Price War Begins: GPT-6 Sol and Luna Halve Model Costs

Today the calendar bunched two launches into the same window. OpenAI said it has added Sol and Luna next to Astra to form a three-model family; there is no Terra on the list. The announcement page refreshed with a dedicated animation while stressing that Astra remains the most capable model at the top. I had already found the Astra family fluent in cross-tool work, and now I am watching whether Sol and Luna can match that natural coworker feel. Anthropic landed on the same day with Opus 5.5; the industry looked locked to one date and the price competition lit up.

Pricing is the core of the story. GPT-6 Sol cuts the previous Sol generation from four dollars to two dollars on input and from twenty dollars to ten dollars on output — exactly half. Input tokens are what you send to the model and output tokens are what it generates, like an order and the dish that arrives. GPT-6 Luna goes further: it was already eighty percent off and now adds another fifty percent, with cache reads discounted up to ninety percent. Cache reads mean avoiding recomputation when the same context is reused — when an agent reads the same file a hundred times, ninety-nine of those reads cost almost nothing. For day-to-day work that makes Luna notably cheaper and at the same time smarter, which shrinks the API bill directly.

Positioning inside the family is clear. Astra stays at the top as the reference for demanding reasoning, Sol serves complex coding and agentic flows, and Luna is the most efficient option for focused high-volume tasks. Think of a kitchen: Astra is the head chef, Sol the sous-chef, Luna the fast and precise assistant — you do not need the whole brigade for every question, the right specialist steps in. The docs describe Sol for complex code and agent workflows and Luna for bulk work, with knowledge cutoffs listed as April 2026 for Astra and May 2026 for Luna. If I take my earlier Astra experience — faster and more effective communication than Fable — as a baseline, whether Sol and Luna reach the same line will decide everyday comfort.

One of the places these claims were measured is AutomationBench. Built by Zapier for realistic business workflows, it checks whether agents can coordinate across apps, discover the right API on their own, and stay within policy. A single test has limits — the vendor chart rightly puts what it sees as most effective at the top, but field validation is still needed. The question I care about is whether a model scoring 33 in the lab runs at the same pace on real customer data. So an AutomationBench score alone does not settle it; it should be read alongside companion tests for computer use and software engineering.

Start with the smallest member, Luna. Versus the prior Luna there is a jump from 9.1 toward 14 — near double — while another scale shows a modest 20 versus 17. The cost side is clearer: the 0.070 dollar band against 0.037 dollars and around 0.021 with a fifty to fifty-five percent edge. A small example: if you route tens of thousands of short tasks a day (such as ticket tagging or report summarization) to Luna, you close the same work at roughly half the cost and higher accuracy. Luna now behaves like an agent that gathers in the background, and that character fits the numbers: more work, smaller invoice.

The real gap appears with Sol. On OpenAI's chart the new Sol in extra-high mode shows 33.2 points for 0.27 dollars, while the prior Sol at its max setting managed only twenty-eight percent for 0.67 dollars and twenty-six percent for 0.54 dollars in extra-high. The new Sol is therefore both smarter and close to twice as cheap. To make it concrete in three steps: 1) define the same automation job, 2) run at the same difficulty setting, 3) put score and bill side by side. The result is that Sol delivers a higher score for less money than its predecessor. Because the per-million token price halved on both input and output, savings compound across agent loops.

The comparison with Astra makes it more telling. Astra Medium runs 34.1 points for 1.27 dollars while GPT-6 Sol lands nearby — roughly 33 points — for 0.27 dollars, so almost the same score at one-fifth the cost. On the chart Sol sits between Astra Light and Astra Medium, which matters because many teams find Medium sufficient. On the rival side Fable 5.1 spends 2.45 dollars for 31.4 points while Sol spends 0.27 dollars for 33 points — nine times more expensive and still behind. Anthropic itself frames Fable 5.1 as its most capable model and Opus 5.5 as a forty percent cheaper alternative for Fable-level work; which holds sustainably in the field will become clear only with repeated testing beyond a single graphic.

Other fronts are mixed but instructive. On frontier coding tests the leaderboard shuffles quickly; costs climb while scores stay close. On computer use the chart shows 5.6 Sol passing Astra Light at one point — that is the model's ability to operate a browser or desktop on its own, such as filling forms, moving files or clicking inside apps. OpenAI describes a more native coworker tone for Sol and Luna in that area — less unnecessary verbosity, more context-aware brevity — which saves time in daily agent flows. For long-horizon suites like OSWorld 2.0 with 108 desktop tasks, we still need independent results to see how durable that edge is.

Error rate and cache economics tell a quieter but more important story than price alone. Luna drops to a very low error band on the scale while Astra remains at the lowest band; by error I mean the model misleading or misdirecting. When you run a company you cannot master every employee's job step by step — results and trust matter — so a small misdirection from an agent that quickly produces output on an unfamiliar topic can get expensive. That is why falling error directly lifts trust. On the cache side the discount of up to ninety percent trims the line that dominates agent bills when the same context is read hundreds of times; reading the same research file across ten agent steps now costs about as much as a single read.

Access is spelled out: GPT-6 and Luna are available in ChatGPT Work, Codex, Plus, Pro, Enterprise and Education and for Go and open users via the desktop app; not on every chat surface, but directly by model identity on the API. According to OpenAI help pages GPT-6 Astra rolls out in chat as GPT-6 Pro for Pro 100 and Pro 200 plus Business and Enterprise on a gradual rollout; there is no chat inclusion for Plus. For developers the model identifiers are distinct, with cutoffs in May and April 2026. Pricing also lists separate lines for long context and cache writes, so it pays to count input plus output plus cache reads and writes together when budgeting.

What does this mean in practice? For everyday work, research and intermediate tasks Luna is now the obvious candidate; it offers similar smarts at far lower cost so the price of experimentation falls and serving more people becomes realistic. Sol, at near-Astra scores for one-fifth the bill in code and agent loops, is positioned to become the default for teams. On the Anthropic side Opus remains expensive and often overkill; the two vendors are clearly aiming at different spots. Today's winner is the user — cheap, premium and smart options sit on the same table. I will run Sol and Luna against Opus 5.5 in scenarios that matter for your work and share the results in the coming days; together we will sort where to chase cheap and where to chase solid reasoning.