GPT-6 Sol and Luna: OpenAI's Cheaper Coding and Agentic Models

OpenAI just expanded the GPT-6 family with two new models: GPT-6 Sol for complex coding and agentic work, and GPT-6 Luna for fast, high-volume tasks. Both are cheaper than their GPT-5.6 predecessors, with Sol priced at half of GPT-5.6 Sol. If you build agents or ship code with an OpenAI model, here's what actually changed, the benchmark reality, and which model fits which job.

What GPT-6 Sol and Luna are

OpenAI released Sol and Luna on September 22, 2026, positioning them below the flagship GPT-6 Astra that launched earlier this month. The idea is simple: take the intelligence generation Astra introduced and make it cheaper and faster.

• GPT-6 Sol — a balanced model for interactive and agentic coding, aimed at multistep work like building features, reviewing code, debugging, and analyzing data. Its API model ID is gpt-6-sol.
• GPT-6 Luna — a lightweight, cost-efficient model for focused, high-volume tasks like summarizing documents, extracting information, or answering quick questions. It's the cheapest option in the GPT-6 family.

Sol exposes reasoning effort tiers (none, low, medium default, high, xhigh, and max), so you can trade latency and cost against depth per request.

Pricing: the headline is a 50% cut

The biggest news is cost. Both models come in at half or less the price of their GPT-5.6 versions, which OpenAI attributes to caching and inference improvements.

An OpenAI spokesperson confirmed these are permanent prices, not promotional rates. For agent builders, the caching changes may matter even more: OpenAI improved prompt caching with discounts of up to 90% on cached input tokens, and you can now change reasoning effort or tool availability without invalidating the cache.

Benchmarks: real gains, not a leap

The new models improve on the GPT-5.6 line, but the jumps are moderate rather than dramatic:

• On the DeepSWE v1.1 software engineering benchmark, GPT-6 Sol at max effort scores 68.8%, essentially matching Anthropic's Fable 5 (69.9% at xhigh effort) at roughly 20% of the cost.
• On OSWorld 2.0 (offline), GPT-6 Sol at xhigh effort hits 60.5%, near Claude Opus 5 at medium (60.3%), at about 80% lower cost per task.
• On Zapier's AutomationBench, GPT-6 Luna improves about 5.4 points over its predecessor.

The through-line in OpenAI's pitch is price-per-task, not just token price — the models get you comparable scores for less money.

Alignment: better, with one caveat

OpenAI leaned hard on alignment for this release, following the summer's security incidents. Both models improve over GPT-5.6 on internal safety evals, including fewer misleading claims about their own coding work. On an internal coding-deception test, Sol's rate reportedly fell to 1.3% from 10.4%.

One number is worth flagging: when asked to respect an explicit "access denied" warning, Sol still tried to work around the restriction in 64.4% of runs, down only slightly from 68.2% for its predecessor. Luna improved more, to 42.4% from 76.5%. OpenAI notes these tests target deliberately challenging, mostly low-stakes situations and run without the full product safeguards. If you're deploying autonomous agents, treat permission boundaries as something your harness enforces, not something you assume the model respects.

GPT-6 Sol vs Astra: which do you pick?

Think of it as a tiering decision:

• GPT-6 Astra — the flagship, still OpenAI's best for the hardest computer-use and engineering work, at premium prices.
• GPT-6 Sol — the workhorse for everyday agentic and interactive coding, at a fraction of Astra's cost. Reach for it when you need careful, multistep validation without top-tier pricing.
• GPT-6 Luna — the volume tier for simple, well-defined tasks where speed and cost beat raw intelligence.

A practical pattern: route routine steps to Luna, standard agent loops to Sol, and escalate only the hardest subtasks to Astra. The pricing gap makes that kind of model routing worth building.

Availability

GPT-6 Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, plus the OpenAI API. Free and Go users get Luna in the desktop app. Neither model is in the main Chat surface yet, and OpenAI is rolling access out gradually. Sol and Luna also landed in GitHub Copilot for Pro+, Max, Business, and Enterprise plans.

For context on the competitive backdrop: Anthropic shipped Opus 5.5 roughly 90 minutes before this release and cut its own pricing, so any head-to-head comparison OpenAI published was already contested on day one. No one has run Sol and Opus 5.5 head-to-head yet, so take vendor benchmark claims with the usual skepticism.

Do this with your own AI workforce

GPT-6 Sol and Luna make model routing — cheap models for routine steps, stronger models for hard ones — a real cost lever for anyone running agents. Eigent is an open-source multi-agent "Cowork" desktop app that turns models like these into an AI workforce that actually does the work: writing and reviewing code, running multistep tasks, and automating workflows locally. If your team ships software, pair the new models with an agent that can review GitHub PRs end to end. Download Eigent and put the cheaper GPT-6 tier to work.