Three tabs, one deadline
You have Laguna S-2.1, Claude Opus 5, and GPT-5.6 Luna Pro open in three tabs, and a migration ticket due today. Poolside just priced its first frontier-tier model, and the instinct is to assume new entrant means cheaper entrant. That instinct is wrong often enough that it’s worth fifteen minutes to check before you rewrite a routing config.
New models don’t launch cheap because they need share. They launch at whatever rate signals “we belong in this tier.” A lab with one flagship product and no ad-supported consumer app to subsidize compute has less reason to race to the bottom than an incumbent already amortizing training cost across a much larger user base. So check the actual listed input and output rates on all three pages before assuming Laguna S-2.1 undercuts anything.
What the headline rate hides
The input rate is the number everyone screenshots. It’s also the number that matters least for most real workloads, because output tokens are priced at a multiple of input on every closed frontier model tracked on the site, usually somewhere in the 4-6x range. If Laguna S-2.1 posts a lower input rate than Opus 5 but a comparable or higher output rate, and your workload is generation-heavy (long completions, chain-of-thought, code that gets written rather than just read), the input discount gets erased fast.
This is the trap that catches teams migrating on the strength of a single quoted number. A summarization or code-completion pipeline moves to the model with the lower sticker price, and the bill goes up, because the new model’s average completion runs longer per request and the output rate carries none of the input-side cache discount the old model had. Nobody notices until the first invoice, because nobody pulled the actual input-to-output token ratio from their own traffic before switching. Pull yours before you touch the routing config. It’s the same ratio that determines whether a cache-heavy agent workload behaves nothing like the benchmark prompt the vendor used to set the rate.
None of the three model pages here disclose active parameter counts, so there’s no memory-based math to lean on the way there is for MoE models like GPT-OSS 120B or Qwen3 235B A22B, where total and active params tell you almost everything about both hosting cost and self-host feasibility. Closed frontier models don’t give you that lever. You’re comparing quoted rates and measured latency, full stop, and the vendor controls both numbers.
The single-source problem
Open-weight models on this site routinely show enormous spreads between the cheapest and priciest provider for the exact same weights. As of this writing, Llama 3.1 405B and Qwen3 235B A22B both show spreads over 20x depending on which provider is quoting, because a dozen providers compete to host the same open checkpoint and some of them are running clearance-sale hardware. That spread is arbitrage you can act on today.
Laguna S-2.1, Claude Opus 5, and GPT-5.6 Luna Pro don’t have that option. They’re single-sourced through their labs’ own APIs (or a small number of licensed resellers), so there’s no competitive spread to shop. The rate you see on the model page is close to the only rate you’ll get. That changes the decision entirely. With an open model, a bad quote from one provider just means checking two more on the providers page. With a closed model, a bad rate means the model itself is expensive for your workload, not that you’re shopping wrong.
Run your ratio, not the headline
What this comparison can’t tell you is Poolside’s actual cache-read discount policy, or whether Laguna S-2.1 even offers one. That detail, if it exists, lives in the fine print on the model’s own page, and a single headline rate won’t capture it. Check it directly before modeling anything off input price alone.
Pull your last week of production traffic on whichever incumbent you’re currently running, Opus 5 or Luna Pro, and compute your actual input-to-output ratio and your average tokens per completion. Run that exact mix through the calculator against all three models’ live rates. If Laguna S-2.1 comes out ahead on your real ratio and not just on the headline input number, migrate a small slice of traffic and measure latency and output length before moving everything. If it only wins on the number you’d screenshot for a slide, stay put. A new tab in your pricing spreadsheet isn’t a reason to touch a working pipeline.