GPT-6 lands with Luna at half the price of 5.6

What happened

OpenAI shipped GPT-6 today in two SKUs: Sol, positioned as the reasoning and engineering model, and Luna, the cheaper general-purpose sibling. The launch post frames Sol as the successor to the 5.6 Sol model that a lot of developers had quietly settled on as their daily driver, and Luna as the volume workhorse for chat, search, and lightweight coding.

The headline number is pricing: GPT-6 Luna lands at roughly half the per-token cost of GPT-5.6 Luna, a cut steep enough that Simon Willison called it "a really big deal" within an hour of the announcement. Sol's pricing is more conservative — a modest step down from 5.6 Sol — but it now sits well underneath comparable Claude Opus tiers on every axis: input, output, cache reads, and cache writes.

The launch also ships the usual pelican-SVG canary tests, refreshed tool-use behavior, and — per the post — meaningful improvements in structured output reliability. The community reaction on Hacker News was unusually pragmatic: less "benchmarks!" theater, more spreadsheet math about what this does to monthly bills.

Why it matters

The interesting thing about this release isn't a new capability. It's that OpenAI has decided to compete on price at the exact tier where Anthropic has been most vulnerable. A back-of-the-envelope comparison posted in the HN thread has Claude Opus 5.5 at $4/M input and $20/M output; GPT-6 Sol undercuts that meaningfully while GPT-6 Luna sits at a fraction of it. For any team running agent loops — where output tokens dominate the bill and cache reads are the difference between viable and unviable — that gap compounds quickly.

There's a second, quieter story in the community reaction. Developer `m_fayer` described GPT-5.6 Sol as "a colleague you click with" — the first model whose engineering instincts they could actually predict. That kind of loyalty is rare, and it's the reason Anthropic's retention has held up despite pricing pressure. The open question with GPT-6 Sol is whether OpenAI preserved that feel or optimized it away. Early anecdotal reports are mixed: faster tool calls, tighter structured outputs, but a few people already saying the "jam session" quality feels different. If OpenAI broke what made 5.6 Sol lovable, the price cut won't save them with senior engineers.

The subscription-tier math is where this gets brutal for Anthropic. `jeffnash` laid it out: Claude Code 20x and Codex Pro 20x are nominally comparable products, but Codex's usage windows are more generous and less obscure, and Codex Pro users are already reporting they can do a full day of agent work without hitting a wall. With GPT-6's pricing baked in, the effective work-per-dollar delta between Codex Pro and Claude Code isn't marginal — it's the kind of gap that shows up on an engineering VP's monthly report and triggers a meeting.

There's also a signal here about where OpenAI thinks the market is going. Splitting the lineup into Sol and Luna is a bet that the median use case doesn't need reasoning — it needs a fast, cheap, reliable model that follows instructions. That's a very different bet than Anthropic's, which has effectively been "everyone should be running Opus and the price is the price." Both bets could be right. Only one is friendly to indie developers.

What this means for your stack

If you're running agent workloads on Claude, the honest move is to price out the same pipeline on GPT-6 Sol this week. Not to switch — the migration cost is real and the vibes matter — but to know the number. If Sol is 40% cheaper on your actual traffic and the quality delta is within a rounding error, you have a leverage conversation with your Anthropic rep or a migration plan. Either outcome is useful.

For lightweight tasks — classification, summarization, routing, the RAG-adjacent grunt work that eats surprising amounts of budget — GPT-6 Luna is probably the new default, full stop. The Claude Haiku tier isn't competitive at this price point, and Gemini Flash is a different set of trade-offs. Rewrite the model constant in your config, run your eval suite, ship it if the numbers hold.

For coding agents specifically, the Codex Pro 20x subscription is now genuinely the value play if you can tolerate the workflow. Claude Code's UX is still better in several ways — the diff review flow, the file-scoped tool calls, the way it handles multi-repo context — but at some point the price gap becomes the entire argument. That point may be now.

Looking ahead

The interesting question isn't whether Anthropic responds — they will, probably within the quarter — but whether they respond on price or on capability. A price cut concedes that GPT-6 Luna is a real substitute; a capability push (longer context, better tool use, better agent primitives) argues that Claude is a different product class. The latter is the harder, more defensible move. Watch which one they pick. It'll tell you what Anthropic thinks it's actually selling.