/economy of AI thread: this is the productization half of the price-cutting OpenAI has been running since the GPT-5.6 Luna −80% cut — free-tier text chat going unlimited is the cost curve reaching the point where OpenAI treats basic conversational capacity as a loss-leader rather than a metered resource, matching the “abundant intelligence” framing already tracked. The Reasoning Slider is a cost/economy feature at the UI layer too — exposing the compute-vs-quality tradeoff directly to the user rather than fixing it per-model, the same “harness config as capability multiplier” pattern already seen in ARC-AGI compaction settings. Modest fit with agent-layer convergence: correcting the user instead of “agreeing when it wouldn’t help” is a small trust/verifiability nudge inside the chat product itself.