Claude Opus 5.5, GPT 6 Sol, GPT 6 Luna, and a new price war
Jessica and Cathy dig into Simon Willison’s read on Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna as a pricing war with real product consequences. They focus on the practical shift: cheaper default models, cache economics, and whether Opus 5.5’s communication improvements matter more than its benchmark aura. They also poke at the weird failure mode where “max” thinking literally breaks on a silly SVG prompt.
Transcript
Jessica Okay, the weird part is that this is basically a price war disguised as a model launch. GPT-6 Sol and Luna got cut in half, Opus 5.5 got cheaper too, and suddenly the boring question is which one you leave on by default.
Cathy Right, and that’s the actual story. Not the pelicans, not the naming chaos, just the fact that the middle tier got squeezed hard enough that older pricing starts looking kind of silly.
Jessica Exactly. Simon’s read is useful because he doesn’t pretend the leaderboard is the point. He’s looking at what people will actually ship against, and that’s where the price table starts doing the work.
Cathy Mm-hm.
Jessica I mean, look at Luna. Half of GPT-5.6 Luna again is not a cosmetic change. If you’re building something that does a lot of routine calls, that’s the difference between a toy budget and a real one.
Cathy And the cache pricing matters even more than the headline token price for long agent loops. If most of your input is cached, the difference between twenty cents and ten cents or two cents and one cent stops being abstract very fast.
Jessica Yeah, and that’s where I think this becomes a product story instead of a model-story. The model that feels cheap enough to keep on all day is the one people will remember, not the one that won a benchmark by three points.
Cathy Sure, but I’d keep one eyebrow up. Benchmarks still tell you something about the ceiling, and Simon’s comparison grid is really about feel as much as raw scores. He’s basically saying, these families have a different shape now.
Jessica Okay, but the shape does matter. He’s saying the 5.6 family was bolder and brighter in those pelican renders, while the 6 family got muted. That sounds goofy, but it’s his way of showing the vibe changed with the cost curve.
Cathy No, that part is genuinely useful. People laugh at the pelican test, but it catches a bunch of little behavior shifts that don’t show up in a single score. It’s a dumb prompt, which is why it’s kind of good.
Jessica And then there’s Opus 5.5, which feels like Anthropic heard the complaints and went, fine, we’ll make it cheaper and clearer. That’s such an Exploring Next take, by the way, but it’s also probably right.
Cathy Yeah, that line from Thariq is doing a lot of work there. Cheaper per token, clearer communication style, same intelligence class as Fable 5.1, better at Blender. That’s not just a cosmetic refresh.
Jessica What I like is that it sounds aimed at people who actually use the thing. Not just, ‘here is the next frontier crown.’ More like, okay, your agent should stop being so annoying to talk to.
Cathy Mm.
Jessica But the funniest bit is the max-thinking failure. A pelican on a bicycle prompt should not be a stress test, and yet Opus 5.5 max just kept reasoning until it hit the output ceiling and died.
Cathy That’s the part I care about technically. If the highest effort setting can chew through the entire 128,000 token budget on a trivial SVG, then max is not a universal upgrade. It’s a mode with sharp edges.
Jessica And Simon’s reaction makes sense to me. He’s basically saying, if it breaks on a stupid prompt, why would I trust it on a real one? That’s not benchmark theater, that’s just a nasty practical signal.
Cathy I’d be a little careful there, though. One bad demo doesn’t mean the mode is useless everywhere. But it absolutely means you need to know when higher effort is buying quality and when it’s just buying delay and failure.
Jessica Right, and that’s the real product implication. The winning model is not the one with the fanciest reasoning mode on paper. It’s the one where the effort dial actually maps to something a developer can predict.
Cathy Exactly. That’s why the pricing and the default-effort behavior matter together. If medium works well and max sometimes self-destructs, the sensible setup is obvious, at least for a lot of workloads.
Jessica Also, I love that he says he’s using GPT-6 Sol and Claude Opus 5.5 as his default models now. That’s the line I’d remember if someone asked me what changed. Not the pelicans. The defaults.
Cathy Yeah, and he upgraded the Datasette Agent demo to GPT-6 Luna because it’s fast and competent at SQL and building HTML and JavaScript. That’s exactly the kind of boring win that actually ships.
Jessica So for me the takeaway is: the battle moved down the stack. It’s not just who can impress you once, it’s who can be cheap, decent, and stable enough to sit in the loop all day.
Cathy Mm-hm. And the annoying part is that he’s probably right, which is rude, because now we have to care about pricing tables again.
Jessica I cannot believe we are spending another Wednesday-ish conversation on token tables, Cathy. But okay, the pelican grid did make me laugh, so I’ll allow it.
Cathy Stop it.
Jessica No, seriously, if the next episode is just us arguing over cache reads, I’m blaming you. Anyway, I’m going to steal your exact face when I see a model overthink a bicycle.
Cathy Fair. And for once, your product optimism is not completely annoying.
Jessica Wow. I’m framing that. Okay, come back next time with fewer pelicans, more actual defaults.