Key takeaways:
OpenAI introduced GPT-6 Sol and GPT-6 Luna on 22 September 2026, which brought the GPT-6 family under the flagship group Astra that was introduced on 3 September.
Both models are approximately half the cost per token relative to the GPT-5.6 models, which OpenAI claimed to be a fixed rate, not a limited-time offer.
Sol: Conquers the hardest code and professional work. Luna: Gets the heavy work out of the way, whether it be summarization, information extraction, or customer service. Neither replaces Astra for the hardest jobs.
For German and EU businesses, using these models still means checking data residency, GDPR processing terms, and EU AI Act transparency duties before routing customer data through the API.
Early benchmark results show Sol and Luna roughly matching rival models from Anthropic at a fraction of the cost per completed task, though OpenAI's own figures are the only numbers available so far.
OpenAI released two new models this week: GPT-6 Sol and GPT-6 Luna. Both sit below the flagship GPT-6 Astra model, which launched earlier this month. Together, the three models now form OpenAI's full GPT-6 lineup, with no GPT-6 Terra announced.
The headline is price. Both new models cost about half of what their GPT-5.6 predecessors charged, and OpenAI says the cut is permanent this time. Here's what's actually in OpenAI's own announcement, and what it means if you're deploying these models from Germany or elsewhere in the EU.
What OpenAI Announced
OpenAI describes Sol and Luna as models built with training methods similar to Astra's, aimed at bringing Astra's improvements to faster, cheaper versions. Both are API-only, live now as gpt-6-sol and gpt-6-luna, with no downloadable weights to self-host.
The GPT-6 family now has a clear three-tier shape. Astra handles the hardest, most complex work: multi-step reasoning, difficult science and math problems, and tasks spanning multiple types of media. Sol is positioned for complex coding and professional tasks performed repeatedly. Luna is built for fast, high-volume, simpler jobs like summarization, extraction, and straightforward Q&A, occupying roughly the same role in the lineup that GPT-5.4 Mini and Nano played in the previous generation.
OpenAI attributes the price cut to improvements in caching and inference efficiency, which let the company serve these models more cheaply and pass the savings on directly.
Pricing Breakdown
Model | Input (per 1M tokens) | Output (per 1M tokens) | Vs. GPT-5.6 equivalent |
GPT-6 Sol | $2.00 | $10.00 | Down from $4.00 / $20.00 |
GPT-6 Luna | $0.10 | $0.50 | Down from $0.20 / $1.20 |
Sol's cached input rate comes in at $0.20 per million tokens for prompts up to 272,000 tokens, with larger prompts priced at double the input rate and 1.5 times the output rate. Batch and Flex processing run at half price, and a faster response mode costs double the standard rate.
Luna's cached input rate sits even lower, at $0.01 per million tokens. OpenAI also rolled out an improved prompt caching system across both models, with cache hit discounts of up to 90% and a 30-minute reuse window for repeated prompt prefixes, a detail that matters most for teams running agentic workflows that resend the same instructions and tool history on every turn.
Where the Benchmarks Land
OpenAI's own figures show both models performing close to, or on par with, competing frontier models, at a noticeably lower cost per completed task. A few of the headline comparisons:
On the DeepSWE v1.1 coding benchmark, Sol scored 68.8% at maximum effort, close to Claude Fable 5's 69.9% at its highest effort setting, at roughly 20% of the cost per task.
Luna scored 66.6% on the same benchmark, which OpenAI says is comparable to Opus 5 and Fable 5 running at medium effort, at a much lower cost.
On the OSWorld 2.0 computer-use benchmark, Sol edged out Opus 5's medium-effort score (60.5% vs. 60.3%) while costing roughly 80% less per task.
Worth flagging: every one of these numbers comes from OpenAI's own testing, as reported by outlets including TechCrunch, VentureBeat, and MarkTechPost. Independent, third-party verification of these specific scores wasn't available at the time of writing, so treat them as the vendor's account until outside evaluators weigh in. For more on how Opus 5 itself performs, see our Claude Opus 5 review.
Factuality and Alignment Claims
OpenAI also reports fewer factual mistakes in this generation. Using an internal test built from de-identified ChatGPT conversations where users had flagged an error in an earlier model's answer, Sol made roughly half as many mistakes as GPT-5.6 Sol on the same set.
OpenAI is upfront that this test set was deliberately chosen for being error-prone, so it doesn't represent typical everyday use. On alignment testing, both models reportedly scored better than their GPT-5.6 predecessors, including making fewer misleading claims about their own coding work.
Competitive Context
The timing here is notable. Anthropic released Claude Opus 5.5 the same day, reportedly about 90 minutes before OpenAI's announcement went out, according to The Next Web's report citing TechCrunch. Neither company appears to have coordinated the timing, but it puts both companies' pricing and performance claims side by side for anyone comparing options this week.
OpenAI also said earlier this week that it will now let outside groups run technical safety evaluations during training, rather than only reviewing models before release. That's a meaningful shift in how frontier labs are handling external scrutiny, and it fits a broader industry pattern we cover in our piece on governance and sandboxed execution in agentic AI.
What This Means for German and EU Businesses
None of the excitement around pricing changes the compliance homework. A few things worth checking before routing any business data through GPT-6 Sol or Luna.
GDPR still applies in full. OpenAI is a US-based provider. If personal data of EU residents passes through the API, you need a valid data processing agreement, clarity on where that data is actually processed, and an appropriate transfer mechanism such as Standard Contractual Clauses or reliance on the EU-US Data Privacy Framework. We could not confirm at time of writing whether GPT-6 Sol and Luna specifically support EU data residency; check OpenAI's current terms directly before sending customer data through either model.
EU AI Act transparency duties already apply to OpenAI as a provider. General-purpose AI model transparency obligations, including summaries of training content, have been binding since August 2, 2025. High-risk system obligations, which would matter more if you're using these models for things like automated hiring decisions or credit scoring, remain deferred under the May 2026 Digital Omnibus agreement to December 2027 and August 2028, depending on the system type.
Human review still matters. OpenAI's own factuality test shows a real, if reduced, error rate on exactly the kind of task the test was built to catch. For any German business publishing or acting on model output, particularly anything customer-facing, that's still a reason to keep a person checking the work.
If you're comparing GPT-6 against other providers available in Germany, our breakdown of suitable AI subscription plans in Germany covers how ChatGPT, Claude, Gemini, Perplexity, and Grok stack up for local buyers.
Our Initial Assessment
The actual news here isn't a capability leap. Both models land close to their predecessors and close to rival models on most benchmarks. The real story is cost: OpenAI is making frontier-adjacent performance meaningfully cheaper to run at scale, which matters most for teams running high-volume agentic workflows where token cost adds up fast.
For most businesses already using GPT-5.6 Sol or Luna, switching over is likely worth doing simply for the price cut, provided the compliance basics above are already in place. For businesses still evaluating which AI vendor to build on, our guide on finding the right AI language model is a good starting point, since this release is one more data point in a market where frontier labs are now competing on cost per task as much as raw capability.
Weighing whether GPT-6 Sol or Luna fits your stack, or need a second opinion on the compliance side before rolling either out? Our guide to how AI consulting helps businesses in Germany covers what that process typically looks like, or book a discovery call and we'll walk through your use case, data flows, and what actually needs sign-off before go-live.
Related service: AI Development
Sources
This piece draws on OpenAI's official announcement and reporting from TechCrunch, VentureBeat, The New Stack, MarkTechPost, and The Next Web.
Frequently Asked Questions
Is GPT-6 Terra available?
No. OpenAI released only Sol and Luna this round. As of now, there's no GPT-6 Terra in the lineup.
Are the new GPT-6 Sol and Luna prices permanent?
Yes. An OpenAI spokesperson confirmed these are standard rates going forward, not promotional or introductory pricing like GPT-5.6's launch prices were.
Can I self-host GPT-6 Sol or Luna?
No. Both are API-only models. There are no downloadable weights, so self-hosting isn't an option for either model.
Do GPT-6 Sol and Luna replace GPT-6 Astra?
No. Astra remains OpenAI's top model for the hardest, most complex tasks. Sol and Luna sit below it.
Does the EU AI Act apply to businesses using GPT-6 Sol or Luna?
Yes, for OpenAI as provider now. High-risk deployer duties depend on your specific use case and remain mostly deferred until 2027–2028.