Wednesday, October 7, 2026
spot_img
Artificial IntelligenceChinaFeatured

Model Reasoning: What “Max Reasoning” actually means, and whether AI reasoning levels can be measured

Jennifer Evans
Jennifer Evans

As generative AI evolves, the separation between the discourse of practitioners and what the public actually experiences widens to the point where the two are almost unrelated. The options available to the individual user become more layered and more complex, while our general ability to understand distinctions, diminishes. PNC’s internal transaction data show that 2.2% of its customer households paid for a generative-AI subscription in May 2026. The public discourse is almost entirely focussed on the future; AGI / superintelligence scenarios, solving of open mathematical problems and future or high level alignment/safety capabilities and incidents that apply to the individual, but in ways they can’t touch and in ways they are not able to fully understand. Whose responsibility is it to monitor and manage this gap between user experience and frontier lab progress? No one, it seems.

The Rapidly Shifting Reasoning Landscape

If we look at what is happening a level down from the discourse, the level that actually affects users, it’s confusing to say the least. ChatGPT Work now offers settings including Light, Medium, High, Extra High, Max and, for some users, Ultra. Claude offers low, medium, high, xhigh and max. Gemini offers Standard thinking, Extended thinking and Deep Think. Grok offers non-reasoning, reasoning and multi-agent modes. Every interface suggests a simple progression: each level thinks harder than the one before it.

That is broadly true until the top of each ladder, where the control stops describing depth and starts describing architecture. It is also much less measurable than the labels imply.

Each vendor can define what these settings are intended to do. None of them publicly define them as fixed quantities of computation, time or intelligence.

The unit underneath the dial

Reasoning models generate internal reasoning tokens before and during a response. These tokens help the model break down a problem, consider alternative approaches, plan actions, use tools, recover from ambiguity and work through multiple steps.

Reasoning tokens are separate from the answer displayed to the user. OpenAI does not expose the model’s raw internal reasoning, although API customers can see how many reasoning tokens a completed request consumed. Anthropic returns a summary of the thinking process while billing for the full internal generation, so the visible token count and the billed token count differ. (The gain is inspectability of the emitted trace, not a verified record of the model’s causal computation.) Google returns thought summaries when asked, and requires thought signatures to preserve reasoning state across turns.

Individual problems can require anywhere from a few hundred to tens of thousands of reasoning tokens. That provides some sense of scale, and it establishes no fixed allowance for Medium, High or Max.

Three variables inside one control

What users experience as “how hard the system worked” involves at least three separate variables.

The model. Sol, Terra and Luna are different GPT-5.6 models with different capability, speed and cost profiles. Haiku, Sonnet, Opus and the Mythos-class models are different Claude models. Flash, Flash-Lite and Pro are different Gemini models. Changing models is not the same as changing reasoning effort.

Reasoning effort. The effort setting controls how much reasoning the selected model applies to the task. Higher levels generally allow more planning, exploration, synthesis and checking. On Claude, the same parameter also governs how many tool calls the model makes, so lowering it shortens the whole loop rather than only the thinking.

Orchestration. A task can be completed by one agent or divided among several. Each subagent performs its own model and tool work before returning findings for synthesis. This variable is what makes the top tiers structurally different from everything below them.

A ladder with eleven operating states, not three

The range runs from constrained single-agent work through expanded single-agent work to multi-agent work. Each of those three groups contains distinct positions that behave differently in practice.

Group one: one agent, constrained cognition

Rung 0, reasoning disabled. The model answers directly with no internal deliberation. Gemini 2.5 allows a thinking budget of zero on Flash variants. Grok’s non-reasoning mode does the same. Gemini 3 removed the option to switch thinking off entirely on its Pro models.

Rung 1, minimal. As few thinking tokens as the architecture permits while preserving reasoning state. Gemini’s MINIMAL level is the default for 3.1 Flash-Lite and still requires thought signatures to be passed back.

Rung 2, low or light. Extraction, classification, routing, simple rewriting, straightforward execution. On Claude, low effort means the model may skip thinking on easy problems altogether and will make fewer tool calls when tools are available.

Group two: one agent, expanded cognition

Rung 3, medium. OpenAI’s recommended starting point and the GPT-5.6 default. Claude and Gemini both treat medium as a deliberate step down from their defaults rather than as the baseline. Suitable for research, planning, judgment, documents and ordinary agentic work.

Rung 4, high. Complex debugging, deeper planning, multi-step research, work where quality matters more than latency. This is the default on Claude Opus 5, Claude Sonnet 5 and Claude Fable 5, and the default thinking level on Gemini 3 Pro.

Rung 5, extra high. Called xhigh on both the OpenAI and Anthropic APIs. Intended for deep research, security review, asynchronous workflows and long-running agentic tasks. Anthropic recommends starting here for coding and agentic use and stepping down only when evaluations show quality holds.

Rung 6, max. The maximum reasoning effort available to the selected model, appropriate for the hardest quality-first tasks requiring more exploration and verification. Max describes the effort assigned to an agent, and it says nothing about how many agents are involved.

Group three: many agents

Rung 7, parallel reasoning inside one response. Gemini’s Deep Think runs parallel reasoning paths and takes minutes to return. Grok 4 Heavy applies parallel test-time compute across a 256,000-token context. The user sees one answer produced by several simultaneous attempts.

Rung 8, user-directed subagents. At most intelligence levels, users can explicitly ask the system to delegate independent work. Grok Build spawns up to eight concurrent subagents, each in its own Git worktree on a separate branch. Anthropic’s documentation advises setting a large output ceiling when running Opus at xhigh or max because the model may act across subagents and tool calls.

Rung 9, system-directed subagents. Ultra’s distinguishing feature is proactive delegation: the system decides for itself that parallel agents would materially improve the task, divides the request into workstreams, spawns specialized agents and synthesizes their findings. Grok 4.20’s multi-agent mode assigns four roles, a coordinator plus research, logic and contrarian analysis, and has them contest each other’s reasoning before one writes the answer. Heavy mode scales the same architecture to sixteen.

Rung 10, restricted frontier tiers. Access here depends on approval rather than on a setting. Anthropic’s Mythos-class models sit above Opus, and Mythos 5 is available only to a small number of approved organizations through Project Glasswing while Fable 5, the same underlying model with additional safeguards, ships publicly.

Where the same request lands on different systems

Two requests that look identical produce different architectures:

“Analyze this issue at Max reasoning.”
“Analyze this issue at Max reasoning using separate subagents for the evidence, counterarguments and fact-checking.”

The second explicitly asks for a multi-agent workflow even though Ultra was not selected. The same distinction holds on Grok, where multi-agent mode can be chosen directly, and partially on Claude, where subagent behaviour rises with effort rather than being selected outright.

What Anthropic ships

Claude’s five effort levels are low, medium, high, xhigh and max. High is the default on Opus 5 and Sonnet 5 through the API and Claude Code. Effort control appears as a dropdown beside the model picker in claude.ai and in Cowork, and it reached all plans with the Opus 4.8 launch.

Two details are key for the measurement question. The first is that Anthropic describes effort as a behavioural signal rather than a strict token budget. At lower levels the model still thinks on sufficiently difficult problems; it simply thinks less than it would at a higher level on the same problem. Adaptive thinking lets the model decide per request whether to think at all, and at high, xhigh and max it almost always does.

The second is that the one genuinely numeric control has been withdrawn. Manual extended thinking with a specified budget_tokens value is deprecated on the 4.6 generation, and models from 4.7 onward reject requests that use it with a 400 error. Adaptive thinking is the only thinking mode available on Fable 5 and Mythos 5. The industry moved from a countable allowance to an uncountable disposition on purpose, because fixed budgets performed worse across mixed workloads.

What replaced it is quantified at a different layer. Task budgets, in beta on Fable 5, Mythos 5 and recent Opus models, give the model an advisory token budget for an entire agentic loop, including thinking, tool calls, tool results and output. The model sees a running countdown and prioritizes accordingly. The number describes the work, not the cognition.

What Google ships

Gemini is no longer a two-position switch. Gemini 2.5 used model-specific numeric thinking budgets, reaching 32,768 tokens on Pro, with negative one signalling dynamic allocation. Gemini 3 replaced that with thinking_level, offering MINIMAL, LOW, MEDIUM and HIGH, with dynamic thinking at HIGH as the default on Pro models and MINIMAL as the default on 3.1 Flash-Lite. Mixing the two parameters in one request returns an error.

In the consumer app the same ladder appears as Standard thinking, Extended thinking and Deep Think. Deep Think is described as maximum parallel reasoning, requires the Pro model, takes minutes to return, and is restricted to AI Ultra subscribers.

What xAI ships

Grok’s divisions are real and they are architectural rather than numeric. The current shape is non-reasoning, reasoning, and multi-agent. Multi-agent mode runs a coordinator alongside specialists for research, mathematics and contrarian analysis, which argue before consensus. Heavy mode runs the same design at sixteen agents and sits behind the $300 monthly tier. Grok Build, the coding CLI, exposes agent count directly through a parallel flag capped at eight.

xAI officially documents four- and sixteen-agent configurations for Grok 4.20 Multi-Agent. It also says the sixteen-agent configuration uses “significantly more” tokens, not exactly four times as much. 

Who is allowed to buy which architecture

The paywall lands in a consistent place across all four vendors, and it lands on parallelism.

OpenAI restricts Ultra to Pro and Enterprise users in ChatGPT Work, and to Plus and higher in Codex, as of July 2026. Google restricts Deep Think to AI Ultra at $99.99 and $199.99 per month. xAI restricts sixteen-agent Heavy to SuperGrok Heavy at $300 per month. Anthropic makes effort control available on every plan while restricting Opus-class models to Pro and above.

Depth controls are comparatively cheap to give away because they are self-limiting. An adaptive system asked a trivial question spends little regardless of the setting, so handing a free user a Max button costs the vendor almost nothing on most prompts. Parallel agents remove that safety valve. They increase compute exposure with the number and length of independent trajectories, although the multiplier is neither fixed nor necessarily linear.

Whether the gate is really about compute

Partly, and the vendors have started saying so directly. Google replaced fixed daily prompt counts in the Gemini app with compute-based usage limits that refresh every five hours against a weekly cap, calculated from prompt complexity, features used and conversation length, with premium models and higher thinking levels drawing down faster. Anthropic pools chat and Claude Code usage against shared limits and moved non-interactive agent usage onto a separate monthly credit in June 2026.

What subscribers purchase is closer to permission to run a shape of workflow than to a quantity of computation. No vendor sells reasoning by the token at the consumer tier, and none publishes what a given level costs them. The price sorts customers by the architecture they are allowed to invoke, and the meter behind it reconciles the difference afterward.

The gate that money does not open

A fourth criterion now operates alongside plan tier and compute. Claude Fable 5 and Claude Mythos 5 were suspended in June 2026 to comply with US Department of Commerce export controls, and access was restored on 1 July after those controls were lifted. Fable 5 returned publicly. Mythos 5, which shares the same underlying model with fewer safeguards in biology, cybersecurity and machine learning research, is being reintroduced only to vetted organizations through Project Glasswing.

That tier is not for sale at any price on the consumer ladder. The criterion is authorization, and it suggests where the top of these ladders is heading: the highest rung stops being a setting and becomes a licence.

What Chinese Models Ship


Alibaba, DeepSeek and Moonshot face the same latency and cost pressures as the American labs and made the opposite choice on legibility. Their reasoning controls are numeric, their reasoning traces are returned in full, and the parallel architecture at the top of the ladder is sold as a product rather than as a plan tier. Qwen kept the number; DeepSeek and Kimi exposed the trace.


Qwen retained the integer. The thinking_budget parameter caps reasoning tokens directly, and when the ceiling is reached the model stops reasoning and answers immediately, defaulting to the model’s maximum chain-of-thought length. Qwen3.6 Plus documents budgets up to 81,920 tokens per request. Thinking itself toggles through enable_thinking, and in hybrid models it can be switched mid-conversation with /think and /no_think suffixes, with the most recent instruction taking effect. Reasoning arrives in a reasoning_content field beside the answer.


DeepSeek published the argument in its own documentation. Thinking mode defaults to enabled, effort defaults to high, and effort escalates automatically to max for complex agent requests coming from harnesses such as Claude Code and OpenCode. Then the compatibility note: low and medium both map to high, and xhigh maps to max. Two of the five rungs on DeepSeek’s ladder are aliases for a rung above them. The five-position scale exists so that Western clients can send familiar parameters, and it sits on top of two real internal states. A vendor building for interoperability treated the industry’s vocabulary as an interface convention, which is what it is.


Moonshot made the trace structural. Kimi K3 exposes a top-level reasoning_effort parameter, and at launch on 16 July 2026 only the max level was available, with thinking permanently enabled. The model was trained in preserved-thinking-history mode, so an agent harness must pass the entire reasoning trace back on every turn; when it fails to do that, or when a session started on another model switches to K3 partway through, output quality destabilizes. Parallelism ships as a second product rather than a higher subscription: K3 Max handles chat and single-track agent work, K3 Swarm Max coordinates fleets of sub-agents across batch, search and repository-scale tasks, and both bill at $3 per million input tokens and $15 per million output. Artificial Analysis placed K3 at 57.1 on version 4.1 of its intelligence index, fourth among tested configurations.


A model trained to receive its own reasoning history cannot treat that history as a secret. The trace is an input. K3’s current interface requires clients to preserve its emitted thinking history. That is a product design choice, not an architectural necessity.


Open weights move the budget into an inference stack the operator runs. Moonshot committed to publishing K3’s full weights under a Modified MIT licence by 27 July 2026, which had not happened as of 26 July. Once weights ship, a thinking cap becomes a number the person serving the model sets, and no vendor is positioned to withhold it.


Selling tokens rather than seats removes the reason to gate architecture behind a plan. A lab with no subscription ladder has nothing to sort customers into, and no revenue reason to make sixteen agents a privilege of the top tier.


Commensurability still does not follow. Counting tokens on Qwen and reading traces on DeepSeek and Kimi improves the legibility of a single run, and it does nothing to make high on one lab’s model comparable to high on another’s, because DeepSeek’s high absorbs two lower labels and Kimi’s max is the only setting on offer. The gain is auditability within a system, not a shared unit across systems.


The American labs defend adaptive allocation on performance grounds, and the argument is a reasonable one: fixed budgets do behave worse across workloads that mix trivial and difficult requests. The Chinese releases establish that the tradeoff is a choice rather than a requirement, since a model can hard-stop at a declared token ceiling and still land in the top five. What the open-weight route surrenders is the ability to keep the highest architecture away from anyone who can afford the hardware to run it.

Why the differences stay hard to see

A straightforward question may produce essentially the same answer at every level because there is little additional reasoning to perform. Adaptive systems are designed to skip work that would not improve the result, so the top setting does not force the model to spend its allowance on an obvious answer.

Differences surface when a task requires the system to reconcile numerous sources, test competing explanations, discover hidden contradictions, plan a long sequence of actions, debug a complicated system, verify an extended chain of assumptions, or divide work among specialists.

Even then, higher effort creates the opportunity for better reasoning and guarantees nothing. A weak prompt, unavailable evidence or a mistaken premise still limits the answer.

What holds and what does not

Definable today:

  • Light, Medium, High, Extra High and Max form an ordered reasoning-effort scale, and the equivalent Claude and Gemini scales are ordered the same way.
  • Each vendor publishes a default: Medium for GPT-5.6, high for Claude Opus 5 and Sonnet 5, HIGH for Gemini 3 Pro.
  • Higher effort generally means more complete reasoning, higher token use and greater latency.
  • Ultra, Deep Think and Grok’s multi-agent modes add parallelism rather than depth.
  • API users can measure reasoning tokens consumed after a request completes.
  • Qwen enforces a caller-specified thinking-token ceiling, and DeepSeek and Kimi return the full reasoning trace rather than a summary of it.

Undefined by any American vendor:

  • A fixed reasoning-token budget for each level.
  • A fixed amount of time spent at each level.
  • A multiplier such as “High thinks twice as much as Medium.”
  • A guaranteed number of planning, checking or verification passes.
  • A fixed number of agents at the multi-agent tiers.
  • An equivalent amount of reasoning across different prompts or model generations. OpenAI states there is no exact mapping between GPT-5.5 and GPT-5.6 levels, and Anthropic notes that lower effort on a newer model often exceeds the highest effort on an older one.
  • Chinese models: A reasoning-token quantity attached to a published level. Qwen accepts a thinking budget from the caller and enforces it as a hard stop, and that number is an instruction the user supplies rather than a disclosure of what Medium or High contains.
  • Plan tier determines which of those architectures a user may invoke on the closed platforms, and open weights remove that gate for anyone with the hardware to serve the model.

What we can interpret from this

Auditing a single run has become easier where the full trace comes back. Comparing two runs across vendors has not.

Reasoning levels are configuration controls, not standardized units of cognition. They tell the system how much computational effort to favour, and they do not tell us how many tokens, seconds, reasoning passes or agents will be involved.

The most accurate interpretation now has four parts:

  • Model choice determines the underlying system.
  • Reasoning effort influences how deeply each agent works.
  • Orchestration determines how many agents contribute.
  • Plan tier determines which of those architectures a user is permitted to invoke at all.

Task complexity then decides how much of the available capacity gets used. That is a useful control system, and it is a pricing structure, and it remains a long way from a scientific measurement of how much an AI thought.

Sources

Featured

Previous article
Best B2B Software Comparison Platforms for Evaluating SaaS Products
Next article
How Designated Representatives Manage Workplace Safety
Jennifer Evans
Jennifer Evanshttps://patternpulse.ai
Principal, patternpulse.ai, and cofounder, Tech Reset Canada. AI policy, research and analysis. Entrepreneur since 2002, marketer since 1998, machine learning since 2009. Based in Toronto and Southeast Asia.

We bring confidence to decision makers.