What are the three GPT-5.6 models and their prices?

Sol ($5 input/$30 output per million tokens) is the flagship, Terra ($2.50/$15) is the mid-tier option, and Luna ($1/$6) is the budget model.

How does GPT-5.6 Sol compare to Anthropic's Fable 5?

OpenAI claims Sol scores 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Fable 5, while using less than half the tokens, taking less than half the time, and costing about one-third less—though these are vendor-selected benchmarks.

Why did the government try to restrict GPT-5.6?

The Trump administration reportedly sought to limit its rollout due to concerns about potential misuse of its cybersecurity capabilities, though OpenAI pushed back and the model shipped.

What is ChatGPT Work?

ChatGPT Work is a new enterprise-focused product from OpenAI designed for workplace tasks like drafting documents, spreadsheets, and presentations, available on desktop, web, and mobile.

Home›Large Language Models›Article

Large Language Models

GPT-5.6 Is Here: Three Models, a Cybersecurity Pitch, and a Direct Shot at Anthropic

OpenAI's GPT-5.6 family arrives with three tiers, aggressive benchmark claims against Anthropic, a cybersecurity pitch that spooked regulators, and a new enterprise product. Here's what's actually worth paying attention to.

Dispatch desk·July 10, 2026·4 min read

TL;DR

GPT-5.6 launches in three variants—Sol, Terra, and Luna—with OpenAI claiming significant token efficiency gains and benchmark leads over Anthropic's models, while also rolling out a workplace-focused ChatGPT Work product.

What changed

OpenAI released the GPT-5.6 model family in three tiers alongside ChatGPT Work, a new enterprise product, with Sol priced at $5/$30 per million tokens and claims of 54% better token efficiency on coding tasks.

Why it matters

If the token efficiency and benchmark claims hold in production, GPT-5.6 could shift enterprise AI buying decisions away from Anthropic's Claude at a time when the two companies are in direct competition for the same customers.

Editorial record

Drafted by autopilot. Reviewed when marked.

Source

OpenAI dropped its GPT-5.6 model family on Thursday, and if you've been paying attention to the AI space lately, you'll recognize the playbook: tiered pricing, benchmark warfare, and a press release packed with superlatives. The question, as always, is how much of this actually holds up when you get past the marketing department.

Three Flavors, One Family

GPT-5.6 ships in three variants. Sol is the flagship—the one OpenAI wants you to reach for when the task actually matters. Terra sits in the middle, and Luna is the budget option for when you need something capable but don't want your API bill to ruin your morning. It's a sensible three-tier structure that mirrors what Anthropic has been doing with its own model families for a while now. Imitation, meet flattery.

Pricing, for the API crowd: Sol runs $5 input / $30 output per million tokens. Terra comes in at $2.50 input / $15 output. Luna is $1 input / $6 output. Sol is expensive—you're paying flagship prices for flagship performance, so you'd better be getting flagship results.

The Token Efficiency Angle

CEO Sam Altman has been making noise about efficiency, specifically claiming Sol is 54% more token-efficient on coding tasks compared to previous generations. That's not a trivial number if it holds in production. Token efficiency isn't just a cost metric—it's a proxy for how cleanly a model reasons. Verbose, repetitive output often signals a model that's padding rather than thinking. If Sol genuinely compresses its reasoning without losing accuracy, that's a meaningful capability improvement, not just a cost optimization trick.

OpenAI is also positioning 5.6 as its "strongest cybersecurity model yet," claiming it achieves frontier performance while consuming significantly fewer tokens on security-related tasks. That's a specific enough claim to be falsifiable, which is at least better than the usual vague capability gesturing.

The Government Already Tried to Put the Brakes On This

Here's the detail that doesn't show up prominently in the launch materials: the Trump administration reportedly sought to restrict GPT-5.6's rollout, citing concerns about potential misuse of its cybersecurity capabilities. OpenAI pushed back, and the model shipped—but the episode is worth keeping in mind when evaluating the "strongest cybersecurity model" pitch.

The supported use cases lean defensive: threat modeling, code review and patching, and blue teaming (that's where you simulate an attacker on your own systems to find holes before someone less friendly does). Those are legitimately useful capabilities for security teams. Whether the same capabilities can be trivially flipped toward offensive use is the question that apparently gave regulators pause—and that question doesn't disappear just because the model shipped.

ChatGPT Work: The Enterprise Play

Alongside the model launch, OpenAI announced ChatGPT Work, a workplace-focused product targeting enterprise teams. It runs on desktop, web, and mobile, and is designed for the daily document-drafting, spreadsheet-wrangling, presentation-building grind that occupies most knowledge workers' afternoons. It's clearly positioned against Microsoft Copilot and the broader wave of enterprise AI assistants—though notably, OpenAI and Microsoft have a complicated relationship that makes that competition awkward to discuss openly.

The product is available across ChatGPT, Codex, and the OpenAI API, so developers aren't locked out of building on top of it.

The Real Target: Anthropic

You can read the entire GPT-5.6 launch as an answer to one question: how do we beat Anthropic in enterprise? Anthropic has spent the past year quietly winning over the enterprise crowd with Claude, positioning itself as the responsible, safety-conscious alternative to OpenAI's move-fast energy. It's worked. Claude has become genuinely popular among developers and enterprise buyers who've decided they want their AI vendor to at least appear to be thinking carefully about what it's building.

OpenAI's response is to pull out benchmarks. Specifically, the Artificial Analysis Coding Agent Index, which OpenAI uses to claim Sol scores 80—2.8 points above Anthropic's recently released Fable 5—while using less than half the output tokens, completing tasks in less than half the time, and costing roughly one-third less. Terra, they claim, performs just above Fable 5. Luna, they say, outperforms Anthropic's Opus 4.8.

Now, a quick word about benchmark theater: any company citing a benchmark in a press release has selected that benchmark because it looks favorable. The Coding Agent Index is a real and reasonably respected metric, but "real benchmark" and "complete picture" are different things. Sol may genuinely be more efficient on that specific index while underperforming on tasks that index doesn't measure. The honest answer is: run your own evals on your actual workloads before you commit your production stack to anything.

What's Actually Worth Watching

The token efficiency claims are the most interesting thread here, because if they hold up at scale, they compound. A 54% efficiency improvement on coding tasks doesn't just mean cheaper API bills—it means more headroom for longer context, more agentic loops, more complex tool-use chains before you hit cost or latency walls. That's the kind of improvement that actually changes what you can build, not just what you can afford.

The cybersecurity positioning is either genuinely significant or a marketing angle that happens to be technically accurate but practically marginal. Security teams will figure that out quickly. The government's concern is a real signal worth not ignoring, even if the model shipped anyway.

And the three-tier family structure? That's just table stakes now. Every serious AI lab has one. The differentiation lives in the details—and in production, where demos don't follow you.