What OpenAI launched: GPT-5.6 Sol, Terra and Luna
On June 26, 2026, OpenAI previewed the GPT-5.6 series: three models designed to cover every performance/cost ratio.
Sol is the flagship, the most capable of the lineup. Terra is the balanced model for everyday work, with performance close to the previous generation at roughly half the cost. Luna is the fastest and cheapest, built for high volumes.
The launch wasn't a classic public release: initial access was limited to 20 trusted partners, with coordination with the U.S. government and a reinforced "safety stack." General availability is expected in the following weeks. Even the way it shipped says something: frontier models are now treated as sensitive technology — a theme we also covered in the Fable and Mythos case.
Where GPT-5.6 is aimed: coding and cybersecurity
The two areas where OpenAI pushed hardest with GPT-5.6 are coding and security.
On coding, Sol sets a new state of the art on Terminal-Bench 2.1, the benchmark that measures command-line workflows: planning, iteration and tool coordination. It's the sign of a model built for development agents, not just for code completion.
On cybersecurity, Sol is OpenAI's most capable model ever, including the discovery and exploitation of vulnerabilities. It's precisely this power that drove the launch restrictions. Terra and Luna, below the flagship, round out the scale with an eye on cost for high-volume tasks.
Where Claude stays strong
Claude — with Opus 4.8 and the 4.x family — holds on to its strengths, which aren't the same ones GPT-5.6 leans into.
Quality of reasoning on complex tasks, analysis of long documents thanks to its large context window, and structured writing remain areas where Claude is among the best on the market. On top of that sits a mature enterprise ecosystem: Model Context Protocol for integrations, the Agent SDK for agents, and Claude Enterprise with SSO and audit logs.
Then there's a distinctive trait in its approach: safety built into the model (Constitutional AI) and the guarantee that it won't train on your business data. To find your way among the Claude models, see Opus, Sonnet and Haiku compared.
GPT-5.6 or Claude for your business?
30 minutes to discuss your specific case.
Pricing: OpenAI's new grid vs the Claude scale
GPT-5.6's announced prices are aggressive. Per million tokens: Sol $5 in and $30 out, Terra $2.50 / $15, Luna $1 / $6. It's a scale designed to route each task to the right model based on cost.
Claude covers the same logic with Haiku (low-cost), Sonnet (balanced) and Opus (top of the range). The exact comparison depends on the input/output mix and the use case, so a blunt "which is cheaper" is misleading. The right yardstick is still cost per result, not price per token: on a complex task, a model that solves it well on the first try costs less than a cheap one that needs three attempts and a review. For the Claude numbers, see how much Claude costs for businesses.
The availability and sovereignty question
There's a practical difference the benchmarks don't capture: availability.
At the time of the announcement, GPT-5.6 is in preview for 20 partners, not for everyone, and comes with usage restrictions tied to cyber risks. For a company that has to decide and get moving today, that's a real constraint.
Claude, by contrast, is available and deployable right now: via API, Claude Enterprise, or on European regions through Amazon Bedrock and Google Vertex for those with data sovereignty requirements — an increasingly decisive topic, from GDPR to Switzerland with the revFADP. When compliance is at stake, availability and guarantees count as much as a benchmark score.
Which to choose for your business
Choose GPT-5.6, once it's available, if: your priority use cases are advanced coding or cybersecurity, or if you're already in the OpenAI/Microsoft ecosystem and want to take advantage of the Sol/Terra/Luna scale.
Choose Claude if: you need deep reasoning and document analysis, a mature enterprise and agent ecosystem that's available today, and guarantees around data sovereignty and non-training.
In practice, many companies will keep both and route each task to the best-suited model. The right question isn't "who wins," but "which model for which job." If you want to figure out what makes sense in your context, let's talk.