← Back to feed

On July 14, OpenAI CEO Sam Altman posted on X that GPT-5.6 Sol had seen “insane” growth since its public launch and that the inference team had done “heroic work” to support demand — before adding that service hiccups were a real possibility. “We are going to move mountains to continue to scale, but it is possible there are some hiccups soon,” Altman wrote.

The warning is notable on two fronts. Altman made it voluntarily, before any reported outage. And the problem he named — inference capacity, not model quality — points to the gap between building a frontier model and operating one at the scale that puts it at the centre of developer and enterprise workflows.

The Specific Bottleneck

Altman’s language was precise about what is constrained. He did not flag quality issues, safety incidents, or cost problems. The concern is inference: the compute infrastructure required to run the model in production at the rate users are demanding it.

Training a model requires a fixed amount of compute executed once. Inference requires that compute to run continuously, in parallel, at whatever rate the user base demands, with low latency. For a model at the capability level of GPT-5.6 Sol — used for extended reasoning, multi-step coding tasks, and agentic workflows — inference is compute-intensive relative to the chatbot-tier workloads that earlier infrastructure was sized for.

OpenAI’s challenge is that Sol’s public launch happened on a compressed timeline. The model was developed in June, held back for several weeks at the request of the Trump administration while regulators reviewed its advanced cybersecurity capabilities, then released to broad public access. The demand curve that followed was not paced — it was a step function.

The Timing Context

GPT-5.6 Sol sits at the top of OpenAI’s current model family and is the highest-capability option available through ChatGPT, Codex, and the developer API. It competes directly with Claude Fable 5 and leads several important benchmarks including Terminal-Bench 2.1 (84.7%), Agents’ Last Exam (53.6%), and the AA Coding Agent Index. At that capability tier, users who activate it tend to run long-context, compute-heavy sessions.

The demand signal Altman described is consistent with what has been observed across the frontier tier more broadly. Fable 5 Max Effort costs $1.57 per task on LiveBench against Sol’s $0.59 — Sol is meaningfully cheaper for equivalent work, which amplifies adoption at the margin. Token consumption from frontier models has been growing at roughly 14% per month across enterprise accounts; at that rate, a newly launched flagship can hit its capacity envelope within weeks.

What “Hiccups” Actually Means

Altman did not specify what degradation users should expect. In practice, inference capacity constraints at OpenAI’s scale manifest as rate-limit tightening, slower response times under peak load, or temporary throttling of the highest-compute variants. The “xHigh” and “Ultra” modes for Sol — which run parallel subagent chains and consume significantly more compute per request — are the most likely targets for any capacity management.

The warning also has a secondary function. By flagging the possibility proactively, Altman sets expectations before users encounter problems, which is a better outcome than unexplained degradation in a model many organisations are now treating as production infrastructure.

The Structural Point

The underlying dynamic is consistent with every major model launch at the frontier since GPT-4. Training timelines have compressed. Release-to-demand curves have steepened. But inference infrastructure must be built before demand arrives, because scaling compute takes time that a demand spike does not provide. Every lab with a successful frontier launch faces the same sequence: model performs well, demand exceeds projections, capacity scramble follows.

What is different now is the magnitude. The user base for frontier AI in mid-2026 is substantially larger than it was at any prior launch. Enterprise adoption of GPT-5.5, Fable 5, and related models has created a base of heavyweight users whose sessions consume orders of magnitude more compute than a casual ChatGPT prompt. When those users migrate to a new flagship, the baseline demand is already large, and the growth multiplier compounds on top of it.

OpenAI’s inference team, as Altman described it, is doing heroic work. The question is whether heroic work is enough when the demand signal is insane.