跳到主要内容

GPT-5.6 August 2026 Update: Effort Slider, Free Think Button & the 272K Price Cliff

进阶

On 6 August 2026 OpenAI shipped a mid-cycle refresh of GPT-5.6 that most coverage collapsed into one sentence: "Sol got smarter, free users got unlimited chats." The real update is more interesting — and more consequential if you are a Claude user who occasionally reaches for the OpenAI stack. Three moving parts changed at once: an effort control now runs both in the ChatGPT UI (as a slider) and the API (as a six-tier parameter); a free-tier Think button ships the following week with unlimited Luna chats; and the long-context pricing cliff at 272K input tokens interacts awkwardly with Fast mode, which only runs on Sol. Each shifts a real decision.

This page pulls those decisions out of the release-notes fog and shows what actually changes for a builder who thinks in Claude terms.

What you'll learn
  • See exactly what shipped on 6 August 2026 for GPT-5.6 — and what shipped the week of 10 August for free users
  • Understand the six-tier effort parameter (`none, low, medium, high, xhigh, max`) and map it onto Claude's thinking-effort model
  • Learn the 272K input-token pricing cliff — and how it can silently double a job you thought you had budgeted
  • Know when Fast mode helps, when it does nothing, and which requests it silently refuses
  • Decide when the update is a real reason to switch a workload from Claude to Sol — and when it's not

What actually shipped on 6 August 2026​

Three concrete things. The rest is polish.

Guided walkthrough1 of 3
  1. Plus and Pro users see a new slider on web, mobile and desktop that sets how much reasoning ChatGPT uses per response. Under the hood, the API exposes the same knob as `reasoning.effort` with six positions — `none`, `low`, `medium`, `high`, `xhigh`, `max` — defaulting to `medium` when omitted. This replaces the previous 'model decides' behavior: you now pick the tier and pay for it.

The effort slider — mapping the six tiers to a Claude mental model​

If you have been thinking in terms of Claude's five-level effort setting (introduced with Opus 5 in July), the mental model is similar but the labels don't line up cleanly. GPT-5.6 has six tiers where Claude has five, and the semantics differ at both ends.

GPT-5.6 tierWhat it doesRoughly like Claude Opus 5
noneSkips reasoning entirely; instant answerNo direct equivalent — Claude always plans a little
lowShort chain-of-thought, cheapminimal effort
medium (default)Balanced planning; the shipped defaultlow–medium
highDeeper planning; slower, dearermedium
xhighLong deliberation; multi-step reasoninghigh
maxFull reasoning ceiling — expensivemaximum

Two practical consequences:

  • The default moved. On the API, if you omit reasoning.effort, you get medium. Previously many wrappers relied on the model auto-selecting effort; that behavior is gone. Explicit is now required for predictable cost.
  • none is a real button. It is genuinely useful for classifier prompts, structured extraction, or tool-router calls where reasoning tokens are pure overhead. Claude has no cleanly-equivalent 'off' switch.

A safe request template that pins effort explicitly

{
"model": "gpt-5.6-sol",
"reasoning": { "effort": "low" },
"messages": [
  { "role": "system", "content": "Extract only. Output JSON." },
  { "role": "user", "content": "..." }
]
}

The 272K token pricing cliff — a real hazard​

This is the change most likely to burn a bill unexpectedly.

The rule: if a request's input crosses 272,000 tokens, the entire request is re-priced at 2× the input rate and 1.5× the output rate, not just the tokens above the line. There is no proration.

Two things about that rule matter:

  1. It's a cliff, not a slope. A prompt at 271,999 input tokens pays the base rate. A prompt at 272,001 pays 2×/1.5× on every input and output token in the call — including the reasoning tokens the model generates.
  2. Long conversations creep across it. Multi-turn chats and agent loops accumulate history. A session that starts under 272K can silently cross the line on turn 12 and re-price every subsequent turn.
Guided walkthrough1 of 3
  1. GPT-5.6 Sol base pricing is $5 / $30 per million (input / output), with cached input at $0.50. Above 272K input tokens, the same call costs $10 / $45. On a 300K-token document with 4K output, base cost is $1.62. Cross the cliff and it becomes $3.18 — nearly 2× for one extra token past the threshold.

Fast mode — small print worth reading​

Fast mode is a Sol-only option that trades price for latency: up to 2.5× faster at 2× the API price. Terra and Luna don't offer it.

Three things the launch pitch downplays:

  • 'Up to' is a ceiling, not a norm. Wall-clock speedup depends on your workload — tool calls, network hops and streaming rate all cap the realized gain well below 2.5×.
  • Fast mode and long context don't cleanly mix. OpenAI hasn't published a mutual-exclusivity rule, but early guides report Fast mode requests behave badly (or silently degrade) once you cross the 272K input threshold. Treat them as incompatible until OpenAI documents otherwise.
  • Reasoning tokens still count. Fast mode makes throughput faster, but a max-effort call still generates the same volume of reasoning output. The economics only work when you're paying for latency you can actually feel.
Pro tip
  • Fast mode's break-even is roughly 25% more throughput per dollar at the 2.5× ceiling — worse when the ceiling isn't hit.
  • Prefer Fast mode for interactive chat where the user waits. Skip it for batch, background agents and long-context RAG.
  • If you were going to switch a batch pipeline from Claude to Sol for cost, don't turn on Fast mode — you'll erase the saving.

Free tier: unlimited Luna + the Think button​

For a Claude user this matters less than the API changes, but it's the change most likely to show up in a friend's screenshot and start an argument.

  • Unlimited text chats on GPT-5.6 Luna for free ChatGPT users, from the week of 10 August 2026. Previously capped.
  • Think button on free chats, same window. Adds a deeper-reasoning pass to a specific message, subject to abuse protections. It is not the same UI as the Plus/Pro effort slider — it's a per-message toggle, not a session-wide setting.

The practical read: OpenAI is aggressively defending the free tier now that Claude and Gemini both have generous free reasoning options. Expect the abuse-protection cap to be tight for the first few weeks.

Where this update actually changes a Claude user's decision​

For most Claude workloads the answer stays the same. For three specific patterns it moves.

Guided walkthrough1 of 3
  1. `reasoning.effort: none` on Sol Luna is now a plausibly cheaper option than a min-effort Claude Haiku call for high-volume routing, tag extraction, or structured labelling — provided your prompts stay well below 272K. Benchmark it on a representative sample before switching.

Migration notes for existing GPT-5.6 API code​

If you shipped against Sol before 6 August, two things need to change:

  • Set reasoning.effort explicitly on every call. Omitting it now costs you medium; you probably wanted low for most.
  • Add a 272K guard at the boundary of your pipeline. Reject or chunk requests whose input tokens push past the threshold. A dumb check of len(input) / 4 as an approximation is enough as a first line of defence.

A guardrail one-liner for Node.js wrappers

if (estimateTokens(input) > 260_000) { chunkOrReject(input); }

What this doesn't include​

Two rumors that were circulating around this update are not part of it:

  • GPT-6 (Astra). As of 11 August 2026 OpenAI had not confirmed a launch date, model card, or pricing. It has since shipped: GA on September 4, 2026 at $10 / $50 per MTok with the same 272K pricing cliff described on this page. See the GPT-6 Astra field guide.
  • Sol pricing drop. The base $5 / $30 rate has not changed. Some coverage confused the 272K cliff (a surcharge) with a price change.

Try it​

Check yourself

0/4
  1. You send a 300,000-token input prompt to GPT-5.6 Sol at base pricing. Roughly, what happens to the cost compared to sending 271,000 tokens?
  2. On the API, you omit `reasoning.effort` on a GPT-5.6 Sol call after the 6 August update. What effort tier do you get?
  3. Which of these workloads benefits most from Fast mode?
  4. Which behavior change did OpenAI ship for GPT-5.6 Sol on 6 August 2026?

Sources & further reading​

Cross-links inside AILmanac:

Next​