A new frontier model has shipped almost every week this fall. If the next ones are much smarter, which of today’s habits still pay off? I tested five common techniques on a small, a mid-size and a top model. The prompt tricks faded. The tools and checks didn’t.
Most AI best practice was written for weaker models: long instructions, expert personas, worked examples, “think step by step.” Teams have built libraries of prompts around them, and new hires learn them as craft.
Meanwhile the frontier keeps moving. In September alone, OpenAI, Anthropic and Google each shipped or announced a new top model. If those tricks were compensating for weak models, they are about to become dead weight, and the effort spent on them should go somewhere that lasts.
You can’t test tomorrow’s model today. But you can test today’s small, middle and top models side by side, and see which techniques lose value as intelligence rises. That trend is the best forecast available.
Every build on this site leaned on some technique: calibration, quote checks, decision models, thinking levels. Before the next generation of models arrives, I wanted to know which of those habits are investments and which are crutches.
The ideaUse model size as a time machine. Whatever stops helping the top model today will stop helping every model soon. Put effort into what still helps at the top: tools, verification and good tests.
Sources are listed at the end of the page.
The same 16 problems in every cell, all at reasoning effort “low.” Pick a measure, then a cell to see which problems it got right.
Log scale. Each bar is the cost of one right answer, at published prices (Astra at the flex tier).
What is actually shipped, with prices, and what is only rumored. Benchmarks are the vendors’ own unless noted.
| Model | Released | Price per 1M tokens (in / out) | Notes |
|---|---|---|---|
| Gemini 4 Argon | Announced Sep 30 | $2 / $10 introductory, then $4 / $20 | Leads 13 of 18 benchmarks Google disclosed, such as GraphWalks long context 84.2% against Astra’s 71.8%; trails on some coding and terminal tests. First on Vals Index (68.9%). Access limited to vetted cyber defenders for now. |
| Claude Opus 5.5 | Sep 23 | $4 / $20 | Terminal-Bench 4.0 66.4%, OSWorld 2.0 81.8%. Thinking always on, medium by default. Performs at Fable 5.1’s level on most work. |
| GPT-6 Sol and Luna | Sep 22 | $2 / $10 and $0.10 / $0.50 | The mid and small GPT-6 models used here. |
| GPT-6 Astra | Sep 3 | $10 / $50 | OpenAI’s top model, used here at the half-price flex tier. |
| Claude Fable 5.1 | Sep 1 | $10 / $50 | Anthropic’s most capable generally available model; same model as Mythos 5.1, which has lighter safeguards for vetted organizations. Humanity’s Last Exam 65.0% with tools. |
Anthropic“Fable 5.5” is a community name. As of October 5, Anthropic had announced no Fable 5.5 or 6, date or model ID.
OpenAINothing about a successor to GPT-6 is confirmed. A “Bel” pretraining leak from August is unverified; one analysis estimates a GPT-7 in the second half of 2027 at the earliest plausible pace.
GoogleArgon is announced and benchmarked but not broadly available, so independent tests are still to come.
Sorted by the evidence on this page and the builds before it. “Fades” means its benefit shrank or vanished as models got stronger.
The skill shifts from coaxing a model to specifying, equipping and checking it.
Prompt libraries tuned to one model’s quirks lose value with each release. Knowing what good looks like, owning the tests that prove it, and deciding what a model may do on its own become the core of the job.
Check the practices your team relies on today.
Check a few practices to see where your effort is going.
# The model writes the program; the sandbox runs it; the output is graded instructions: "Do not solve the problem yourself. Write one short, self-contained Python 3 program that computes the answer and prints only the answer." model → ```python … ``` # about 150–270 tokens instead of 400–700 sandbox → python -I solve.py (no network, 20 s limit) → "512" → correct
Sixteen problems, one run per cell, and problems with exact answers, which flatters the code technique: open-ended writing or judgment can’t simply be computed. With the top model at 16 of 16 everywhere, this set can’t show whether prompt tricks hurt the strongest models, only that they no longer help. Small differences of one or two problems are within run-to-run noise (see How Hard Should It Think?). Projecting today’s size trend onto future models is an argument, not a measurement.
The longer argument, about what stays scarce when intelligence is abundant, is in the companion essay: When intelligence is abundant.