The headline number is impressive: Llama 5 is training on roughly ten times the compute used for Muse Spark, which was Meta’s first major Llama 4-era model released in April. That‘s a massive leap in scale. The problem is that GPT-5.5 has been out for months. Matching GPT-5.5 in internal benchmarks is a milestone — but it’s a milestone that lands Meta in the same position it‘s been in for years. Playing catch-up.
OpenAI is already shipping GPT-5.6 today. Google’s Gemini 3.5 Pro is expected July 17. Anthropic‘s Fable 5 is back online. Meta’s internal goal for Llama 5 is to compete with these models, not to lead them. By the time Llama 5 releases, the frontier may have moved again.
Meta‘s AI leaders have openly discussed relying on external models from Google and OpenAI as a temporary bridge until Llama 5 catches up. Employees already use Anthropic’s Claude for internal coding, even though Meta builds its own models. The company‘s own developers don’t trust Llama to do the work. Llama 5 is a necessary step forward. But it‘s a step that should have been taken months ago. Meta is spending more to get less — or at least, to get it later.
The Code Model Gambit: Splitting the Product Line
Llama 5 won’t be a single model. According to the leaks, Meta is preparing a dedicated coding model — internally referred to as “Llama 5 Codex” — alongside the general-purpose flagship. It‘s being positioned as an Opus-level code specialist, trained on AI-generated data and led by Meta’s new superintelligence chief, Alexandr Wang.
This is a departure from the “one model fits all” approach. Meta is betting that the future of AI is specialized, not monolithic. A model that writes code better than a general-purpose model could capture a different slice of the developer market — and that slice is growing fast.
But there‘s a caveat: Meta’s own engineers are reportedly using Claude Code for development work, not Llama. The company has admitted that its internal code assistant relies on external models, including Anthropic‘s Claude, to get the job done. If Meta’s own developers don‘t trust their model for coding, it’s unclear how much traction a new code model will gain with external developers, regardless of its specs or positioning.
The Open vs. Closed Tension: What Happens to Llama‘s DNA?
The biggest question around Llama 5 isn’t performance. It‘s licensing. Llama has been open-source. That was its defining feature. Developers chose Llama not because it was the best, but because it was free and auditable.
The leaks suggest a shift. Meta’s new AI organization has reportedly considered a two-track strategy: open-source the less capable versions and keep the frontier models closed. This would align Meta more with OpenAI than with its own open-source legacy.
Alexandr Wang, who now leads Meta‘s AI efforts, has been described as someone who “wavers” on open-source. His background is at Scale AI, not in open-source advocacy. If Llama 5 goes closed, it loses the one thing that made Llama relevant. Developers will have little reason to choose it over GPT-5.6, Gemini, or Claude. If it stays open, it may not be the version that actually competes.
Timing the Window
Llama 5 is reportedly targeting an August preview and September release. That’s a two-month window. But in AI, two months is enough time for competitors to release, iterate, and move on.
By the time Llama 5 arrives, GPT-5.6 will have been out for almost two months. Gemini 3.5 Pro will have had its launch moment. The market may be ready for a new model — or it may be looking ahead to whatever comes next.
Meta’s history doesn’t help. Llama 4 was criticized heavily, and the negative reception reportedly affected morale internally. Some employees have expressed doubts about the company‘s AI culture, with one departing staffer describing a “culture of fear” and a lack of conviction in Meta’s AI mission. Llama 5 needs to be a reset, not just an iteration.
At the same time, Meta has begun charging for enterprise AI agents built on Llama 5, suggesting the company intends to monetize this generation, not just give it away. That‘s a departure from the “give the model away and sell the platform” strategy that worked for Llama 2 and 3. Whether that’s a shift toward sustainability or a sign that open-source isn‘t working is still unclear.
P.S. Meta is training Llama 5 on 10x more compute, but its own engineers are reportedly still using Claude Code to get work done. The gap isn’t just compute. It‘s trust. That’s not a problem GPUs can solve.
