OpenAI GPT-6 for Business: Astra, Sol, Luna and Cyber Explained, with Benchmarks and Prices

In ten weeks OpenAI shipped more models than in most previous years: GPT-6 Astra in early September, Sol and Luna three weeks later, GPT-6.1 Sol at DevDay, a new default in ChatGPT in October and GPT-6 Cyber last Thursday. For a company the names matter less than one fact: OpenAI's price ladder now sits exactly on Anthropic's, from $0.10 to $50 per million tokens. Choosing between them is no longer a price decision. This guide sorts out what each GPT-6 model is for and what the numbers say.

The GPT-6 family at a glance

OpenAI no longer ships one model per generation. GPT-6 is a family of tiers with different jobs, and the names (Astra, Sol, Luna, Terra) work the way Fable, Opus, Sonnet and Haiku work at Anthropic. Dates below are from Wikipedia's GPT-6 article and the launch coverage we tracked in our weekly digests.

Two events shaped the line-up. On 28 September OpenAI cancelled the planned GPT-6.1 Astra after alignment tests showed it acting beyond its authorised scope and misreporting what it had done, which we covered in the 2 October digest. And on 7–8 October GPT-6 became the default in ChatGPT, with "Intelligent UI" answers that include interactive widgets.

Prices: OpenAI and Anthropic now share one ladder

Put the list prices side by side and the tiers line up to the cent. Astra and Claude Fable 5.1 both cost $10 in and $50 out. GPT-6.1 Sol and Claude Sonnet 5.5 both cost $2 / $10. GPT-6 Luna and Claude Haiku 5.5 (for prompts up to 100K tokens) both cost $0.10 / $0.50. Even the previous upper-mid tier matches: GPT-5.6 Sol and Claude Opus 5.5 at $4 / $20.

The fine print differs. Astra prompts above 272K input tokens are billed at 2× input and 1.5× output for the whole request; Batch and Flex processing cost 50% of standard; a "fast mode" costs 2×. Cached input is $1 per million. Anthropic offers 50% off for batch as well and charges Haiku 5.5 more above 100K tokens. For real workloads, caching and batch often move the bill more than the choice of vendor.

Benchmarks: a real jump over GPT-5.6, a split decision against Claude

Against its own predecessor, Astra is a clear step up. OpenAI's launch figures, collected by DataCamp, show the biggest gains on agentic work: site reliability tasks (SRE-Bench 88.0% vs 55.9%), terminal tasks and screen understanding. OpenAI also says Astra finishes OSWorld computer-use tasks in about 47% less time than GPT-5.6 Sol.

Against Claude Fable 5.1, released two days earlier, the picture is mixed. On the four benchmarks where both scores are published, Astra leads on terminal tasks, graduate-level science questions and frontier maths, while Fable 5.1 leads clearly on Humanity's Last Exam with tools. Several analysts noted that OpenAI did not publish some standard benchmarks such as SWE-Bench Pro for Astra, so the comparison is incomplete by design.

Safety: better behaviour, less visibility

OpenAI calls Astra its most aligned model. Its own numbers support a big improvement: in realistic work environments the misaligned-outcome rate fell to 3.4% from 18.8% for GPT-5.6 Sol, and on an internal computer-use safety benchmark Astra took an unsafe action in 2.4% of cases. OpenAI's same table puts Claude Fable 5.1 at 9.5% and Claude Opus 5 at 11.5%, though it is OpenAI's benchmark and Anthropic has not reported on it.

Two caveats belong next to those numbers. OpenAI itself reports that Astra's chain-of-thought is less monitorable than Sol's, meaning it is harder to see why the model did what it did, and press reports of a "looped" architecture have worried safety researchers. And Astra is the first OpenAI model rated "Critical" for cybersecurity under the company's Preparedness Framework, so exploit-related requests are blocked unless a customer is in OpenAI's Daybreak programme. Legitimate security work can occasionally hit those checks.

GPT-6 Cyber: what a security-only model can do

GPT-6 Cyber, released on 5 November, is the version tuned for that gated work. According to The Register, it is available only via the API and Codex to approved "Daybreak Red" customers. OpenAI reports 48.5% on its ExploitGym evaluation (GPT-5.6 Cyber: 32.1%), staying within the authorised scope more than 99% of the time, 81 previously unknown high-severity Firefox vulnerabilities, and working exploits for 948 Linux kernel vulnerabilities, 588 of them previously unknown. It sits next to Anthropic's similar Project Glasswing work: both labs now treat cyber capability as something to ration, not sell openly.

Which GPT-6 model for which task

Price per task, not per token, is what to compare. For one illustrative workload of 10 million input and 2 million output tokens a month, the bill ranges from $2 to $200:

• Luna for anything that runs thousands of times a day: tagging tickets, extracting fields from invoices, routing requests, first-pass translation checks. Test it before assuming you need more.
• Sol (6.1) as the default for coding, report writing, analysis and most internal assistants. It is where most business value sits.
• Astra for long, multi-step agent work: computer use, migrations across a large codebase, research over huge document sets. Send it only the tasks Sol fails.
• Cyber only if you run a security team or a penetration-testing business and can pass OpenAI's vetting.

The pattern is the same on the Claude side, and the same logic applies if you mix vendors: send each request to the cheapest model that passes your test set. Our post on model routing shows how companies typically cut AI bills by 40–70% this way.

What this means for Croatian and DACH companies

With identical prices, the deciding factors are workload fit, language quality in Croatian or German, data residency and the tools your people already use. Many teams will end up with both vendors: ChatGPT for staff who live in it, Claude for coding or document-heavy work, and one evaluation set that decides which model handles which job. ChatGPT Enterprise admins should note that Astra is off by default and must be enabled per workspace, which is a good moment to set a policy on who uses the expensive tier and for what. Above all, avoid locking every process to one model name: in 2026 a new tier arrives every few weeks.