What is the difference between model size and effort level in AI?

Model size picks which set of trained weights answers you, which governs what the system knows and how deep its pattern recognition runs. Effort level is a per-request setting that tells the same model how much work to do: how long to think, how many files to read, how much to verify, and how far to push through a multi-step task before checking in. Model is capability; effort is diligence.

When should I raise the effort level instead of switching to a bigger model?

Fix the context first. If the model had the right information but skipped steps, failed to verify, or bailed early on a multi-step task, that is a diligence failure: raise effort on the same model. If it worked hard, showed its work, and was still confidently wrong or missed the abstraction, that is a capability ceiling: upgrade the model.

Is a more expensive AI model always more expensive to use?

No. Per-token price and per-task cost are different numbers. On work that stretches a smaller model, the larger model can finish in far fewer tokens or avoid a failed run entirely, so its total cost per task comes out lower. At August 2026 list prices, Claude Fable 5 is five times Sonnet 5's output price, so it is cheaper per task whenever it finishes in under a fifth of the tokens, or whenever the smaller model's retries eat that gap.

Which AI providers offer effort or thinking-level controls?

As of August 2026 all the major labs do. Anthropic exposes low, medium, high, xhigh, and max effort across Claude Sonnet 5, Opus 5, and Fable 5. OpenAI's GPT-5.6 family (Luna, Terra, Sol) exposes reasoning effort up to max and a multi-agent ultra mode. Google's Gemini API uses thinking_level (minimal, low, medium, high). Alibaba's Qwen models expose enable_thinking, thinking_budget, and a reasoning_effort level on Qwen3.8-Max.

Essay

Beyond the Model Picker: Mastering the Two Dials of Modern AI

By Doug Liles · August 21, 2026

Open Claude, ChatGPT, or Gemini this month and look closely at the model picker.

It has grown a second control.

Next to the familiar list of model names there is now a slider, a dropdown, or a toggle labeled something like “effort,” “reasoning effort,” or “thinking level.” If you use Claude Code, the /model menu carries an effort slider and the session header tells you which level is live.

For a casual user it reads as clutter. For anyone who runs AI as infrastructure, it is the most important interface change since tool use.

The one-dimensional question, “which model should I use?”, has become a two-dimensional one. And most people are still answering it on one axis.

Capability and diligence are different things

Here is the mechanical picture, stripped down.

When you send a prompt, the text is split into tokens and each token is mapped to an integer. Those integers run through the model’s weights, a long chain of matrix multiplications, and probabilities come out the other end. The weights were fixed during training. Your prompt steers them; it does not change them.

The first dial, model size, chooses which set of frozen weights answers you. Sonnet, Opus, or Fable on Anthropic’s side. Luna, Terra, or Sol on OpenAI’s. A larger model has more parameters, which in practice means more encoded knowledge, deeper pattern recognition, and a higher ceiling on abstraction. That is capability.

The second dial, effort level, tells the model you already picked how hard to work on this request. Anthropic’s Lydia Hallie, on the Claude Code team, describes the mechanism plainly: the effort level is sent to the model alongside your prompt, and the model was trained to behave differently at each level, so that learned behavior is itself baked into the frozen weights. That is diligence.

Effort is not a thinking timer, and it is not a token budget. Anthropic’s API documentation is explicit that it is “a behavioral signal, not a strict token budget.” It governs how long the model thinks, yes, but also how many files it reads, how much it verifies its own work, how many tool calls it makes, and how far it pushes through a multi-step task before coming back to you.

It also cuts the other way. The training teams watch for “overthinking,” where extra reasoning tokens stop adding accuracy and start adding cost and latency. Anthropic’s own docs warn that the max level “may show diminishing returns and is prone to overthinking” and should be tested before you adopt it broadly. Raising effort is not free, and it is not always better.

Figure 1. Model size chooses who answers; effort level chooses how long they work. Three cells from Anthropic's staffing analogy.

Step zero is context. The dials come after.

Before you touch either dial, check what you handed the model.

Hallie’s guidance for Claude Code, which generalizes to every agentic system I operate, is that when the output is wrong your first instinct should not be a knob. Is the prompt vague? Is the model connected to the right tools? Are the files, specs, or statutes it needs actually in the context window?

No amount of capability fixes a missing input. No amount of effort fixes a question you did not ask.

Once the context is sound and the answer is still wrong, there is exactly one diagnostic question: did it not try hard enough, or did it not know enough?

If the model had the right information but skipped steps, failed to run the tests, or bailed on a long task before finishing, that is a failure of diligence. Raise effort on the same model.

If the model clearly worked hard, read everything, ran the checks, showed its reasoning, and was still confidently wrong or missed the abstraction entirely, it hit a capability ceiling. More tokens will not move the answer. You need a bigger brain.

And once the hard task is done, go back to defaults. Hallie’s framing is that effort is “a manual override to scale how hard and long Claude works,” not a permanent setting. The default on every current Claude model is high, with one exception, Opus 4.7, which defaults to xhigh.

Figure 2. The diagnostic. Context first; then one question decides which dial to turn.

Generalist, expert, specialist

The cleanest way to hold both dials in your head at once is the staffing analogy Hallie uses, which I have adopted for my own agent teams.

Sonnet at high effort is a really good generalist with the whole afternoon. No deep prior experience with your specific problem, but diligent enough to read every file and run every test until the task is understood.

Opus at low effort is five minutes with an expert. Deep pattern recognition from training does the work; the answer comes fast and is usually right, but nobody looked closely, so granular details can slip.

Fable, even at low effort, is the specialist glancing at the problem. Anthropic released Claude Fable 5 on June 9, 2026, as a new tier above Opus, priced at $10 per million input tokens and $50 per million output, and built for the most demanding reasoning and long-horizon agentic work. (Its sibling, Mythos 5, shares the same capabilities but is available only through Anthropic’s limited-release Project Glasswing.) Anthropic’s claim is that Fable completes long jobs “that Opus and Sonnet can’t reach at any effort level.” I have not independently benchmarked that, but the pattern matches what I see: the specialist catches the edge case in the first pass that the generalist would have found on the fourth.

Effort sets how long any of them spends on your task. Model sets who showed up.

The productivity paradox: when expensive is cheap

The most common procurement mistake I see, in startups and in government alike, is buying on price per token.

Price per token is the sticker. Cost per task is the bill.

A smaller model grinding near its capability ceiling can burn through dozens of iterations and a great many reasoning tokens, and then fail, which means you paid for every token and got nothing. A larger model can often reach the same solution in a fraction of the steps. Hallie puts it directly: you are paying more per token for the larger model, “but on tasks that genuinely stretch the smaller one, the total cost per task can come out lower.”

The arithmetic is simple enough to carry around. At August 2026 list prices, Claude Sonnet 5 is $2 in and $10 out per million tokens, Opus 5 is $5 and $25, and Fable 5 is $10 and $50. Opus costs 2.5 times Sonnet; Fable costs 5 times. So Opus is cheaper per task whenever it finishes in under 40 percent of Sonnet’s tokens, and Fable whenever it finishes in under 20 percent, or whenever Sonnet’s retries eat that gap.

The effort dial swings the other variable. In Anthropic’s own illustration, the same prompt at high effort generated roughly seven times the tokens of low effort to reach a higher-confidence answer. Seven times the tokens on the small model can cost more than one pass on the large one. Both dials are cost dials. One moves the price; the other moves the volume.

Figure 3. Per-token price is the sticker; per-task cost is the bill. Break-even is list-price arithmetic, not a benchmark.

Those break-even figures are arithmetic on list price, not a benchmark, and the 7x is one illustration, not a law. Measure your own task mix before you route on it. But measure it. The teams I see winning on AI cost are not the ones with the cheapest model; they are the ones who know their cost per completed task by model and by effort level, and route accordingly.

The industry has converged on two dials

This is no longer an Anthropic idiom. Every major lab now ships the second axis, and the naming is settling.

Provider	Model tiers (Aug 2026)	Effort / thinking control
Anthropic	Haiku 4.5 · Sonnet 5 · Opus 5 · Fable 5 (Mythos 5, limited release)	`effort`: low · medium · high (default) · xhigh · max. Claude Code adds `ultracode`, which runs at xhigh and orchestrates multi-agent workflows.
OpenAI	GPT-5.6 Luna (fast, $1/$6) · Terra (balanced, $2.50/$15) · Sol (flagship, $5/$30)	Reasoning effort through `max`; `ultra` coordinates four agents in parallel by default. `max` is a settings toggle in ChatGPT Work and Codex; `ultra` is Pro/Enterprise.
Google	Gemini 3.x families	`thinking_level`: minimal · low · medium · high. Replaced the older token-count `thinking_budget`, which is still accepted for backward compatibility.
Alibaba (open weights)	Qwen3.8 series, Qwen3.8-Max	`enable_thinking` toggle, `thinking_budget` cap, and on Qwen3.8-Max a `reasoning_effort` level (low · medium · xhigh).

Prices are per million input/output tokens at list. Sources are linked at the end.

OpenAI released the GPT-5.6 family on July 9, 2026, with celestial names that map to tiers: Sol is the flagship, Terra is “a balanced model for everyday work,” and Luna is the fastest and cheapest. Google moved Gemini from a raw thinking-token budget to a named level, the same design decision Anthropic made. Alibaba’s Qwen line, one of the strongest open-weight families, now carries all three control styles at once.

Notice a third axis forming at the top of the table. OpenAI’s ultra and Claude Code’s ultracode are not effort levels in the strict sense; they are parallelism. They spend tokens on breadth, spinning up several agents at once, rather than on depth. OpenAI’s unreleased next model, Astra, is reported to be built around that idea, coordinating multiple agents on long-running problems. OpenAI says an internal version of Astra produced solutions to ten open problems in mathematics and theoretical computer science, each formalized as a machine-checkable Lean certificate, for roughly $2,000 in tokens at Sol API rates. The Lean certificates are checkable by anyone and mathematicians have called the results significant, but Astra has no release date, so treat it as a direction, not a product. The direction is clear enough: model, effort, and soon breadth.

From expensive user to savvy operator

The practical posture is not complicated, and it is the same one I apply to agent platforms under the tiered-routing principle: the cheapest model that reliably completes the task, with explicit escalation paths.

Use the defaults for routine work. Every provider tuned them for exactly that, and the default effort on a mid-tier model is the right answer for most of what most people do most of the day.

Route the small stuff down. Summaries, formatting, extraction, first-pass classification: Luna, Haiku, Sonnet at medium. Spend nothing you do not need to.

Escalate on the diagnostic, not on frustration. Context first. Then ask whether it failed to try or failed to know. Raise effort for the first; change the model for the second. Reset when the hard task is done.

Measure cost per completed task, not cost per token, and keep it by model and by effort level. That number is the only one that tells you whether the specialist is actually expensive.

Two years ago the only lever was “pick a better model.” Now you have two, and a third is forming. The question is no longer just what you will ask the AI. It is how you will set the dials.

Will you give the generalist the whole afternoon, or is it time to call in the specialist?

Sources

Primary sources, checked August 21, 2026.

Lydia Hallie, Anthropic, “Choosing a Claude model and effort level in Claude Code”, July 7, 2026. Source of the frozen-weights explanation, the diagnostic question, the generalist/expert/specialist analogy, the 7x token illustration, and the per-token vs. per-task framing.
Anthropic, Effort parameter documentation and Claude Code model configuration: effort levels, defaults, “behavioral signal, not a strict token budget,” overthinking guidance, ultracode.
Anthropic, Introducing Claude Fable 5 and Claude Mythos 5 and pricing: June 9, 2026 release, Project Glasswing, list prices.
OpenAI, GPT-5.6: Frontier intelligence that scales with your ambition, July 9, 2026: Sol/Terra/Luna, max and ultra; pricing via Simon Willison’s write-up.
OpenAI, Ten advances in mathematics and theoretical computer science, August 1, 2026; context and caveats via The Decoder.
Google, Gemini 3 developer guide and Gemini thinking: thinking_level values and thinking_budget legacy status.
Alibaba Cloud, Qwen thinking mode: enable_thinking, thinking_budget, reasoning_effort.

Break-even percentages are my arithmetic on list prices, not a vendor figure.