tech-news · 22 September 2026
GPT-6 Sol and Luna Shift the Agent Cost Curve
OpenAI's new mid-tier and low-cost GPT-6 models halve token prices and improve prompt caching, changing the economics of long-running agents more than the raw capability frontier.
OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September, filling out the cheaper end of its GPT-6 family. The headline is not a dramatic new intelligence record. It is that both models arrive at half the token price of their GPT-5.6 predecessors, alongside caching changes aimed squarely at persistent agents that repeatedly carry large instructions, tool definitions and working context.
That combination matters now because an agent rarely makes one neat model call. It plans, invokes tools, reads results, retries and keeps revisiting the same context. A modest saving on each step can become a material saving across a long session, while a cache miss can quietly erase much of the benefit.
What changed this week
OpenAI positions GPT-6 Sol and Luna as the mid-tier and high-volume members of the family, below GPT-6 Astra. Sol is priced at US$2 per million input tokens and US$10 per million output tokens, down from $4 and $20 for GPT-5.6 Sol. Luna drops to $0.10 input and $0.50 output, from $0.20 and $1.20. OpenAI says the reductions come from improvements in inference and caching rather than introductory discounts.
The intended split is practical. Sol targets recurring, relatively demanding jobs such as feature development, debugging, code review and data analysis. Luna is aimed at narrower work that runs at volume: extraction, classification, summarisation and short answers. Both are available through the API as gpt-6-sol and gpt-6-luna , and in ChatGPT Work and Codex for paid plans. Luna also reaches Free and Go users through the desktop app, although neither model launched in the ordinary Chat surface.
OpenAI reports gains in coding, computer use, factuality and professional-work evaluations, but the strongest part of its case is cost per completed task rather than a clean sweep on raw scores. That distinction is important. Vendor benchmarks can indicate where to test, but they do not substitute for running a representative workload with its real tools, prompts and acceptance checks.
The second half of the release is a revised prompt-caching system. According to OpenAI's caching announcement , eligible shared prompt prefixes reused within 30 minutes can receive up to a 90% cached-input discount. A new dashboard shows cached and uncached input, while a diagnostics tool reports why a request missed the cache and estimates how many tokens were affected.
Developers also gain explicit cache breakpoints and a way to change reasoning effort between turns without invalidating earlier reusable context. Tool definitions can remain stable while allowed_tools narrows what the model may call on a given turn. There is also a prewarming option for known instructions, schemas or reference material before the first user request.
Why it matters
For agent builders, these details point to a more useful design question than “which model is smartest?” The better question is which stage needs which level of intelligence, and how much repeated context each stage forces the system to process.
A workflow might use Luna to classify an incoming request, Sol to investigate and implement it, and reserve Astra for the smaller number of cases where deeper reasoning changes the result. That model-routing pattern is not new, but Luna's price makes routine calls cheap enough to reconsider tasks that were previously poor candidates for hosted inference. Sol, meanwhile, may offer a workable middle ground for coding agents that need more capability than a lightweight classifier but run too many turns to justify flagship pricing everywhere.
Caching could have an equally large effect. Long-running coding and research agents often resend repository guidance, tool schemas, policies and conversation history. Keeping that stable material at the front of the prompt, appending changes rather than rebuilding the whole prefix, and monitoring hit rates can reduce both latency and spend. It also introduces operational work: prompts become cache-aware artefacts, tool ordering matters, and a small integration change can have a cost impact that deserves an alert.
There are reasons not to accept the launch framing at face value. The Decoder's review cites early Artificial Analysis results showing Sol and Luna close to GPT-5.6 levels overall, with gains in some tests and regressions in others. It also notes that OpenAI selected favourable comparisons and that different reasoning levels make simple model-to-model conclusions difficult. The release may therefore be best understood as an efficiency move first and a capability move second.
The published cache discounts also describe eligible input, not the full cost of a job. Output tokens, uncached context, retries, tool execution and human review still count. A cheaper model that needs more retries can lose to a dearer one that gets the task right first time. Likewise, prewarming and a 30-minute reuse window help active sessions more than sporadic workloads.
What to watch next
The next useful evidence is not another launch chart. It is production data showing total cost per accepted result: cache-hit rate, latency, retries, review time and failure rate on stable workloads. Independent evaluations should also test Sol and Luna with the same agent scaffold, tools and reasoning budget rather than comparing whichever settings produce the best score.
It will also be worth watching whether developers actually adopt explicit cache management or find it too brittle, and whether OpenAI extends these models into regular Chat without changing their economics. For now, GPT-6 Sol and Luna make a credible case that the next phase of agent competition will be fought as much over efficient context reuse and workload routing as over the top benchmark number.