GPT-5.6 Max vs Ultra: What Actually Changes?

Max and Ultra can appear next to each other in a model picker, but they are not the same control turned up twice. In OpenAI’s terminology, Max gives the selected model more time to reason about a single task. Ultra uses automatic task delegation so a root agent can send separate parts of the job to subagents and combine their results.

That distinction matters before you spend more time, plan usage, credits, or API tokens. Start with the GPT-5.6 Sol vs Terra vs Luna tier guide, use the Codex profile for the coding surface, and read What Is ChatGPT Work? if the job is a longer research or deliverable workflow rather than software development.

Max and Ultra change different things

Max belongs to the reasoning-effort ladder. For GPT-5.6, max sits above xhigh and allocates more reasoning work to one model invocation. Ultra changes the topology of the run: the system can delegate work, let agents proceed in parallel, and then synthesize their findings. OpenAI’s selector language describes Ultra as maximum reasoning with automatic task delegation, not merely a stronger Max.

The two choices can still share ingredients. An Ultra run can give its agents substantial reasoning, while a Max run can use tools. The decisive question is not which label sounds higher. It is whether the work benefits more from deeper attention along one chain or from several genuinely separable workstreams.

Choose Max for one hard problem

Use Max when the problem is difficult but fundamentally singular. Examples include tracing one subtle failure across a request path, reconciling a specification with one implementation, checking a proof or financial model, resolving a design contradiction, or producing a migration plan whose steps depend tightly on one another. One agent can preserve the full chain of assumptions and revise it as new evidence appears.

Max is a quality-first choice, not an automatic default. OpenAI recommends comparing max with xhigh on representative tasks. More reasoning can uncover edge cases and correct an early approach, but it does not promise linear gains. A long run can still follow a bad premise, overcomplicate a simple answer, or spend extra tokens verifying details that do not affect the decision.

A good Max prompt should therefore define the decision, evidence boundary, and acceptance test. If the task can be solved reliably at Medium, High, or Extra High, keep the lower setting. Escalate to Max when failures at lower effort are caused by insufficient depth rather than missing data, unclear instructions, unavailable tools, or a dependency outside the model’s control.

Choose Ultra for parallel workstreams

Use Ultra when the job has independent branches that can make useful progress at the same time. A repository audit can assign different subsystems to separate agents. A research project can split official documentation, market evidence, and risk analysis. A larger implementation can separate bounded components, tests, and review, provided ownership is clear and concurrent writes will not collide.

The potential speed advantage is wall-clock speed: several branches can run together instead of waiting in a single queue. It is not evidence that each subagent reasons more deeply than a Max agent. If one branch depends on the result of the previous branch, delegation adds coordination without creating useful parallelism. A single slow external service can also remain the bottleneck no matter how many agents are available.

Ultra introduces failure modes that Max does not. Agents may duplicate the same investigation, interpret the boundary differently, return conflicting recommendations, or inherit the same faulty assumption from the root prompt. The root agent then has to judge and compress several outputs; important caveats can be lost during synthesis. Parallel code changes are especially risky when agents touch shared files, schemas, or mutable state.

Hands-on verification required: Official documentation establishes the execution model and published eligibility, but it does not predict the delegation trace, elapsed time, usage depletion, or result quality for a particular account and workload. Test a representative run before treating Ultra as the production default.

When normal effort or one agent is better

OpenAI’s guidance is that most tasks do not need Max or Ultra. Medium is the general balance of speed and intelligence, while lower effort suits quick, well-scoped work. High or Extra High often provides enough depth for difficult tasks without the full latency of Max. A well-instructed single agent also avoids the handoff and synthesis costs of orchestration.

Prefer normal effort or one agent when:

• the answer is short, routine, or easy to verify;
• each step depends on the exact output of the previous step;
• several workers would compete over the same files or mutable state;
• the job is blocked on one slow tool, approval, database, or external service;
• you need one coherent voice, one audit trail, or deterministic control over a fixed workflow; or
• the prompt is still ambiguous enough that parallel agents would multiply the ambiguity.

The cheapest correction is often better task design: narrow the scope, supply the missing source, define the output contract, or split the work manually at a known boundary. Increase reasoning or add agents only after the simpler run exposes a real depth or throughput problem.

Speed, usage, cost, and failure modes

Max normally takes longer than lower reasoning efforts because one agent is allowed to explore and verify more. In the API, reasoning tokens are billed as output tokens, and a high reasoning budget can consume the output allowance before a complete visible answer is produced. In ChatGPT Work and Codex, the same deeper run consumes plan usage or credits according to the product’s usage system rather than an API invoice.

Ultra can reduce elapsed time when branches are parallelizable, but it normally increases aggregate consumption because the root and subagents all do work. ChatGPT Work and Codex share the same plan-usage and credit structure, and consumption varies with model, context, reasoning, tools, retrieval, and caching. OpenAI does not publish a universal Max or Ultra multiplier, so do not promise a fixed cost or a linear quality return.

API billing is a separate purchase path. Published GPT-5.6 rates are per one million tokens: Sol is $5 input and $30 output, Terra is $2.50 input and $15 output, and Luna is $1 input and $6 output. Multi-agent usage includes tokens generated across the root and its subagents; reasoning tokens count as output. ChatGPT or Codex plan allowances and credits do not pay an API token bill.

Availability depends on the surface

ChatGPT Work. Max is documented for users who have GPT-5.6 access. Ultra is documented for ChatGPT Pro and Enterprise. Work is rolling out across ChatGPT surfaces, and the visible model or mode can still depend on platform, workspace settings, administrator controls, and rollout state. Its usage follows the same structure as Codex, but Work is the surface for longer research and finished deliverables rather than the dedicated software-engineering workspace.

Codex. Max is likewise documented for users with GPT-5.6 access. Ultra is documented for Plus and higher plans. This entitlement is specific to Codex; it should not be copied onto ChatGPT Work without checking Work’s separate plan matrix. Codex subagents are most useful for read-heavy exploration and clearly partitioned work, while concurrent write-heavy tasks need stronger ownership boundaries.

Responses API. The API reasoning-effort value is max; there is no ultra value for reasoning.effort. The related orchestration feature is named Multi-agent [beta] and is available for GPT-5.6 models through the Responses API when the organization has the required model and beta access. It is enabled separately, its schema can change during beta, and its documented default allows three concurrent subagents—distinct from the four-agent default described for product Ultra.

Do not transfer an entitlement, rollout state, or billing rule across those rows. A plan that exposes Codex Ultra does not by itself establish ChatGPT Work Ultra. Access to either product mode does not grant API Multi-agent beta, and an API key does not turn ChatGPT plan credits into token balance. Check the exact surface, account, workspace, model, and billing route.

A practical decision rule

Begin with the lowest effort that reliably passes the task’s acceptance test. Move upward only for an observed reason:

1. Use Medium or lower for routine, bounded, easily checked work.
2. Use High or Extra High when one agent needs more care but the full Max latency is unnecessary.
3. Use Max when one hard problem benefits from deeper exploration, verification, and revision in a single context.
4. Use Ultra when the job contains multiple independent workstreams whose parallel progress is worth the added usage and synthesis risk.
5. Use API Multi-agent only when you intentionally want beta orchestration and have measured its separate token bill.

For an important recurring workflow, run a small matched evaluation rather than choosing by label. Compare correctness, completeness, elapsed time, total usage, edit conflicts, and review burden. The best setting is the least expensive execution shape that repeatedly clears your quality bar—not the option with the most impressive name.