NEW DELHI, September 23, 2026, 5:04 PM IST — OpenAI has launched GPT-6 Sol and GPT-6 Luna, extending its newest model family into lower-cost tiers aimed at software development, agentic workflows and high-volume automation.
The release matters for engineering teams because the headline change is not simply another benchmark cycle. OpenAI has cut the listed API price of both tiers by 50% compared with the promotional pricing of their GPT-5.6 predecessors, while adding the tool use, long-context support and adjustable reasoning controls needed for production agents. The result is a wider price-performance range for teams that have been reserving frontier models for only their hardest tasks.
What OpenAI confirmed
According to OpenAI’s launch announcement, GPT-6 Sol is positioned for complex coding and recurring agentic work, while GPT-6 Luna targets focused, high-volume jobs such as extraction, summarisation, classification and routing. Both models are available through the OpenAI API as gpt-6-sol and gpt-6-luna.
Standard short-context pricing for Sol is $2 per million input tokens and $10 per million output tokens. Luna is listed at $0.10 per million input tokens and $0.50 per million output tokens. Cached input reads cost one-tenth of ordinary input, while requests with more than 272,000 input tokens use higher long-context rates. Those thresholds make prompt shape and cache efficiency important parts of any realistic cost comparison.
The models each support a 1.05-million-token context window and up to 128,000 output tokens, according to the GPT-6 Sol and GPT-6 Luna model pages. Both accept text and image input and support function calling, structured outputs and OpenAI’s built-in tools through the Responses API. Fine-tuning is not currently supported.
OpenAI also says the models are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna is available to Free and Go users in the desktop app, but neither model is yet offered in the main Chat experience.
A two-tier model for production agents
Sol and Luna are intended to sit below GPT-6 Astra, which remains OpenAI’s highest-capability tier. OpenAI describes Sol as the default choice for demanding reasoning, coding and multi-step tool use. Luna is designed for repeatable work at scale, where latency and unit economics can matter more than squeezing out the final increment of capability.
That split mirrors how mature platform teams already route workloads. A repository-wide refactor, incident investigation or complex infrastructure change may justify Sol. Log classification, ticket triage, document extraction and deterministic transformations may fit Luna, provided application-level evaluations show acceptable accuracy. Astra remains the escalation path for unusually difficult or consequential tasks.
OpenAI reports that Sol at maximum reasoning effort scored 68.8% on DeepSWE 1.1, within 1.1 percentage points of the Claude Fable 5 result cited in the company’s comparison, at an estimated 80% lower cost per task. Luna at maximum effort scored 66.6% in the same evaluation. These are vendor-reported benchmark results, not a guarantee of production performance, and teams should reproduce them against their own repositories, tools and failure criteria.
Caching becomes an operational control
The launch also changes the economics of long-running agents through improved prompt caching. OpenAI says cached input reads receive a 90% discount and that reasoning effort or tool availability can be changed without invalidating earlier cached context. The company has added a caching dashboard, diagnostics and explicit breakpoints to help teams inspect where reusable prompt prefixes end.
For DevOps and platform engineers, that turns cache hit rate into an observable cost and latency signal. Stable system instructions, repository context and tool definitions can be placed in reusable prefixes, while volatile user input stays later in the prompt. This is closely related to the context and lifecycle discipline covered in GravityDevOps’ LLMOps guide.
OpenAI says GitHub reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of model requests over several months. The company did not publish enough detail in the launch post to treat that figure as a universal saving, so teams should measure their own hit rates and end-to-end request costs.
AWS availability broadens deployment options
Amazon Web Services separately confirmed that both models are generally available through Amazon Bedrock. AWS says the models support up to one million tokens of context in Bedrock and can be governed through established identity, security and model-invocation audit controls.
Bedrock availability gives AWS-centric teams another route to adoption without immediately redesigning their cloud control plane. It does not remove the need to check regional availability, data handling, quotas and feature parity. OpenAI’s documentation says EU data residency for Sol and Luna is available only with Standard processing, and regional processing carries a 10% premium where offered.
What developers and platform teams should do now
The practical next step is an evaluation, not a wholesale model swap. Teams should run the same representative tasks across their current model, Sol and Luna; record success rate, latency, token use, tool-call failures and human review time; and set routing rules from those results. The cheaper model wins only when it meets the workload’s quality and safety bar.
API users should also review integration compatibility. OpenAI recommends the Responses API for built-in tools and function calling. Chat Completions supports function calling for these models only when reasoning effort is set to none. Applications that depend on custom sampling settings, fixed response shapes or provider-specific tool semantics need regression tests before production rollout.
For coding agents, evaluation should include mergeability rather than pass rates alone: test quality, scope discipline, repository conventions, secret handling and rollback behaviour. The same principle applies to CI automation. A lower per-token price can increase total spend if it encourages longer runs, excessive retries or broader permissions. GravityDevOps’ comparison of CI/CD tools provides useful context for where approval gates and audit trails belong.
The broader signal
The simultaneous arrival of cheaper GPT-6 tiers and Anthropic’s lower-cost Claude Opus 5.5 points to a new phase of frontier-model competition centred on cost per completed task, not just peak capability. That is constructive for buyers, but benchmark claims from competing vendors use different harnesses, effort levels, fallbacks and pricing assumptions.
For technical decision-makers, the durable advantage will come from portable evaluations and policy-based routing rather than loyalty to a single leaderboard. Sol and Luna make capable agents less expensive to test at scale. Whether they reduce production cost will depend on caching, task design, review burden and the failure rate that matters in each system.
Sources
Primary sources include OpenAI’s launch announcement, official API pricing and model documentation, plus AWS’s Bedrock availability notice. Independent launch reporting was reviewed from TechCrunch.
