You did not pick Claude. You picked a demo.

If your model choice cannot survive ten minutes against the strongest case for the other two, you have not chosen. You have preferred.

The hook

A staff engineer walked into a design review last month with a slide that said "we are using Claude for the agent tier because it handles tool use better." Everyone nodded. The slide moved on. In the next one-on-one, I asked what happened when they tested it against OpenAI and Google's current frontier. Three answers, in decreasing order of frequency: we did not; we ran one demo prompt; we have an eval, but we wrote the tasks after we picked Claude. I have made this mistake myself, more than once. It always feels like a decision at the time.

गरुडाचें वाहकें कासे पीतांबर · He Who Rides Garuda, Robed in Yellow Silk

गरुडाचें वाहकें कासे पीतांबर । सांवळें मनोहर कैं देखेन ॥ १ ॥ ॥ ध्रु. ॥
बरवया बरवंटा घनमेघ सांवळा । वैजयंतीमाळा गळां शोभे ॥
मुगुट माथां कोटि सूर्यांचा झळाळ । कौस्तुभ निर्मळ शोभे कंठीं ॥ २ ॥
ओतिव श्रीमुख सुखाचें सकळ । वामांगीं वेल्हाळ रखुमादेवी ॥ ३ ॥
उद्धव अक्रूर उभे दोहींकडे । वर्णिती पवाडे सनकादिक ॥ ४ ॥
तुका म्हणे नव्हे आणिकांसारिखा । तोचि माझा सखा पांडुरंग ॥ ५ ॥
The right to say "this one is not like the others" is earned by having honestly surveyed the others first. Before the survey, what looks like judgement is taste.

What I keep seeing

None of those three answers are decisions. They are preferences retrofitted with the vocabulary of decisions. The staff engineer was not lazy. They were busy, and the demo was convincing, and nobody in the room asked the hard question because everybody in the room had the same vague sense that Claude was the right call. Consensus is not evidence. It is just consensus.

I see this pattern at least once a quarter. A team locks into a model provider based on one impressive demo, one benchmark blog post, or one conference talk. The lock-in hardens over months. By the time someone asks "did we actually evaluate this," the answer is baked into the architecture and the switching cost is real.

The mechanics

The current frontier from the three major labs is roughly comparable on most public benchmarks and differs sharply on specific tasks. The strongest Anthropic model leads SWE-bench Pro. The frontier OpenAI model is stronger on some multi-step reasoning tasks and weaker on long-context retrieval. Google's current Gemini is stronger on multimodal input and cheaper per token at scale. None of these facts, alone, tell you which one to build on.

A real choice looks like this. You write ten to twenty task descriptions that describe your product, not a generic capability. For each task, you write a grader: a small program that decides whether an output is acceptable. You run each candidate model on each task, twenty to fifty times, at your production temperature. You look at pass rate, tail latency, cost per successful task, and failure modes. The output is a spreadsheet with numbers. The choice is the row with the best numbers, weighted by what your product actually needs.

If you do not have that spreadsheet, you did not choose. A vendor demo chose for you.

Where Tuka comes in

The abhanga does something specific in its structure. Tuka spends four verses surveying every element of the tableau: Garuda beneath, the rain-cloud complexion, the crown blazing like ten million suns, the Kaustubh at the throat, Rukmini on the left, Uddhava and Akrura on either side, the Sanakadikas singing. Only after the full survey does the closing line land: "नव्हे आणिकांसारिखा" (not like any of the others). The claim is small because the survey was large. Read the line without the survey and it is a boast. Read it after the survey and it is discrimination.

Model selection has this exact shape. The right to say "Claude is not like the others" is earned by having honestly written the graders for OpenAI and Google too, and looked at the results without flinching. Before that survey, what you have is a vendor preference dressed as an engineering conclusion.

What I would actually do

For every model-tier decision, write the graders before you write the argument. Twenty tasks, twenty graders. Run the current frontier from each of the three labs. Publish the spreadsheet inside the team. Redo it every model release. When you announce a choice, point at a row. If a senior person cannot defend the choice by pointing at a row, they have not chosen. Send them back. This will feel slow. The alternative is paying for a vendor demo for eighteen months and calling it architecture.

Chetan Dhandal