GPT-5.6 Sol Gets a Reasoning Slider in ChatGPT

> OpenAI merged Instant and Thinking into GPT-5.6 Sol with a reasoning slider for Plus/Pro, while free users move to GPT-5.6 Luna as the new default model.

ShareX LinkedIn

🎧 Listen — ~6 min

Ready · GPT-5.6 Sol Gets a Reasoning Sli

0:00 / 6:00

Verified by Essa Mamdani

OpenAI pushed a significant ChatGPT update on August 6-7, 2026, and it changes how tens of millions of people will interact with the assistant every day. The company merged its "Instant" and "Thinking" experiences into a single model, GPT-5.6 Sol, controlled by a new reasoning slider — while quietly moving free-tier users onto a different, weaker model behind the scenes.

What actually shipped

For Plus and Pro subscribers, GPT-5.6 Sol now powers both quick replies and deep, multi-step reasoning. Previously, users had to pick a mode in advance: fast-but-shallow GPT-5.5 Instant, or slower GPT-5.6 Thinking. That binary choice is gone. In its place is a slider in the composer with five settings — Instant, Medium, High, Extra High, and Pro — that lets you dial reasoning effort up or down mid-conversation without switching models.

OpenAI says the change reduces factual errors by roughly 68% in finance, medicine, and legal domains compared with GPT-5.5 Instant, tightens formatting (less bullet-point sprawl, fewer unnecessary headers), and makes Sol more willing to push back instead of reflexively agreeing with the user. The company frames this as closing the "personality gap" between quick answers and long research sessions — the tone and style no longer shift depending on which mode happened to answer.

Free and Go users are being moved to GPT-5.6 Luna as the new default, with a "Think" button arriving for harder questions. Critically, Think on the free tier runs on Luna, not Sol — the cheapest, fastest model in the 5.6 family. Free users get unlimited text chats, which sounds generous, but the practical tradeoff is that free and paid users are now running meaningfully different models rather than the same model with different rate limits. Go users can enable Thinking manually but are capped at 10 messages per five hours, and that mode still doesn't touch Sol.

Why the "Instant vs Thinking" merge matters

The bigger structural story here isn't the specific benchmark numbers — it's OpenAI abandoning the idea that users should have to know, in advance, how hard a question is. Manually picking "Thinking mode" for a hard prompt requires the user to correctly predict difficulty before seeing the answer, which is exactly the kind of task LLM interfaces are supposed to remove from people. A continuous slider — or better yet, an automatic switcher, which is also part of this release — puts that judgment back where it arguably belongs: with the model, not the user.

Automatic switching is opt-out, not opt-in. On eligible paid plans, ChatGPT will now silently escalate from Instant to Medium reasoning when it detects a complex request, and users can turn this off from a "Configure" option in the model picker if they want manual control. This is worth flagging for developers building repeatable workflows against ChatGPT's UI (not the API) — automatic reasoning escalation means the same prompt phrased the same way might get materially different treatment depending on how ChatGPT scores it that day, and there is currently no user-facing signal that tells you which reasoning tier actually ran unless you check usage logs.

Pricing and access tiers

This release also draws a much sharper line between ChatGPT plans than before. According to OpenAI's help center documentation:

Free and Go: GPT-5.6 Luna only, no Sol access at any reasoning level.
Plus: Sol at Medium and High reasoning; no Extra High, no Pro.
Pro: Sol at every level, including Extra High and Sol Pro for the longest-running tasks.
Business and Enterprise: full access, subject to admin-configured workspace limits.

Notably, GPT-5.6 Terra and Luna are not selectable inside standard ChatGPT conversations regardless of plan — they're only reachable through ChatGPT Work, Codex, or the API. That keeps the standard chat interface simple (three effective tiers instead of five separate models) while preserving Terra as the cost/capability middle ground for API users building agents, and Luna as the cheap, high-volume workhorse.

If you hit a GPT-5.6 reasoning limit, ChatGPT quietly falls back to GPT-5.4 Thinking mini rather than blocking the request outright — a detail that matters if you're benchmarking response quality and don't realize a fallback occurred mid-session.

What this means for API and agent builders

For anyone building on the API rather than the ChatGPT UI, nothing about model availability changed — Sol, Terra, and Luna are all still directly selectable, and none of the slider/auto-switching behavior applies; that's a ChatGPT product-surface feature, not a model-family change. But the pricing signal is worth noting: Luna's positioning as an extremely cheap, high-volume model (OpenAI has priced it well under $1 per million output tokens in prior announcements) makes it a natural target for classification, extraction, and other high-throughput tasks where Sol-level reasoning would be wasted spend. Terra remains the practical default for routine coding and agent orchestration where you want Sol-adjacent quality without Sol's cost profile.

The competitive backdrop matters too. This update lands the same week Alibaba's Qwen3.8 Max (a 2.4-trillion-parameter MoE with a 1M-token context) and reports of an imminent Claude Fable 5.1 have kept pressure on OpenAI's pricing and capability claims. The GPT-5.6 Sol changes read as a direct response to that pressure: cut hallucinations, unify the UX, and make sure free users still feel like they're getting something better than before — even though what they're getting is Luna, not Sol.

Rollout mechanics and what to check first

The update is landing gradually rather than all at once, and OpenAI has been explicit that account-level availability lags the announcement by days in some cases. If you don't see the slider yet, check the model picker for a "Configure" option before assuming your account hasn't been touched — some workspaces on Business and Enterprise plans still show the legacy picker until the new controls propagate to that tenant. Codex users need a minimum client version to see any of this (desktop app 26.707.30751 or Codex CLI 0.144.0), so an outdated local install is a common reason someone thinks they're on GPT-5.6 Sol when they're actually still running the previous default.

Regional availability follows the existing ChatGPT and API country lists rather than introducing a new one, with one wrinkle: GPT-5.6 is not currently supported for workloads pinned to UAE inference residency, even though UAE-based users can access it without that residency setting. That's a narrow edge case, but it's the kind of detail that breaks a compliance-sensitive deployment if nobody catches it during rollout.

The practical takeaway

If you're a Plus or Pro user and haven't noticed a new slider in your composer yet, the rollout is gradual and may not have reached your account. If you're building products against ChatGPT rather than the API, budget time to test both manual reasoning selection and the auto-switch behavior, since the two can diverge from what you'd expect a single "model" to do. And if you're comparing free-tier ChatGPT against a paid competitor for an internal eval, make sure you're actually comparing against Sol — as of this week, free ChatGPT users are one weight class down, permanently, by design.