Devlog 006: After a long grind, the model upgrade is finally live

Last time (Devlog 005), I wrote that switching AI models feels like onboarding a new fortune-telling apprentice — smart, but you have to teach them the rules all over again. This one is the follow-up: the rules finally stuck.

It was more of a slog than I expected. I thought picking a new model would be a one-line config change. Instead I fell down a rabbit hole and burned through several late nights.

A small promise

We've always had a quiet pact with you: paid reports always run on the strongest AI model available at the time — not the cheapest, not the merely-good-enough, but the very best one out there right now. So every time a new flagship ships, we evaluate it seriously and put it to work. This round, that means two of today's top contenders.

Who's on the roster now

• Claude Opus 4.8 — the new main model behind paid reports. What I love most is that it actually stays anchored to the chart: every palace is explained with the real stars, brightness, and transformations, instead of dissolving "Tian Tong and Tai Yin in fall" into a pretty but empty line. The writing is nuanced too — it reads like someone who genuinely gets you, not a formula reciting itself.
• GPT-5.6 — another top-tier new generation, wired in as an option to give report generation an extra layer of resilience.

Why it took so long

Honestly, a few invisible traps.

Some models are smart, but their output is like a coin flip — same chart, one run comes out sharp and deep, the next goes strangely blank with all the chart detail gone. Quality can't ride on luck; a report you paid for can't be hit-or-miss.

Some models have hidden parameter rules — one setting out of place and the whole request bounces back, so it quietly falls through to a backup model. You can't catch a culprit like that without writing experiment after experiment, turning off one variable at a time.

So most of this stretch went into making the models fight it out on the same set of test charts — comparing paragraph by paragraph who's accurate, who's making things up, who's stable, who only looks good once in a while. Only the winner gets to play.

What it means for you

• Interpretation is sharper and better-written at the same time — palaces connect into one coherent story instead of each talking past the others.
• After a report is generated, there's still a layer that checks whether it claims anything the chart doesn't actually contain. Caught → rewrite; still failing → we refund your credits. We won't hand you a report built on guesswork.

If something still goes wrong

No amount of safeguards makes it perfect. If you ever hit a report that keeps failing to generate and stays stuck — and your credits weren't refunded automatically — please don't just eat the loss. Email [email protected] with roughly what happened (an order number or report link helps), and I'll check it and refund you as soon as I can.

Thanks for trusting us, and for waiting while we got this right. More soon.