OpenAI's GPT-6 Sol and Luna: cheaper on paper, but nerds aren't so sure

OpenAI has quietly rolled out two new models, GPT-6 Sol and GPT-6 Luna, pitching the smaller Luna variant as a much cheaper alternative to its predecessor, GPT-5.6 Luna. The launch has racked up over 1,400 points and 680 comments on Hacker News in half a day - which, on that site, usually means either something genuinely useful has shipped, or someone’s marketing has outpaced reality. Probably a bit of both.

What OpenAI actually announced

The headline claim is straightforward enough to check: GPT-6 Luna is priced at roughly half the rate of GPT-5.6 Luna. That’s a real, stated price cut, not a rumour. Developer Simon Willison, known for benchmarking new models by asking them to draw an SVG of a pelican riding a bicycle, put both Sol and Luna through their usual paces, alongside GPT-6 Astra for comparison. His side-by-side grid of every model at every “effort” setting threw up one properly interesting, verifiable detail: the GPT-6 family defaults to noticeably duller colours than the GPT-5.6 family. Trivial, but a genuine, observable difference rather than a hyped one.

So who is actually getting the “insane value”

Here’s the catch. Commenters piling in with enthusiasm - one calling 6-Luna “insane value” on cost grounds - are largely reasoning from list-price-per-token, using rankings from sites like OpenRouter that show 5.6 Luna already as the most-used model on the platform this month. That tells you about popularity, not efficiency.

A developer running actual coding sessions through the OpenCode tool pushed back hard on the “cheap” framing. Their 30-day usage logs, comparing a $10 subscription covering several open models against what the same work would cost on Luna’s per-token API pricing, worked out at roughly $184 versus $40 - Luna coming out far more expensive in that real workflow, not less. Another commenter countered that this wasn’t a fair fight, since one rival model (DeepSeek V4 Flash) simply burns through more tokens to do the same job, and pointed to Artificial Analysis benchmarks showing Luna coming out cheaper overall.

The original poster wasn’t moved. Distrusting third-party benchmarks, they ran their own head-to-head coding task - explaining a machine learning concept and building a working tic-tac-toe player - on both a rival model and Luna. The rival did it more cheaply on token counts. Their conclusion: the “significantly cheaper” claim for Luna “may not really hold” once you look at what a model actually does with real code, rather than at headline pricing or aggregated leaderboards.

What this means for anyone not running an OpenCode session

If you’re an ordinary reader who occasionally pokes ChatGPT for help drafting an email, none of this changes anything for you today. The pricing dispute matters mainly to developers building products or automated coding pipelines on top of these APIs, where token efficiency - not the sticker price - decides the real bill. For that crowd, the lesson from this thread is an old one: vendor pricing pages and generic benchmark rankings tell you what a model costs in theory; only running your own workload tells you what it costs in practice.

The takeaway

OpenAI’s price cut for GPT-6 Luna is real and confirmed. Whether it translates into cheaper actual usage is, so far, contested by the very people testing it - which is exactly the kind of gap between headline claim and lived experience worth watching before anyone rushes to switch.