Maximum Effort Loses

If You Only Read One Thing

AI pricing now has two expiration dates: the point where extra reasoning stops improving the answer, and the date a provider's discount disappears. The Effort Dial Lies uses new same-weight runs to show compute is not a quality ladder; The Discount Expires shows why OpenAI's temporary Sol cut makes cost a dated contract, not a model property.

The Effort Dial Lies

Maximum reasoning is not the safest setting. It is a model-specific configuration that can spend more, take longer and produce a worse result than the setting below it.

The evidence changed on August 21. Artificial Analysis added lower-effort runs for Qwen3.8-27B and Grok 4.6 to its public evaluation changelog. The weights stayed fixed. Only the requested reasoning effort changed.

Qwen's low setting scored 42.87 on the Intelligence Index, against 52.02 at xhigh. Low preserved 82.4% of the score while using 26.2% of the output tokens and 36% of the evaluation cost. Artificial Analysis reports 43 million output tokens and a $211.65 suite cost for low effort.

This is configuration-specific capability. Think of reasoning effort as a budget for internal search, not a larger brain. The model weights stay fixed while the inference service gives the model more room to explore, reconsider and sometimes abandon an earlier path. More search can find a correction. It can also create more places to get lost.

The token split makes that mechanism visible. Reasoning consumed 54.1% of Qwen's output at low effort and 93% at xhigh. Answer tokens fell from 19.6 million to 11.5 million across the suite while internal reasoning rose from 23.1 million to 151.8 million. The model talked to itself more and wrote less.

At Alibaba Cloud's listed rates, the difference compounds. A 10,000-task daily workload using the observed average output would cost about $769 a day at low effort and $1,415 at xhigh. The annual difference is roughly $236,000 for those nine additional index points. That is not a universal workload estimate, but it gives the dial a financial scale.

Grok supplies the sharper contradiction. SpaceXAI documents high as Grok 4.6's default and xhigh as its maximum depth for tasks where quality matters more than response time. Yet Artificial Analysis scored high at 60.92 and xhigh at 60.01. The xhigh run used 2.6 million more output tokens and cost $192 more across the suite.

The difference is small enough that noise remains a serious explanation. The result matters because the agentic and knowledge subscores moved in the same direction, and because this is not an isolated pattern.

An independent 94-item on-prem evaluation found that one open-weight model scored 93.4 at medium effort and 84.6 at its high default. High consumed 81% more completion tokens. Disabling reasoning scored 78.5, so the finding was not that thinking is useless. The useful amount had a peak.

The strongest counterargument is workload specificity. A composite index mixes coding, tool use, long context and difficult knowledge tasks. Maximum effort can still win on a narrow class of hard problems, and Qwen's nine-point gain from low to xhigh is real even if it is expensive. One run per configuration cannot establish the best setting for every application.

That concession changes the conclusion, but it does not restore the ladder. Last Sunday's local Qwen test showed a default reasoning setting turning one task into a 21-minute wait. Tuesday's Grok analysis showed that fewer turns and less accumulated input can matter more than a low token price. The new data completes the argument: a model name is no longer a sufficient deployment unit. The unit is the model, effort setting, harness and task distribution together.

This shifts value toward systems that can estimate difficulty, choose an effort budget and verify the result. The model provider sells possible search. The runtime decides how much search a task receives. The evaluator determines whether those extra tokens bought an accepted answer.

The concrete watch is observability. By September 30, two major coding-agent products exposing the chosen per-turn effort in an exportable trace would make this configuration testable in production. If effort stays hidden behind labels such as “auto” or “maximum,” the largest variable in reasoning cost will remain invisible at the workflow level.

The Discount Expires

OpenAI did not launch a new GPT-5.6 Sol on August 21. It changed the economics of the existing model for exactly three months.

The updated GPT-5.6 release page says OpenAI cut Sol's API and credit pricing by more than 20%. OpenAI announced no accompanying model or capability update; the existing model name and published effort menu remained in place. The quoted service became cheaper because the rate schedule moved.

That makes the rate card part of the model version. A benchmark's cost per task multiplies observed token use by the price posted on the day of the run. When the scheduled window reaches November 21, a router using August costs can change its model choice even if every model produces identical answers.

A preregistered paired study of Claude Sonnet 5 shows the other half of this contract. Explicit high effort added $0.01031 per call on average. Researchers did not detect an accuracy gain, although their interval could not rule out an improvement of up to 4.67 percentage points. The study's point was deliberately narrow: the purchased unit includes the served model, requested effort, output rail and price schedule, not just a name.

This is a dated inference contract. Think of it as an airline fare attached to a specific flight, cabin and purchase date. “GPT-5.6 Sol” identifies the aircraft. It does not tell you which effort setting ran, which service tier delivered it or which price was in force.

The strongest counterargument is transparency. OpenAI stated the duration, and customers receive a real saving while it lasts. That is true. The risk sits one layer downstream, where gateways and leaderboards often preserve a cost estimate longer than the price that produced it.

The payoff is a stricter deployment record. An evaluator or router needs the effective price date and served effort beside the model ID before it can reproduce a cost claim. Model providers can change the schedule; the runtime decides whether that change silently rewrites model selection.

The clean test arrives on November 21. If OpenAI restores the prior rate, any public comparison still showing August per-task costs without an effective-date warning will be stale. If the cut becomes the base price, the promotion was price discovery rather than a temporary ranking boost.

The Contrarian Take

Everyone says: Maximum effort is the conservative quality setting, while a lower API price simply makes the same model better value.

Here's why that's wrong (or at least incomplete): Extra reasoning expands the search space; it does not order answers by quality. Fresh Grok results put xhigh below high, while Qwen's low setting keeps most of the aggregate score at barely one-quarter of the output. OpenAI's temporary Sol cut changes measured cost without changing model behavior. “Maximum” describes a budget ceiling, and “price” describes a dated schedule; neither is a durable model property.

Under the Radar

Green is not done — Aident recorded 3,277 successful executions among 4,187 production tool calls, a 78.3% terminal success rate. The telemetry cannot say whether the agent chose the correct tool, supplied the right arguments or completed the user's task. A green executor is one layer below a successful agent. (Cohort)
Stable can still mean source-only — llama.cpp's first non-prerelease semantic version gives downstream packagers a real library version, but v0.2.0 ships without binaries. The release label stabilizes dependency semantics before it stabilizes distribution. (Release)

Quick Takes

SGLang shipped the release that version scanners were looking for. Version 0.5.18 is stable and installable, with 710 pull requests from 212 contributors. Its opt-in checkpoint staging cuts one Qwen3-32B H100 startup from 84.8 to 35.6 seconds, while a 47% language-model-head kernel gain improves full token time only 3.5%. The small end-to-end number is the useful one. (Source)
One local model's high default failed at tool selection. PureTensor's 94-item evaluation scored tool selection at 62.5 under high effort versus 87.5 with no extra preamble at medium. The result is one model on one private task set, but it shows where overthinking can hurt agent work: extra reasoning changes which action gets called, not just the prose around it. (Source)
Grok's default is already a policy choice. Grok 4.6 defaults to high, not xhigh, even though SpaceXAI describes xhigh as the setting for the hardest problems. The provider has separated the ordinary quality default from the maximum compute setting; the API label makes that allocation visible. (Source)

The Thread

Model labels are becoming lossy summaries of a longer contract. “Maximum” names an allocation of compute, not guaranteed quality. A promotional price names a three-month schedule, not enduring cost. “Successful” names an executor state, not a completed task. “Stable” names a version boundary, not a binary.

That changes where operational truth lives. Providers sell a model name and possible search. Runtimes and evaluators hold the state that makes the name meaningful: effort, price date, harness, artifact and acceptance test. The more configurable the stack becomes, the more value moves toward the systems that can preserve that state and prove which configuration produced accepted work.

Predictions

New predictions:

I predict: By September 15, Artificial Analysis will publish a same-model sweep across at least three effort settings for another top-20 model, and at least one aggregate score will fail to rise monotonically with effort. (Confidence: medium; Check by: 2026-09-15)

Issue date: August 23, 2026 · Generated: 04:02 AM ET