GPT-5.6 offers two ways to spend more compute on difficult work: max reasoning and ultra mode. The names suggest a simple ladder—ordinary, max, then ultra—but they represent different execution strategies.
Max gives a GPT-5.6 model more time to reason, explore alternatives, check its work, and revise. Ultra coordinates several agents across parallel workstreams. One goes deeper inside a single reasoning process; the other goes wider by dividing work.
OpenAI introduced both with the GPT-5.6 general-availability release on July 9, 2026. Max appears in the documented API reasoning options and is also available in ChatGPT Work and Codex for users with GPT-5.6 access. Ultra is presented as a product mode for eligible ChatGPT Work and Codex plans. For API builders, OpenAI separately offers Multi-Agent orchestration as a beta. It is safer to keep those product and API concepts distinct than to assume “ultra” is an undocumented API reasoning value.
What max reasoning actually changes
The GPT-5.6 model catalog lists reasoning efforts from none through low, medium, high, xhigh, and max. OpenAI says max gives the model more time than xhigh to explore alternatives, run checks, and revise.
That makes max a good fit for problems with a tightly connected chain of thought:
- debugging a failure that crosses several layers of one system;
- reconciling contradictory requirements into one coherent design;
- constructing a proof or technical argument with dependent steps;
- reviewing a high-impact plan for hidden assumptions;
- refining one artifact against many interacting constraints.
In these tasks, splitting the problem can destroy useful context. The result of step three changes how step two should be interpreted, or several constraints must be held together throughout. One reasoning process with more time may outperform a committee of partially informed agents.
Max does not grant new tools, a larger context window, or guaranteed correctness. It changes the inference budget. That normally means greater latency and more reasoning tokens, so the setting should be justified by the expected value of a better result.
Ultra is parallelism, not simply deeper thought
OpenAI describes ultra as its highest-capability setting and says it coordinates multiple agents in parallel. The GPT-5.6 launch post compares a default four-agent ultra setup with a one-agent baseline in several evaluations, and also shows larger configurations for some tests.
The right mental model is an expert team:
- a root agent decomposes the goal;
- focused agents investigate separate workstreams;
- they return findings or artifacts;
- the root reconciles conflicts and produces the final result.
Ultra can help with broad research, exploring different areas of a codebase, comparing several proposals, preparing independent components, or testing multiple hypotheses. It can reduce wall-clock time when those activities genuinely overlap.
It is a poor fit for a single ordered chain, a task where agents would contend over the same file, or work dominated by one slow external operation. Four agents waiting on the same render queue are still waiting on the same render queue.
Max versus ultra at a glance
| Question | Max reasoning | Ultra mode |
|---|---|---|
| Core strategy | More reasoning in one model context | Several agents work in parallel |
| Best task shape | Sequential, tightly coupled | Independent, bounded workstreams |
| Main benefit | Deeper exploration and revision | Faster coverage and specialization |
| Main cost | More reasoning time and tokens | More aggregate tokens and coordination |
| Main failure mode | Overthinking or excessive latency | Duplication, conflict, weak synthesis |
| Shared-state work | Usually simpler | Often problematic |
Neither mode replaces external validation. Max can confidently converge on the wrong assumption. Ultra can produce a polished consensus from several agents that all inherited the same flawed premise.
Availability and naming matter
The API documentation lists max as a reasoning effort for GPT-5.6 Sol, Terra, and Luna. OpenAI’s launch announcement says max can be enabled in ChatGPT Work and Codex wherever the user has GPT-5.6 access.
Ultra has plan-specific product availability: at launch, OpenAI said it was available to Pro and Enterprise users in ChatGPT Work and to Plus and higher plans in Codex. Product eligibility can change, so current plan documentation should be checked before making procurement decisions.
The Responses API Multi-Agent beta exposes model-directed subagents across all GPT-5.6 tiers. Its documented default maximum concurrency is three, which differs from the four-agent ultra configuration described in OpenAI’s launch evaluation. That is another reason not to treat ultra and the API beta as interchangeable labels.
How much does extra compute cost?
OpenAI publishes token prices for Sol, Terra, and Luna, but the relevant cost of max or ultra depends on actual usage. Max may consume more reasoning tokens. Ultra adds contexts, outputs, tool calls, and synthesis across multiple agents. A fixed “ultra costs X times more” claim would be misleading without a specific task and current product billing model.
Measure:
- total input, cached input, reasoning, and output tokens;
- subagent count and duplicated context;
- tool and external-service charges;
- p50 and p95 completion time;
- final acceptance rate;
- human review and correction time.
If max costs twice as much but turns a 70% pass rate into 95% on a high-value task, it may be economical. If ultra saves two minutes on research but triples spend and produces the same decision, it is not.
A decision framework for extra compute
Ask five questions before escalating:
1. Is the task actually difficult?
Do not infer difficulty from prompt length. A long extraction task can be mechanical, while a short instruction may conceal conflicting goals. Run a lower effort first when failure is cheap and detectable.
2. Is the difficulty serial or parallel?
Use max when the solution requires one coherent chain. Use ultra when independent specialists can contribute useful work simultaneously.
3. Can success be evaluated?
Tests, citations, schemas, visual rubrics, and human approval make escalation measurable. Without acceptance criteria, more compute may only produce more persuasive prose.
4. Is an external bottleneck dominant?
Parallel reasoning cannot accelerate a vendor queue, a fixed-duration render, or a human approval. Optimize the critical path instead.
5. What is the failure impact?
High-impact legal, financial, security, identity, or publication actions may warrant stronger reasoning, but they may also require a qualified human. More model compute is not a substitute for authorization.
A tiered strategy beats a permanent toggle
Start routine work at the lowest model and effort level that passes evaluation. Escalate when observable triggers occur: conflicting sources, failed tests, repeated tool errors, an unusually large dependency graph, or a high-risk approval point.
For difficult work, try max before ultra when the problem is tightly coupled. Try multi-agent execution when the work naturally separates. Cache stable instructions, cap agent budgets, and preserve artifacts so an escalation does not repeat expensive external operations.
From extra reasoning to a production media workflow
A complex media package illustrates the difference. Suppose a campaign contains six clips that need trimming, audio treatment, brand placement, and final merging. The editing jobs are external operations, but the plan and review can use different compute strategies.
Max reasoning could help one agent resolve a tightly coupled edit decision: the logo cannot cover captions, the narration must align with a transition, and the final duration must remain under a limit. Ultra could help before that point by assigning separate agents to inspect clip metadata, verify brand rules, check the audio brief, and review usage rights.
After synthesis and approval, Codex or Claude can call a specialist service such as Medux for the actual transformations. Medux is not a native GPT-5.6 component. The Codex MCP video-editing tutorial demonstrates one concrete audio-to-video operation, while the Claude MCP video-merge tutorial shows how ordered clips become a single asynchronous task.
The production rule is simple: never let an escalation duplicate a paid job. Persist upload identifiers, job parameters, task IDs, and completed outputs outside transient agent context. If an agent needs stronger reasoning, give it the existing state and ask it to diagnose or review—not to start over.
Use ultra around independent preparation and QA. Use max for the hardest integrated decision. Use deterministic probes for dimensions, duration, codecs, and file integrity. Keep a human approval before publication. That division captures the value of extra compute without turning every media request into an expensive multi-agent project.