GPT-6 Sol and Luna: half the price, mixed scores

OpenAI's GPT-6 Sol and Luna are a price cut with the scores left mostly where GPT-5.6 put them. Artificial Analysis, in its September 25, 2026 note, puts Sol (max) at an Intelligence Index of 55, up one point from GPT-5.6 Sol (max), and Luna (max) at 40, up two from GPT-5.6 Luna (max). The same note says Sol's list price dropped from $4 / $20 to $2 / $10 per million input and output tokens, and Luna from $0.20 / $1.20 to $0.10 / $0.50, with the 90% cache-read discount and 25% cache-write premium unchanged.

That is the release. Not a new leaderboard king. A cheaper pair that Artificial Analysis says sits on the cost-efficiency frontier, while losing ground on the knowledge-work evals that actually look like office work.

Sol at $1.06 a task, Luna at $0.07

Artificial Analysis prices an Intelligence Index run, not a chat session. GPT-6 Sol (max) costs $1.06 per task, about half of GPT-5.6 Sol (max) at $1.99. GPT-6 Luna (max) costs $0.07, about 60% under GPT-5.6 Luna (max) at $0.18. The cut is the driver. Token use went the other way: Sol's output rose from 29k to 31k tokens per task, Luna's from 41k to 51k. You pay less per token and the model writes a bit more, and the bill still falls.

OpenAI's own post, dated the same day, frames the models as "the smartest and most cost-efficient models" the company has shipped, with Sol described as the frontier model for enterprise and coding and Luna as the smaller, faster one for everyday work and high-volume processing. The API strings are gpt-6-sol and gpt-6-luna. Both are in the API now, plus Codex and ChatGPT. Microsoft put them in Microsoft Foundry the same day, and GitHub added them to GitHub Models in public preview, with Copilot credit billing on eligible plans.

All four rows are from Artificial Analysis's September 25 write-up, using the max reasoning setting and OpenAI's Codex harness for the coding index. Cache reads stay at 90% off and cache writes at a 25% premium, which is the part that matters if your agent loop rereads the same files. OpenAI says priority processing is 2.5x standard rates and keeps the same 15% batch and Flex discount as prior models.

Hallucination fell because Sol started refusing

The cleanest quality change is on AA-Omniscience, Artificial Analysis's knowledge and hallucination benchmark. GPT-6 Sol (max) cuts its hallucination rate from 92% to 60%. GPT-6 Luna (max) goes from 93% to 77%. Sol's AA-Omniscience Index moves from 22 to 27. Luna's moves from -10 to 1.

Read the mechanism before you celebrate. Sol attempts 83% of questions, down from 99% for GPT-5.6 Sol (max). Wrong answers fall by about a quarter. Accuracy also falls, from 59% to 54%. Luna's accuracy is basically flat, 44% versus 43%, while it answers fewer questions. Artificial Analysis is explicit that Sol bought the hallucination drop by declining more often.

Sol achieves this by declining to answer more often: it attempts 83% of questions vs 99% for GPT-5.6 Sol (max), which cuts wrong answers by about a quarter but also lowers accuracy 5 points from 59% to 54%.

That quote is Artificial Analysis, not OpenAI. OpenAI's launch post does not mention AA-Omniscience, attempt rates, or the accuracy dip. If you are wiring these models into a system that treats a refusal as a failure, the hallucination chart and your production error rate will move in opposite directions.

Coding up for Sol, down for Luna

In OpenAI's Codex harness, GPT-6 Sol (max) scores 57 on the Artificial Analysis Coding Agent Index, up 2 from GPT-5.6 Sol (max). The gains Artificial Analysis names are Terminal-Bench 4.0 at 43% versus 37%, and SWE-Atlas-QnA at 58% versus 54%. Cost per coding task is $2.99, about half the predecessor, which is why Artificial Analysis puts Sol on the Pareto frontier of Coding Agent Index versus cost.

Luna goes the other way. GPT-6 Luna (max) scores 41, down 2 from GPT-5.6 Luna (max) at 43. SWE-Atlas-QnA falls from 49% to 44%. DeepSWE v1.1 falls from 66% to 64%. The cost per task is about 60% lower. You are buying a cheaper coding agent that is slightly worse on the index Artificial Analysis uses, not a smaller Sol.

OpenAI's own coding claims point at a different harness. The launch post says GPT-6 Sol is OpenAI's best model on Terminal-Bench 2.0 and Terminal-Bench 2.0 Hard in the Codex CLI harness, and the strongest they have ever produced on an internal agentic coding evaluation. It also cites a 16% hallucination reduction on an internal knowledge-work eval with CharXiv, BrowseComp, and GPQA Diamond, and a 20% reduction in the share of responses containing a factual error on a production ChatGPT sample. Those numbers are OpenAI's, on OpenAI's sets. They are not the same as the Artificial Analysis regressions below.

GDPval-AA dropped about 100 Elo for Sol

Outside Omniscience, Artificial Analysis records a split. Both models improve on AutomationBench-AA (Sol 62% vs 60%, Luna 53% vs 50%) and on Terminal-Bench 4.0 inside the Intelligence Index (Sol 44% vs 40%, Luna 13% vs 12%). Then two knowledge-work evals go backwards.

On GDPval-AA v2.1, adapted from OpenAI's dataset of economically valuable tasks across 44 occupations, Sol drops about 100 Elo and Luna about 75. On AA-Briefcase v1.1, a private eval of multi-week knowledge-work projects with thousands of input files, Luna drops about 45 Elo and Sol is flat. Artificial Analysis says its team manually inspected hundreds of outputs. The pattern they describe is shorter deliverables, weaker presentation, and omitted rubric elements. Not a collapse in reasoning. A model that stops before the brief is finished.

That is the part I would not let a pricing slide bury. A 50% token discount is real money on a coding loop that already worked. It is a worse trade if the job is a GDPval-style deliverable and the new model leaves pieces of the rubric on the table. Artificial Analysis does not publish the raw Elo numbers, only the deltas, so "about 100" is their phrasing, not a score you can recompute from the post.

OpenAI's release notes do name concrete product behavior that lines up with "shorter, then stop." In ChatGPT, both models are more concise, with Sol a little chattier than Luna. Sol has a new /search command for explicit web lookups. Luna is better at staying on task across long conversations. The notes also say both models are more direct about limits: they refuse overly broad requests, say when they lack information, and avoid fabricating specifics when they are unsure. Personality controls are temporarily limited. Instant, thinking time, and thinking depth are gone. Proactive mini-app creation is disabled for both. Sol keeps mini-app creation when you ask. Luna does not, yet.

Four new tools, and a personality OpenAI says is colder

The launch post lists four tools that are new with this generation.

• Tool search. Instead of loading every tool definition into the prompt, the model retrieves tools on demand. OpenAI says this improves accuracy on their internal eval when many tools are available. No public accuracy delta is attached.
• Memory. The model creates, updates, and deletes its own stored memories via tool calls, rather than a separate memory layer. OpenAI says this produced a 55% reduction in overall memory errors on their internal eval. That 55% is OpenAI's number, on OpenAI's eval, and they do not publish the baseline error rate.
• Image gen. An image generation tool with configurable quality and aspect ratio. OpenAI claims a 30% win rate against GPT-5.6 Sol Image Gen on an internal arena of 2,000 prompts. Again, internal.
• Code mode. TypeScript written against tool interfaces, executed in a sandbox, instead of a JSON tool call. OpenAI says this lets the model filter and restructure tool results before they re-enter context.

None of those four come with a public dataset, a prompt list, or a third-party rerun as of the launch posts. Treat the 55% and the 30% as vendor claims with a named internal eval, not as leaderboard scores.

The tone change is the piece users will notice before any Elo chart. OpenAI says prior models were trained to keep users happy, and that this produced flattery and invented certainty. GPT-6 is trained to be more direct. The company quotes its own description of the result as less effusive, more willing to push back, and more comfortable saying "I don't know." OpenAI also says the tone can feel curt or cold, and that it is working on a warmer balance that does not bring the flattery back. That is a product bet, stated in the launch post, not a measured satisfaction score.

Where it sits against Opus 5.5 and Grok 4.7

Three days before this launch, on September 22, Artificial Analysis reported Claude Opus 5.5 at 58 on the Intelligence Index, with a 20% price cut and a larger cache-hit discount. GPT-6 Sol (max) at 55 does not take that spot. Four days before, on September 21, the same index put Grok 4.7 at 46 and said its Coding Agent Index had overtaken GPT-5.6 Sol. GPT-6 Sol (max) at 57 on that coding index is the number that has to be re-compared. Artificial Analysis's GPT-6 post does not reprint the Grok 4.7 coding score, so this post will not invent the gap.

What the September 25 note does say is narrower and more useful: Sol and Luna stay roughly level with their GPT-5.6 counterparts on the Intelligence Index and the Coding Agent Index, with gains in some evals and losses in others, at about half the price. OpenAI's claim that the pair captures "a significant portion of the cost efficiency Pareto frontier" is Artificial Analysis's phrasing, and it is specifically about cost per task on those two indexes.

Access is already wide. API, Codex, and ChatGPT on day one, per OpenAI. Microsoft Foundry the same day, with the company pointing at Fortune 500 deployment. GitHub Models in public preview, billed in Copilot credits if your plan includes them. OpenAI's cookbook lists a phase parameter with commentary, final_answer, and an optional preambles preamble before a tool call, aimed at multi-agent traces where preamble text was being dropped. That is a plumbing change for people running the Responses API, not a consumer feature.

What nobody in these sources has published is a head-to-head of GPT-6 Sol against Opus 5.5 on GDPval-AA, or a public rerun of the 55% memory-error claim, or the raw Elo values behind the "about 100" drop. Until those exist, the decision the sources actually support is simple. If you were going to call GPT-5.6 Sol or Luna at max effort, GPT-6 is the cheaper call and, on Artificial Analysis's coding index, a small step up for Sol and a small step down for Luna. If the work product is a long knowledge-work deliverable, read the GDPval regression before you flip the default.