GPT-6 Sol and Luna are OpenAI’s mid-tier and small models in the GPT-6 family, launched at half the API price of their GPT-5.6 predecessors. Sol is the workhorse for complex coding and professional work; Luna is the cheap, fast option for high-volume everyday tasks. The headline promise is simple. About 50% cheaper per token, and about half as many mistakes. The details underneath those two numbers are where this release actually gets interesting, so this article works through what the independent benchmarks found, what OpenAI’s comparison charts really compare, and what developers said once they started running the models themselves. One note before the numbers. On September 29, 2026, at its DevDay event, OpenAI followed up with a newer GPT-6.1 Sol at the same list price, which enters the story later where it belongs.

What GPT-6 Sol and Luna cost now

The full list price for GPT-6 Sol and Luna is straightforward. Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20. Luna costs $0.10 per million input tokens and $0.50 per million output tokens, down from $0.20 and $1.20. OpenAI attributes the cut to caching and inference improvements, saying the lower serving costs let it pass “those savings directly on to users” as a 50% API price cut versus GPT-5.6 promotional pricing. An OpenAI spokesperson told VentureBeat the new rates are permanent list prices, not an introductory discount.

One row in that table is worth checking with a calculator. Luna’s output price falls from $1.20 to $0.50, which is a 58.3% cut, not 50%. Independent commentator P.K. Sharma, who checked every number in OpenAI’s announcement, noted that the “50% cheaper” label is carried across all four rows and that on the row where it is wrong, it understates OpenAI’s own discount. Astra stays untouched at $10/$50, roughly five times Sol’s price, which tells you this was a mid-tier and low-tier repricing, not a reshuffling of the flagship.

Beyond the list prices, two mechanics matter more than most coverage suggests. The Batch API gets an additional 50% off the already-halved rates, putting Sol’s batch output around $5 per million tokens and Luna’s around $0.25. And cached input reads keep their 90% discount, with Sol’s cached input at $0.20 and Luna’s at $0.01 per million tokens. For long agent sessions that reuse context, GPT-6 Sol and Luna’s cache discount decides more of the bill than the list price does.

The technical specs are plain enough to state in one breath. Both models run through the API as gpt-6-sol and gpt-6-luna, carry a 1,050,000-token context window with 128,000 maximum output tokens, and support reasoning efforts from none through max. At launch they went live in ChatGPT Work and Codex for paid tiers, with Luna also available in the desktop app for Free and Go users. Standard ChatGPT chat did not get them yet; the rollout there was gradual.

The half-the-errors claim, and the mechanism behind it

The company’s internal factuality test, built from de-identified real-world conversations where users had flagged mistakes by earlier models, shows GPT-6 Sol making about half as many mistakes as GPT-5.6 Sol, approaching what OpenAI calls Astra-level reliability at much lower cost. Luna, the company says, improves substantially too. At higher effort it matches GPT-5.6 Sol at about a hundredth of the cost. These are the company’s own numbers on conversations deliberately chosen because they produced errors, so they say little about typical use. Still, they are the claim the title of this article makes, and they deserve an independent check.

The independent check came from Artificial Analysis, and it confirmed the direction with a catch that most coverage skipped. On their AA-Omniscience benchmark, which penalizes confident wrong answers and applies no penalty for declining to answer, GPT-6 Sol’s hallucination rate fell from 92% to 60%, and Luna’s from 93% to 77%. Real, large drops. But a meaningful share of that improvement comes from the models declining to answer, alongside answering more accurately. Sol now attempts 83% of questions, down from 99% for its predecessor. Declining to guess cuts wrong answers by roughly a quarter. It also drags raw accuracy down about five points, from 59% to 54%, because some skipped questions were ones the model would have gotten right by luck.

This is not a worse model. A model that says it does not know instead of confidently inventing an answer is, in most real workflows, the better trade. But it differs from being “twice as smart,” and OpenAI’s launch materials leave that distinction unstated. On the combined AA-Omniscience Index, which factors in both correctness and hallucination avoidance, Sol still climbs from 22 to 27 and Luna from a negative score of -10 up to 1. The improvement is real. The mechanism is part abstention, part accuracy.

Where Sol gained and where Luna slipped

Independent testing of GPT-6 Sol and Luna tells a split story that OpenAI’s framing smooths over. On Artificial Analysis’s Coding Agent Index, GPT-6 Sol improved 2 points over GPT-5.6 Sol, with real gains on Terminal-Bench 4.0 (43% versus 37%) and SWE-Atlas-QnA (58% versus 54%), at roughly half the cost per task. That part matches the company’s own story. GPT-6 Luna went the other direction. It scored 2 points lower than GPT-5.6 Luna on the same index, with declines on SWE-Atlas-QnA (44% versus 49%) and DeepSWE v1.1 (64% versus 66%), even as its cost per task dropped by roughly 60%.

The knowledge-work results complicate things further. On GDPval-AA, a benchmark adapted from OpenAI’s own dataset of economically valuable tasks across 44 occupations, Sol dropped about 100 Elo points and Luna about 75. Luna also dropped around 45 Elo points on AA-Briefcase, a private evaluation of multi-week knowledge-work projects. After manually reviewing hundreds of outputs, the Artificial Analysis team attributed the regressions to reduced presentation quality and deliverables that omit rubric elements the older models used to include. It’s a strange kind of regression. Not dumber, exactly, but lazier at finishing the paperwork. There were improvements elsewhere too, including AutomationBench-AA (Sol 62% versus 60%, Luna 53% versus 50%), which keeps the overall picture mixed rather than bad.

In plain terms, Sol is a real, if modest, capability upgrade at half the price. Luna is mostly a price cut with a small quality tax attached, noticeable mainly if your work leans toward coding. The flat overall reading is why Artificial Analysis summarized the release as halving the cost per task while scores remain roughly level with GPT-5.6, a cost-efficiency release rather than a capability jump.

OpenAI’s charts, and the asterisk on the Claude comparisons

Here is the detail that changes how you read every “beats Claude” claim in this launch. Anthropic shipped Claude Opus 5.5 about 90 minutes before OpenAI’s own announcement went live, according to TechCrunch’s reporting. Yet most comparison charts in OpenAI’s launch materials benchmark GPT-6 Sol and Luna against Claude Opus 5, not the Opus 5.5 model that had been public for roughly an hour and a half by the time OpenAI hit publish. Per the post’s own footnote, some rows use Claude Fable 5 scores where Fable 5.1 figures were unavailable. For the broader context of how these models compare, our rundown of GPT-6 Astra vs Claude covers the flagship tier, and our coverage of Claude Opus 5.5 lays out what Anthropic actually shipped that afternoon.

To OpenAI’s credit, the sourcing footnote sits at the foot of the post, disclosing that rival scores were “taken from publicly available reports,” with Fable 5 numbers standing in where Fable 5.1 figures were unavailable. As P.K. Sharma put it after reading the whole post line by line, five of the six charts carry a Claude series, and OpenAI ran none of those Claude numbers. In every comparison, one side is a measurement OpenAI performed and the other side is a figure OpenAI read somewhere. That’s disclosed, it is normal industry practice, and it differs from a head-to-head test.

The effort settings deserve the same scrutiny. Reasoning effort trades tokens for score, and it moves both axes of OpenAI’s charts at once. The OSWorld 2.0 row shows GPT-6 Sol at its second-highest effort scoring 60.5% against Claude Opus 5 at its middle effort scoring 60.3%, with OpenAI reporting roughly 80% lower cost per task. Both halves of that sentence are true, and they make different claims. A near tie against a model not run at its maximum is a cost finding, not a capability finding. That caveat matters more than it looks, because the announcement’s framing often reads like both.

There is no GPT-6 Terra, and the naming confuses people

One deliberate omission in the GPT-6 Sol and Luna release: no GPT-6 Terra shipped. The old GPT-5.6 lineup had three tiers, and OpenAI skipped the middle one entirely without saying whether a replacement is coming. The gap landed differently depending on who you ask. On Hacker News, commenter jeffnash wrote that Terra had a true workhorse quality for production tasks, doing bulk code changes and following directions “without being too ‘smart,’” while Sol tended to go on “some rabbit hole side quest” to refactor codebases it had been asked to leave alone. That is a real constituency for the middle tier, and right now it doesn’t have a direct replacement.

The other confusion is positional. GPT-6 Sol and Luna are not smaller versions of GPT-6 Astra. Astra stays the flagship at roughly five times Sol’s price, and the AutomationBench numbers in OpenAI’s post still put Astra ahead of Sol even when Sol runs at its highest effort setting. Sol and Luna inherited some of Astra’s training methods, not its scores. Treating them as cut-down flagships will set the wrong expectations on both price and performance.

What developers said once they ran the models

Reading through the launch thread on Hacker News, the reaction that kept pulling my attention was not about price at all. Commenter m_fayer, describing a year of daily agent use, wrote that 5.6 Sol had been “some sort of sweet spot,” that its communication style and engineering instincts “just clicked,” and that it was the first model they had gotten attached to. The worry was not that the replacement would be worse. It was that it might just feel different in a way that is hard to get back. NorthSouthNorth echoed the same sentiment, noting that in 9 out of 10 tasks they still preferred working with 5.6 Sol even after getting Astra access.

That is a different complaint from the usual benchmark debate, and it sat alongside a practical skepticism that ran through the whole thread. Commenter sieve, who posted real 30-day usage numbers showing roughly 6.5 billion cached-read tokens against about 150 million fresh input tokens, put the skepticism directly: “I do not (generally) trust benchmarks. I only trust what a model does with MY code.” The mood, taken together, was warm on price, fond of the old model, and unwilling to let the launch charts settle anything on their own.

The independent developer who tested everything on launch day

The independent check I kept coming back to was Simon Willison’s launch-day writeup. Willison tests new models as they land, and he approached this launch the way he approaches every launch. He ran his standard pelican SVG rendering test across the whole GPT-6 and GPT-5.6 families, published the comparison grid, and wrote up what the price tables meant in practice. His verdict on the pricing was blunt. GPT-5.6 Luna was already his favorite model for building applications, and somehow GPT-6 Luna was half the price of that again.

Then he did something more committed than a test. He adopted GPT-6 Sol and Claude Opus 5.5 as his default models in Codex and Claude Code, and he upgraded his Datasette Agent demo to run on GPT-6 Luna, reporting that it seemed “fast and competent at both SQL queries and building HTML and JavaScript” for the demo’s purposes. An independent developer putting a new model into a production demo on launch day is a stronger signal than any press quote, because it carries the cost of being wrong.

What to check before you switch

If you are already running GPT-5.6 Sol in production, switching to GPT-6 Sol is close to a no-brainer on the numbers. Same or better scores, half the list price. But the bigger lever for your actual bill is probably the caching system OpenAI shipped alongside the GPT-6 Sol and Luna price cut. The new Prompt Caching Dashboard shows your real cache hit rates, and you can now change reasoning effort or tool availability mid-session without silently invalidating the cache, which used to happen constantly. For long agent sessions, GitHub’s report of more than 50% of prompt tokens no longer needing fresh processing across billions of Copilot requests shows how much of a bill lives in caching rather than list prices. Open the dashboard before you re-route anything.

If your workload leans heavily on coding and you are on GPT-5.6 Luna, test before migrating. The independent benchmarks show a real, if small, regression there, not just noise, and Luna’s accuracy gains are flat on some knowledge benchmarks. If you were on GPT-5.6 Terra, there is no direct replacement, and several developers report Sol at a lower reasoning-effort setting comes closest, with the caveat that it over-engineers more readily than Terra did.

Two more items belong on the checklist. The Batch API’s extra 50% discount is worth checking if any part of your pipeline can tolerate batch latency. And do not take “beats Claude” claims at face value without checking which Claude version is cited; the charts in OpenAI’s post use Opus 5, not the Opus 5.5 that shipped the same afternoon. A single question, asked in writing to any vendor making cost claims, covers most of this. Did you run both sides of the comparison, and at what effort setting? P.K. Sharma’s close reading of this launch suggests that question rarely gets a clean answer.

A newer Sol already exists, and that is worth stating directly

On September 29, 2026, at its DevDay event, OpenAI shipped GPT-6.1 Sol with the same $2 per million input and $10 per million output list price, and cached input halved to $0.10. It is positioned as a capability upgrade for agentic coding and computer use. The Sol and Luna launch page on OpenAI’s site carries a banner pointing to 6.1 Sol as the latest model. Nothing in the Sol and Luna pricing or availability changed because of it, but anyone evaluating a switch should test both and compare real task costs, not just list prices.

GPT-6 Sol and Luna remain what the launch promised, an actual pricing reset on capable mid-tier and small models, with real but partial improvements in factuality and a mixed capability record that the independent numbers describe more honestly than the launch charts do. The price cut is the part of the story that holds up under every check. The rest rewards close reading, which is exactly what the developer community gave it.

FAQ