OpenAI's new flagship model already finds real bugs in Chromium and Firefox source code. It just won't chain them into a working exploit, at least not yet.
OpenAI wants you to believe its newest model can already sniff out the kind of browser bug that used to take a skilled human researcher weeks to find. It can, mostly. What it won't do yet, by design, is turn that bug into a working attack.
OpenAI released GPT-5.6 Sol on July 9, its most capable coding and cybersecurity model to date, alongside two smaller siblings, Terra and Luna. Sol costs $5 per million input tokens and $30 per million output tokens. Terra runs $2.50 and $15. Luna, the cheapest of the three, comes in at $1 and $6. That's the whole family, in one breath.
The launch wasn't simple. A limited preview went out on June 26, then stalled. The Trump administration asked OpenAI to restrict the rollout to a short list of vetted partners, citing national security concerns over the model's cyber capabilities, according to Axios and Fortune. OpenAI agreed, even as it said publicly such restrictions shouldn't become the norm, per TechCrunch. Clearance from the Commerce Department didn't arrive until July 8. The wider public got Sol one day later.
None of this happened in a vacuum. Weeks earlier, the government had forced Anthropic to pull back access to its Fable 5 model over similar concerns, jailbreak vulnerabilities that could hand cyber adversaries a shortcut to finding software flaws. Sol landed under the same shadow.
OpenAI ran Sol against real target codebases, Chromium and Firefox among them, to see how far an AI model could get toward an actual exploit. Sol found bugs. It found exploitation primitives, the raw building blocks an attacker would still need to assemble on their own. It did not produce a full working exploit chain in those evaluations. Under OpenAI's own Preparedness Framework, that lands Sol at High capability for both cybersecurity and biological and chemical risk, one notch below the framework's top tier, Critical.
The UK AI Security Institute ran its own red-teaming before launch and came back with a harder finding. It says it found universal jailbreaks in the cyber domain in every single testing round. Some cracked within hours. Those jailbreaks unlocked long-form agentic work on vulnerability discovery and exploit development. OpenAI says it reproduced the specific attacks the UK agency reported and mitigated them before release. The agency's own expectation, according to Fortune's reporting, is that similar jailbreaks will keep surfacing as red teams keep looking.
That's the tension sitting underneath this whole launch. A frontier model that stops just short of writing you a working browser exploit is still a frontier model that a competent jailbreak might push the rest of the way.
A pricing fight OpenAI can't avoid
Sol's $5 and $30 rates sit close to Anthropic's Claude Opus 4.8, priced around $5 and $25. That's not an accident. Moonshot AI's Kimi K3 launched July 16 at $3 and $15, undercutting Opus by roughly 40 percent, and Financial Times reporting has it beating Opus 4.8 on several mainstream coding benchmarks while still trailing Fable 5 overall. DeepSeek's V4 Flash is cheaper again, at a fraction of a cent per output token next to Sol.
Frankly, OpenAI isn't just racing Anthropic anymore. It's racing a wave of Chinese labs willing to sell frontier-adjacent coding performance at a discount, and that changes what a flagship price tag actually has to justify.
Sol's pitch is that it earns its premium on long, autonomous coding and security tasks, not on benchmark scores alone. Whether that holds depends on whether enterprises running agentic pipelines feel a real capability gap when they swap in Kimi K3 or even OpenAI's own cheaper Terra instead. OpenAI is betting they will. The next few months of enterprise contracts will say whether that bet paid off.
Washington isn't done here either. If a UK regulator can find a universal cyber jailbreak within hours of getting access, the next models in this race are going to ship under even tighter government terms, not looser ones.
Also read: SpaceX is reportedly chasing a Pentagon AI compute deal • AWS billing dashboard told a customer who spent 19 cents that they owed 2.5 billion dollars • Moonshot AI's Kimi K3 Beats Claude and GPT Rivals to Top a Coding Benchmark
