TIRSDAG
2026-08-04

Too many projects, too many ideas, too few hours — one learning a day anyway

Cheap tokens only matter if the model shuts up

Luna at $0.20 in / $1.20 out per million tokens is the kind of number that changes what I’m willing to run in a loop. The jobs I’ve been rationing — classify every log line, summarise every changelog, run the boat’s sensor notes through something that actually reads them — stop being a budget question and start being a plumbing question. Terra dropping 20 percent is less dramatic but it’s the tier I’d actually put in front of a Cloudflare Worker.

The part I’d actually check before rewiring anything: heise makes the point that cheap tokens don’t help if the model burns disproportionately many of them. That’s the whole game. Price per million is the sticker; tokens-per-completed-task is the bill. So before I swap a model in any pipeline, I run the same fifty real prompts through old and new and compare total spend, not rate cards.

And it’s worth noting Sol’s price apparently doesn’t move. The floor is getting cheaper; the ceiling isn’t. Plan your fallbacks accordingly.


The story — OpenAI is adjusting prices across its GPT-5.6 line: Luna, the smallest model, moves to $0.20 per million input tokens and $1.20 per million output tokens, an 80 percent change; Terra, the mid-tier model, gets 20 percent cheaper; Sol reportedly stays the same. That puts Luna close to DeepSeek V4 Pro ($0.44 in / $0.87 out) and Terra somewhat under Kimi K3 ($3 in / $15 out) — Chinese models whose low prices have made headlines and which, per reports, more companies are weighing as cost-driven alternatives. Factoring in efficiency, the picture is similar, though the gaps shift. (Source)