Two Answers Four AI Labs Reached in the Same Month: Effort Dials and the Cyber Split

Step back from the individual launches of July 2026 and a different layer comes into view: design philosophy. Two patterns run through the month. First, every one of the four labs shipped effort settings — some way for users to decide how hard the model thinks — a convergence so uniform it looks coordinated, though it was not. Second, on the question of security capability, the same labs split into three openly contradictory answers. One question converged, one split — and both moved work onto your desk. This article maps the two currents across all four vendors.

Effort settings: the dial everyone shipped

Model choice used to be one-dimensional: buy the best tier your budget allows, done. Now choosing the model settles only half the configuration; the other half — how much computation this particular request deserves — is decided by you, every time. Four labs reached that design in the same month because the underlying trade became undeniable: more reasoning time buys quality but costs money and latency, and where to strike that balance depends on the task, which the vendor cannot know. Handing over the dial is the rational endpoint, and all four arrived independently.

Four implementations of one idea

Anthropic went parameter-style: Claude Opus 5 exposes effort settings at five levels — low, medium, high, xhigh, max — defaulting to high, with thinking now on by default and only disablable at high or below. Moonshot did the same with three levels: Kimi K3 takes reasoning_effort at low, high, or max, defaulting to max. OpenAI chose tier-style: GPT-5.6 ships as three separate models — Luna, Terra, Sol — with further depth settings inside the upper tiers, a de facto two-layer dial. Google also went tier-style at the light end, splitting Flash and Flash-Lite.

The styles feel different in practice. A parameter changes without touching the model name — centralize the setting and you can retune everything at once. A tier change is explicit in the code, which aids clarity but makes mixed-depth pipelines your own plumbing. Neither is better; it depends on your existing setup. Since every vendor also names the concept differently, teams running several APIs should define internal labels — deep, normal, shallow — and map each vendor onto them once.

How much the dial actually moves

Two independent measurements give the dial a scale. Artificial Analysis, testing Claude Opus 5 on GDPval-AA v2, found that effort alone swung the score by 407 Elo points, with output token volume spanning roughly 8x from low to max — same model, same questions. And measuring the GPT-5.6 tiers, the firm scored the composite index at Luna 52, Terra 55, Sol 61 (as of early August 2026), with pre-price-cut task costs of 0.21, 0.55, and 1.04 dollars — yet on a coding-agent index the tiers compress to 75, 77, 80. Luna delivers about 94 percent of Sol’s coding result at a fraction of the cost, and the July 30 price cut widened that gap further.

Two conclusions. The range is enormous, and the sweet spot moves by task: raw-reasoning problems reward depth almost linearly, while scaffolded, tool-driven work barely notices it. And more is not reliably better — Artificial Analysis reports that Claude Opus 5 can be more cost-effective between high and xhigh than at max. Parking the dial at maximum is not a safe harbor; details in our Claude Opus 5 analysis.

The responsibility that came with the dial

A dial you control is a dial you can set wrong — and both failure directions are silent. Set it too deep and you burn 8x the tokens with no error message; you find out on the invoice, or never. Set it too shallow and you get plausible answers produced by shallow deliberation, indistinguishable from the outside. The system will not tell you either way. The one-dimensional comfort of just pick the bigger model is gone; per-task judgment is now part of the job.

The practical countermeasures are unglamorous. Pick your three or four most frequent tasks, assign each a depth, and freeze it in a table — fixed rules with judged exceptions beat fresh judgment every time, and a shared table aligns a team. Then review the bill monthly: overdeep settings only ever show up there, and a line-item glance tells you exactly which job to turn down. Setting the dial and auditing the dial are two halves of one habit; the per-model tactics build on the framework in our effort guide.

The question that split three ways

The second current is a fork, not a convergence. As models improve, they get better at security work — finding vulnerabilities, probing weaknesses, assisting defenders — and the same capability invites misuse. July produced three incompatible policies. OpenAI sells it: GPT-5.6 Sol is billed as its strongest cybersecurity model, openly available through the API with defensive uses among its headline applications. Google gates it: Gemini 3.5 Flash Cyber, a dedicated vulnerability-repair model, ships only to governments and trusted partners through the CodeMender pilot. Anthropic refuses it: Claude Fable 5 and Claude Opus 5 trip safety classifiers on such requests — though Opus 5 fires them about 85 percent less often, and a beta called Automatic Fallbacks reroutes flagged requests to a less restricted model instead of erroring.

Sell, gate, refuse — all three inside nineteen days. Which is right is not this article’s question. What matters operationally is that the same request now passes on one model and dies on another. Depth control converged because it is an optimization problem; capability access split because it encodes what each company is trying to protect. Disagreement there is natural — uniformity would be the eerie outcome. The month’s full context is in our July timeline.

Engineering for the refusal case

Classifiers also misfire on work that has nothing to do with security, so production code needs three things. First, know that a refusal is not a transport error: the HTTP response comes back clean with a refusal marked in a status field, so code that unconditionally reads the body breaks on empty or truncated content — branch on status before parsing. Second, decide the fallback before it happens: surface the error to the user, retry on another model, or route to a human. Any of the three can be right; deciding nothing means silent stalls. Third, if you are on Anthropic, the retry-elsewhere path is now a server-side setting rather than custom code. And to be explicit: this is about keeping your pipeline alive after a refusal, not about evading one — the first is engineering, the second is undermining the safety mechanism, and only the first belongs in your codebase. The wider resilience framing is in our dependency-risk piece.

Regulation, the variable beyond performance

One more July pattern: policy moved availability. GPT-5.6 ran two weeks as a trusted-partner preview before its July 9 general release, with TechCrunch reporting a June request from the US administration to restrict the rollout over misuse concerns; the government-only scope of Gemini 3.5 Flash Cyber reads as the same current. The takeaway is not political — it is architectural. Whether a model is available to you is decided by more than benchmarks, so a workflow hard-coupled to one vendor inherits that vendor’s policy exposure.

At the workbench

For small-scale design and 3D printing, the dial mostly touches the bill. Changing one wall-thickness variable in parametric enclosure code does not deserve max — that is 8x tokens for nothing — while re-deriving whether a screw boss clears a board edge does deserve depth. Draw the line by what a mistake costs: a misclassified failure photo is caught later by eye; a wrong dimension surfaces after a print, in wasted filament and hours. Shallow for the first, deep for the second — a rule that survives model generations. Refusals are rarer at the workbench but not unknown: questions about machine safety interlocks or protective circuits in a powered enclosure can trip a classifier, so hobby use just rephrases, but automated pipelines need the fallback path above. And keep prompts and preamble documents vendor-neutral — plain prose, no model-specific formats — because that portability is the cheapest insurance against the availability risks this month demonstrated. The workflow-wide picture is in our AI 3D design roadmap.

Two facts July settled

First: the quantity of intelligence is now a user setting. Four labs, four implementations, one structure — and measured swings of 407 Elo and 8x tokens from the setting alone. Leaving that dial unexamined is not buying performance; it is paying a fee for skipped judgment. Second: which requests pass is now vendor-dependent, permanently. Sell, gate, refuse — three policies shipping simultaneously means a fallback path is part of any serious integration. The common thread: July added less to what the models decide than to what you decide. The sooner your frequent tasks have assigned depths and your pipeline has a refusal plan, the sooner the month’s changes start paying you instead of costing you.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.