GPT-6 Sol and Luna Land on Devin: New Coding Models with Dramatically Lower Costs

OpenAI's two new models, GPT-6 Sol and GPT-6 Luna, are now officially available in the AI coding assistant Devin, accessible through both Devin Desktop and Devin CLI. What makes this update stand out isn't a dramatic leap in capability — it's a fundamental shift in cost structure, resetting the benchmark for price-to-performance in AI-assisted coding.

How Sol and Luna Are Positioned

Sol and Luna follow OpenAI's established pattern of differentiating models within the same generation. Based on their names and actual performance, Sol is positioned as the primary capability-focused model, while Luna is built around extreme cost efficiency. Launching both on Devin simultaneously means developers can switch between them based on task complexity and budget, rather than settling for a one-size-fits-all solution.

For teams that rely on AI-assisted coding day to day, this dual-model strategy offers finer-grained cost control. High-value, high-complexity tasks can go to Sol, while repetitive or bulk coding work can be offloaded to the more affordable Luna — reducing overall model call expenses across the development pipeline.

Key Numbers on FrontierCode 1.1

The data that really tells the story comes from FrontierCode 1.1, a coding-focused benchmark. According to official figures, GPT-6 Sol matches the performance of the previous-generation GPT-5.6 Sol while cutting per-task costs by 61%. In other words, users can achieve the same coding quality at roughly 40% of what they used to pay.

Luna's results are even more striking: it delivers better performance than GPT-5.6 Luna at just one-quarter of the cost, making it the lowest-cost model on the leaderboard. For developers who keep a close eye on model economics, this pushes the marginal cost per coding task into a new, lower range.

Why Cost Reduction Matters

Matching capability at half the price changes real-world usage patterns more profoundly than a pure performance boost. When equivalent-quality coding tasks become cheaper, workflows that were previously too expensive to automate at scale become viable. Bulk code generation, large-scale refactoring, automated testing and fixes in continuous integration — these high-frequency use cases are especially cost-sensitive, and lower prices directly expand the practical reach of AI coding tools.

FrontierCode 1.1 is a benchmark designed to measure AI models' overall capability on software engineering tasks, covering subtasks such as code generation, bug fixing, code completion, and unit test writing. Unlike general reasoning benchmarks (e.g., MMLU), FrontierCode focuses on actual output quality in realistic development scenarios — testing whether a model can produce runnable, logically correct code rather than just recalling facts. Its "per-task cost" metric incorporates model call expenses into the evaluation, making it possible to compare performance and economics on the same scale. That's part of why it has gained traction in the engineering community. That said, no single benchmark is comprehensive: FrontierCode 1.1 primarily reflects performance on specific categories of programming tasks, and dimensions like cross-language support and system architecture design are not fully covered.

Direct Impact on Devin Users

Devin is positioned as an autonomous AI software engineer, and one of its core strengths is completing development tasks end to end. Lower underlying model costs directly affect the economic viability of running long-chain tasks — a complex engineering task involving multi-step reasoning and repeated model calls accumulates costs quickly, so a drop in unit pricing gets amplified across the entire task lifecycle.

Both new models support desktop and command-line interfaces, covering everything from interactive development to scripted, automated pipelines. The CLI integration is particularly noteworthy: it makes it easy to embed Devin into existing development toolchains and CI/CD workflows, allowing AI coding capabilities to integrate into engineering practice in a more automated way.

CI/CD (Continuous Integration/Continuous Delivery) is the modern software engineering practice of automatically building, testing, and deploying code changes. Connecting AI coding tools via CLI to a CI/CD pipeline means tasks like code review, automated fixes, and test case generation can be triggered and executed without human intervention. Devin CLI makes this integration path more direct — developers can call Devin from configuration scripts in platforms like GitHub Actions or Jenkins, treating it as an automated node in the pipeline rather than just an interactive chat tool. The reduction in underlying model costs is especially significant in this context: CI pipelines are triggered frequently, potentially generating multiple model calls per commit, so lower per-call costs produce substantial cumulative savings at scale.

Things Worth Watching

The publicly available information so far is largely centered on FrontierCode 1.1 as a single benchmark. How Sol and Luna perform on other types of coding tasks, multi-language support, and complex system-level engineering still needs broader validation. Matching or exceeding benchmark scores doesn't necessarily translate directly into real-world project experience.

Additionally, cost figures are quoted in terms of "per-task cost" — translating that into actual billing requires factoring in task complexity and call volume. For teams evaluating whether it makes sense to migrate to the new models, a small-scale test against their own typical workloads is advisable before making a full switch. Overall, the core value of this update is that it lowers the price floor for high-quality AI coding capabilities — a clear positive signal for cost-sensitive development scenarios.