GPT-6 lands. The real story is Luna at half the price.

What happened

OpenAI released GPT-6 in two named variants: Sol, the heavier reasoning tier, and Luna, the fast/cheap tier meant for high-throughput and agent orchestration. The launch post led with capability charts, but the developer community's attention landed somewhere else within an hour. Simon Willison's first reaction on Hacker News cut through the marketing: "GPT-6 Luna being half the price of GPT-5.6 Luna is a really big deal."

That's the headline. Not a new modality, not a fresh reasoning trick — a price cut on the model most people were already using for the boring 80% of their pipeline. Willison ran his usual pelican-on-a-bicycle SVG test on both new variants and posted the outputs; the Sol drawing is noticeably more coherent than 5.6's, but the more interesting artifact is that Luna's pelican is now good enough that most teams won't reach for Sol at all.

On the commercial side, the developer benchmark that actually matters right now is the head-to-head against Anthropic. One HN commenter posted the current Claude Opus pricing side-by-side: cache reads $0.20–$0.50 per million tokens, input $4–$5, output $20–$25, cache writes $5–$6.25. GPT-6 Luna sits well under that, and Sol is competitive with Opus 5.5 on the tasks that need it. "I don't see how anyone can be using Claude with prices like this," one commenter wrote — hyperbolic, but the sentiment is now the median take on the thread rather than the fringe one.

Why it matters

Model pricing has been the least interesting part of the frontier-lab conversation for two years. Everyone assumed input/output token costs would drift down at roughly the same rate across labs, and roughly they did. What OpenAI just did with Luna is different: they halved the price of the tier developers actually keep in their hot path, not the flagship they use for demos.

That matters because agent workloads are structurally cost-sensitive in a way single-shot chat is not. If your agent runs a 30-step loop with tool calls, retries, and reflection passes, every step is a full round-trip. Cutting per-token cost in half doesn't cut your bill in half — it cuts your bill in half *and* makes previously uneconomic architectures (longer reasoning traces, more parallel candidates, cheaper self-consistency) suddenly worth trying. Anthropic's response window here is narrow.

The pricing story also compounds with a subscription story. Commenter jeffnash's read on the two 20x plans is worth quoting: "The deciding factors for me between Claude Code 20x and Codex Pro 20x are usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile." Once the fixed-price plans start diverging on effective throughput, you get a switching event that isn't really about model quality — it's about whether you keep hitting the wall by 3pm.

The counter-current is worth taking seriously, though, because it's the same signal that made 5.6 Sol sticky in the first place. Commenter m_fayer captured it: "5.6 Sol was some sort of sweet spot for me. Something about how it communicated verbally and its engineering instincts just clicked for me, and I was able to somehow predict it and jam with it. Like a colleague you click with." That is not a benchmark you can chart, but it is a real switching cost. Models are starting to accumulate the same 'colleague fit' loyalty that IDEs, keyboard layouts, and shell prompts do — and once a working engineer finds a model they can predict, they will pay a premium to keep it.

Which is why the new Sol is the more interesting release even though Luna is the news. If the new Sol *feels* like the old Sol — same rhythm, same instincts, same willingness to push back on a bad idea — then OpenAI gets to compound the pricing story with the loyalty story. If it doesn't, expect a chunk of the power-user base to stay on 5.6 Sol via the API until it's deprecated.

One last data point from the thread: leokennis wrote that "ChatGPT Plus is basically limitless and 'just works' since 5.6. I've yet to give it some task it cannot do." That's the consumer-side commoditization signal. When your $20/month tier can't be exhausted by a normal user's tasks, the shape of the market changes — the frontier model becomes a loss leader, the mid-tier becomes the profit center, and the price cut on Luna suddenly reads as a competitive weapon aimed at exactly that middle.

What this means for your stack

Three specific things worth doing this week.

Re-run your cost model. If you're on Claude Opus 5.5 for orchestration or agent work, price out the same workload on GPT-6 Luna. Not Sol — Luna. For anything that isn't gated on genuinely hard reasoning, Luna is where the money now lives. Expect a 40–60% bill reduction on typical agent traffic, more if you're heavy on cache reads.

Re-evaluate your subscription tier. If your team is on Claude Code 20x and hitting limits, Codex Pro 20x is the current arbitrage per multiple threads. This will change — Anthropic almost certainly reprices in the next quarter — but the current-quarter delta is real. Do the math on your actual usage, not the marketing math.

Don't migrate your prompts blindly. The one place the community is unanimous is that prompts tuned to a specific model's quirks don't port cleanly. If you've spent six months tuning agent scaffolds against 5.6 Sol, budget time for a real bake-off before flipping traffic. Sol-to-Sol upgrades are usually fine; cross-lab moves need a proper eval harness.

Looking ahead

The next move is Anthropic's, and it's constrained. They can cut Opus pricing to match, which reads as reactive and squeezes margin during a capex-heavy year. They can hold the line and lean on the Claude Code developer experience, which has genuine loyalty but is also where Codex is closing fastest. Or they can ship a new Haiku/Sonnet tier priced against Luna and hope the frontier premium survives on Opus 6. Watch for a pricing announcement, not a model announcement, before the end of the quarter.