OpenAI Unveils GPT-5.6 with Sol, Terra and Luna, Takes Aim at Anthropic

OpenAI announced on Thursday its new family of models, GPT-5.6, split into three versions — Sol, Terra and Luna — with a clear message: it wants to reclaim ground from Anthropic in the arena where the AI battle is now being fought hardest, coding and enterprise work. The launch lands in the same week that rivals like Meta and SpaceXAI have made their own moves, in a market increasingly saturated with announcements.

Three models for three budgets

The structural novelty of GPT-5.6 is that it isn't a single model but a tiered lineup by power and price. It's a strategy OpenAI and its rivals have been refining for a while: offer an expensive, highly capable model for demanding tasks and cheaper options for everyday volume.

• Sol is the flagship model, described by OpenAI itself as its "best coding model to date." It costs $5 per million input tokens and $30 per million output tokens (tokens are the units in which models process text; more tokens mean more compute and more cost).
• Terra is the mid-tier option, at $2.50 for input and $15 for output.
• Luna is the budget variant, at $1 for input and $6 for output.

All three are now available in ChatGPT, in Codex — OpenAI's coding environment — and through the developer API.

The central sales pitch is efficiency. CEO Sam Altman told CNBC that Sol is 54% more token-efficient than previous versions on coding tasks. That figure matters more than it might seem: in professional use of these models, spending scales with volume, and a reduction in token consumption translates directly into lower bills for companies weaving AI into their workflows.

The showdown with Anthropic

If there's one rival this launch is aimed at, it's Anthropic. The company, founded by former OpenAI members, has managed to position itself as the race's "friendly" competitor, heavily focused on corporate clients, and has been gaining share by leaning on that reputation.

OpenAI responds with benchmarks. It cites the Artificial Analysis Coding Agent Index, a benchmark that measures model performance as coding agents, to claim its new family beats Anthropic's across the board.

The most direct comparison is between Sol and Fable 5, Anthropic's recent flagship model. According to OpenAI, Sol "sets a new state of the art at 80 points, 2.8 above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about a third less." The company adds that Terra edges out Fable 5 and that Luna outperforms Opus 4.8.

A critical read is warranted here. Benchmarks a company publishes about its own products are, by definition, marketing material: favorable metrics get chosen and compared against whichever rival version suits best. The index cited is independent, which lends the figures some weight, but the real test will come when outside developers put the models through real-world tasks and see whether that edge holds up outside the lab.

Cybersecurity: the politically sensitive part

The most striking aspect of the announcement isn't coding — it's cybersecurity. OpenAI describes GPT-5.6 as its "strongest cybersecurity model to date," with "frontier performance using far fewer tokens."

The company frames these capabilities as defensive: threat modeling, code review and patching, and blue teaming — the exercise of simulating an attack on one's own systems to find weaknesses before real attackers do. Done well, these are tasks that can save security teams a lot of work.

The political wrinkle is worth noting: according to TechCrunch, the Trump administration had previously tried to restrict the model's deployment, apparently over fears of potential misuse. And there lies the underlying tension with these models: a tool capable of finding vulnerabilities to defend a system is, almost by definition, capable of finding them to attack one. The same skill that helps a security team patch its holes can help an attacker locate them. OpenAI insists on defensive use, but the line between the two is thin and hard to police once the model is in the hands of anyone with API access.

ChatGPT Work, the office assistant

Alongside the models, OpenAI unveiled ChatGPT Work, a tool designed as a workplace companion for enterprise teams. It runs on desktop, web and mobile, and is geared toward everyday administrative tasks: drafting documents, building spreadsheets and putting together presentations.

It's a move consistent with the industry's broader strategy. Office productivity is where generative AI promises the most immediate, measurable return, and it's also the ground where Microsoft's Copilot and Google's Gemini compete, both with the advantage of already having their office suites installed across millions of businesses. ChatGPT Work is entering contested territory, and its success will hinge less on the model's raw power than on how well it integrates with the tools companies already use.

What to watch from here

GPT-5.6 confirms two trends that will define the coming months. The first is that the AI war is being fought less on raw capability than on cost per task: whoever delivers the same performance while burning fewer tokens wins favor with enterprises, which are the ones footing the big bills. The second is that AI-assisted coding has become the main battlefield, with Sol, Fable and their peers going head-to-head.

For users and businesses, the practical takeaway is this: there are more options, and on paper, cheaper ones. But before taking the comparisons at face value, it's worth waiting for independent testing. And on the cybersecurity front, the regulatory debate that already surrounded this model isn't going away: it's the kind of capability that forces the question not just of what it can do, but who should be allowed to use it.