GPT-5.6 Reasoning Modes Explained - Medium vs High vs Max vs Ultra

GPT-5.6 lets you control how hard the model thinks before it answers. Medium and High set how deep a single agent reasons. Max pushes that same single agent to its deepest setting. Ultra is different: it runs several agents in parallel on separate parts of one job.

If you remember one line, make it this: Max is depth, Ultra is parallelism. And one rule beats every mode choice: start at the lowest setting that gets the job done, then climb only when the result falls short.

We have run these settings on real coding and research work. Below is the direct mapping, then the detail behind each mode.

GPT-5.6 Reasoning Modes at a Glance

Notice the split in the Type column. Medium, High, Extra High, and Max all sit on one reasoning effort ladder, where a single agent just thinks harder. Ultra is not on that ladder at all. It is a separate multi-agent mode. Keep those two ideas apart and the rest falls into place.

The One Rule That Governs Every Mode Choice

Every mode above solves a specific problem. But the smart default is simple: pick the lowest effort that reliably passes your check, and escalate only the step that failed.

This habit matters for two reasons.

• Higher effort helps a model finish analysis it already understands. It cannot rescue a vague prompt.
• The top modes drain usage limits fast. Reaching for them by default wastes time and tokens.

So when a result misses, ask why before you climb. If the model lacks context, fix the prompt. If it ran out of thinking room, raise the effort. If the job has independent parts, reach for Ultra. Match the fix to the failure.

Quick Decision Guide: Pick a Mode in Seconds

Run through these in order and stop at your first yes.

1. Does the task split into parts that run without waiting on each other? Use Ultra.
2. Is it one deep problem that must stay whole? Use Max.
3. Does it have several steps, sources, or tradeoffs? Use High or Extra High.
4. None of the above? Stay on Medium.

The order matters. Check for splittable work first, since Ultra saves the most time there. Then judge depth. Everything else sits comfortably at Medium or High.

What GPT-5.6 Reasoning Effort Actually Means

Reasoning effort controls how much the model plans, checks, and revises before answering. More effort buys more care. It also costs more time and more tokens.

The API exposes six reasoning effort levels, in this order:

• none
• low
• medium
• high
• xhigh
• max

In ChatGPT Work and Codex, the same idea shows friendlier labels: Light, Medium, High, Extra High, and Max. So the app's "Extra High" is the API's "xhigh." Max sits one rung above that.

Here is the part that confuses people. Once you factor in three models (Sol, Terra, Luna), several effort levels, the Work and Codex surfaces, and standard versus fast speed, the combinations run into the dozens. That sprawl is why the naming feels heavier than the actual choice. In practice, you only ever decide two things: how deep one agent thinks, which is the effort ladder from Medium to Max, or whether to split the work across agents, which is Ultra.

GPT-5.6 Medium Reasoning: The Everyday Default

Medium is the sensible starting point for most work. It gives the model room to plan a few steps without dragging out speed or burning tokens. It is the recommended default, and we keep that habit too.

Reach for Medium when:

• The task is clear and well scoped
• You need light planning, not deep analysis
• Speed and cost still matter

A quick insight from daily use: most tasks never need anything above Medium or High. People jump to Max out of reflex, then complain about slow replies and drained limits. Start here, and let the work tell you when to climb.

GPT-5.6 High and Extra High Reasoning: For Multi-Step Tasks

High extends the reasoning past Medium. The model spends longer weighing options, checking its logic, and working through several steps. Extra High is the top standard effort for a single agent, one rung higher again.

Choose High or Extra High when:

• The problem has several moving parts or sources
• There are genuine tradeoffs to reason through
• A wrong answer costs you real time to unwind

In our experience, High is the sweet spot for debugging a tangled function, planning a small feature, or reasoning across a few documents. It thinks harder without the heavy cost of the top modes.

GPT-5.6 Max Reasoning: Deepest Single-Agent Thinking

Max is the top of the effort ladder. It gives the model even more time than Extra High to reason, explore alternatives, run checks, and revise. It is still one agent on one chain of thought. It simply gets the most room to think.

Use Max for problems where depth beats speed, such as:

• A single tricky migration decision that cannot be split
• A difficult financial model or a checkable math problem
• A complex code review or a deep analysis with clear success criteria

There is a practical catch. If you do not see Max, you may need to switch it on in your app settings. It is open to any user with GPT-5.6 access in ChatGPT Work and Codex, and it is the max effort value in the API.

Our rule for Max: use it when the work forms one unbroken chain of reasoning. If splitting the task would break that chain, Max is the right call. If the job divides cleanly, you want Ultra instead.

GPT-5.6 Ultra Mode: Parallel Subagents Explained

Ultra is the mode people misread most, because it is not a higher effort level. Instead of one agent thinking harder, Ultra coordinates several agents at once. By default it runs four agents in parallel. Each takes a separate part of the job, then the system merges their findings into one answer.

This changes the shape of the run. A single Ultra request decomposes the work, fans it out, runs the parts together, and combines the results. You ask once, and the orchestration happens inside that request.

Ultra earns its cost when the work has real boundaries. On a code review, one agent can trace the code path, another can check test coverage, and a third can audit the docs. Those parts are independent, so running them together saves wall-clock time.

Ultra is a poor fit when the parts depend on each other. If agent two must wait for agent one to decide something, you pay for idle workers. That is the overkill case, and it is easy to stumble into.

Worth knowing: Ultra can scale past four agents. Some benchmark setups pushed it to sixteen. Adding parallel agents raised scores and cut time on tasks that split well. So Ultra rewards genuinely separable work, not just any hard task.

Max vs Ultra: The Difference That Actually Matters

This is the comparison most people search for, so let us make it clean.

They can sit side by side in the picker, but they are not the same dial turned up twice. Max keeps one agent on one problem longer. Ultra sends separate agents at separate parts, then reconciles them. Pick based on whether the work is one chain or many branches.

Where You See Max and Ultra

Standard ChatGPT chat stops at Extra High, then Pro. Max and Ultra only appear in ChatGPT Work, Codex, and the API. This table is all most readers need.

So do not hunt for Max or Ultra in the ChatGPT model picker, since it caps at Extra High and Pro. Because these top modes sit behind the higher tiers, it helps to compare Plus and Pro before you pick a plan. Codex users get Ultra in Codex on Plus and above.

What Each GPT-5.6 Reasoning Mode Costs

The cost gap between Max and Ultra is wider than it looks, and this is where budgets slip.

Max runs one agent. It uses more tokens than lower effort, but the growth stays inside that single chain. Ultra runs several agents, and each one produces its own reasoning and output tokens. Those tokens stack across every subagent. So one Ultra call can burn far more than one Max call on the same prompt.

For context, GPT-5.6 pricing per one million tokens looks like this.

Sol output sits at $30 per million tokens. Now picture four Sol agents each producing reasoning and output. The bill climbs quickly, and reasoning tokens count as output. Some benchmark setups pushed Ultra to sixteen agents, which multiplies the burn further.

Two ways to keep this sane:

• Set a per-session cap on Ultra so one request cannot quietly run away.
• Use prompt caching. If your agents share a big fixed context like a codebase, cache it once and let them read it cheaply.

Also worth a look: the model tier itself. Most tasks do not need Sol. Defaulting to Terra, with Luna for bulk work, trims real money off a monthly bill.

Official vs U7BUY Subscription Prices

The token prices above apply to the API. If you use these modes inside ChatGPT, Work, or Codex, you pay by monthly subscription instead. That plan is what unlocks Medium, High, Extra High, and the Max and Ultra options.

Here at U7BUY, we line our ChatGPT subscription plans up against the official monthly rates, so you can weigh the value yourself.

A few honest notes on that table. Our Go plan lands just under the official rate. Plus sits at parity, so you pay the same and still get instant delivery from us. Our Pro matches the lower official Pro tier, and our Business price undercuts the official monthly seat rate. You can also buy several months at once, with no annual lock-in. Stock moves fast, so check the store for what is live right now.

Which GPT-5.6 Reasoning Mode Should You Use?

Here is the mapping we actually follow. Match the mode to the task, not to habit.

The escalation path is simple. Start at Medium. If the reasoning is thin, move to High or Extra High. If one hard step still fails and cannot be split, raise that step to Max. If the task divides into independent parts, use Ultra.

Mistakes to Avoid With GPT-5.6 Reasoning Modes

• Treating Max and Ultra as the same button turned up twice. They solve different problems.
• Running Ultra on dependent tasks, then paying for idle subagents.
• Jumping to the top mode by default and draining usage limits fast.
• Expecting these modes to fix a weak prompt. They add thinking, not context.
• Looking for Max or Ultra in standard ChatGPT chat, where neither appears.
• Letting the gpt-5.6 alias route you to pricey Sol when Terra or Luna would do.

Conclusion

GPT-5.6 looks more complicated than it is. Once you separate depth from parallelism, the choice gets simple. Medium and High cover almost everything. Max is for one deep problem that must stay whole. Ultra is for big jobs that break into clean, independent parts.

Our advice holds across every project. Start low, check the output, and climb only when the work demands it. That single habit gives strong results without the slow replies and drained limits that come from reaching for the top mode too soon.

And when you decide to step up to the higher tiers, we are glad to help you get there without overspending.

Frequently Asked Questions