Beyond GPT and Claude: Why the telecoms industry needs specialised AI models

The biggest AI model isn’t always the best one for the job. As AI usage scales and inference costs mount, could specialised models provide a more cost-effective, sovereign foundation for an AI-native telecom stack – and how can operators orchestrate the right model for each task?

As generative AI adoption has accelerated, telcos have largely been consumers of AI, using large language models developed by companies such as OpenAI, Anthropic and Google.

General-purpose AI models have come a long way, but they weren’t built specifically with telecoms in mind. Ask one to interpret an obscure industry standard or troubleshoot a complex network issue, and gaps in its domain knowledge can quickly become apparent.

Now, that is beginning to change with the launch of AT&T’s OTel 2.0, “the largest and best performing open-source model built for telecoms”, which has been post-trained on more than 400 billion telecom training tokens.

Built on Google’s Gemma 4 31B-IT, OTel 2.0 forms a key part of AT&T’s broader AI strategy. Its separate AI Gateway routes prompts to the model offering the best balance of cost, speed and quality for each task.

OTel 2.0’s training data includes standards, specifications and technical documentation from the GSMA, 3GPP, ETSI, O-RAN Alliance, TM Forum and ITU, along with contributions from industry partners. Alongside this, the GSMA and Pleias have released the Telco Corpus, a 10-billion-token openly licensed pre-training dataset for the telecoms industry.

Instead of repeatedly supplying large quantities of standards documentation as context, a telco-focused model already understands much of the shared domain knowledge. According to Andy Markus, AT&T’s Chief Data and AI Officer, “only a small percentage of the tasks we run require that level of sophistication. Many can be handled by lower-cost models without sacrificing performance.”

Many operators currently rely on general-purpose models such as ChatGPT or Claude, supplemented with Retrieval-Augmented Generation (RAG) to search internal documentation and standards whenever the model needs additional context.

This approach can work well, but every retrieval-based request requires relevant documentation to be fetched and interpreted, increasing latency, token consumption and inference costs. And as AI usage has scaled, total inference costs have risen sharply, making model choice and token efficiency increasingly important.

Aside from the potential cost savings, the telecoms domain is specialised enough to justify its own models, trained on network architectures, charging systems, BSS/OSS processes, TM Forum standards and industry terminology that general-purpose models may understand only superficially.

At the time of writing, the top two performers on Hugging Face’s Open Telco AI leaderboard are domain-adapted models: AT&T’s OTel-LLM-8.3B-QnA ranks first, followed by Chunghwa Telecom’s CHT-TeleX-8B. Both outperform frontier models including GPT-5 and Claude Opus on these telecom-specific benchmarks.

AT&T says it now processes an average of 45 billion AI tokens a day. At such a scale, inference becomes an operational cost that requires optimisation in the same way operators optimise network capacity or cloud infrastructure.

When AI moves from answering occasional questions to supporting day-to-day operations, consistency matters as much as raw capability. A specialised model does not necessarily know more in every respect; it is optimised to perform a narrower set of tasks repeatedly and efficiently.

However, going beyond GPT and Claude does not mean abandoning them. Frontier models will remain valuable for tasks requiring broader knowledge or deeper reasoning, while telecom-native models can handle repeatable, high-volume domain tasks. The objective is not to identify one universally superior model, but to orchestrate different models according to the task, balancing accuracy, cost, security and control.

An obvious question is: why would a model trained by AT&T benefit other operators?

OTel is trained primarily on industry knowledge rather than AT&T’s own operational data, such as telecom standards, network protocols and architectural principles that are already shared across the industry. Individual operators can then fine-tune the model using their own product catalogues, business processes, operational data and workflows.

It won’t automatically understand Vodafone’s product catalogue or Telstra’s charging rules, for example – that knowledge still needs to come from operator-specific data, integrations and governance.

Rather than depending entirely on proprietary AI platforms, operators can deploy telco models across public cloud, private cloud or on-premises infrastructure. This gives them greater sovereignty over how AI is deployed and governed, helping them maintain operational control, reduce dependence on individual AI providers and retain the flexibility to optimise costs over time.

At the same time, not every operator will want to deploy and manage these models itself. Instead of purchasing individual AI features or models, operators may increasingly consume AI as a managed service, with vendors orchestrating specialised models such as OTel alongside frontier models according to the complexity, cost and security requirements of each task.

This could create an opportunity for BSS/OSS vendors to monetise outcome-based AI services, with effective orchestration, metering and AI cost management critical to delivering them profitably.

There are, however, reasons for caution:

• The telecoms industry is not a monolith
  Standards may be shared, but implementations, regulations, languages and operational practices vary between markets. An English-language model trained primarily on published standards may therefore require significant regional and operator-specific adaptation.
• Knowledge changes rapidly
  3GPP releases, TM Forum ODA and Open APIs, vendor software and regulations evolve constantly. Maintaining a telecom model could become an ongoing engineering challenge.
• General models are improving fast
  New GPT, Claude and Gemini releases may narrow the domain knowledge gap through larger training corpora and better reasoning, reducing the advantage of specialised models over time.

Today’s frontier models are impressive generalists. But the deeper AI becomes embedded in BSS/OSS, the greater the value of a model that starts with telecom expertise rather than having that context supplied to it for every task. A telecom-native model could therefore handle high-volume operational workflows more effectively and economically, including:

• interpreting charging rules
• analysing mediation issues
• explaining billing anomalies
• supporting network configuration
• understanding TM Forum Open APIs and telecom standards

Rather than acting as generic copilots, AI agents could function as telecom specialists.

For example, a product manager could describe a new package in natural language and have an AI agent generate an initial product catalogue configuration, recommend charging rules, identify dependencies and highlight potential conflicts, before passing the proposed configuration to a human for review.

This aligns closely with the move towards agentic BSS, in which responsibilities are divided between agents that can draw on different models according to their domain and task.

An agent may determine that a customer should migrate to a new tariff, but any resulting action should remain subject to existing safeguards and controls. Agents should therefore execute changes through governed interfaces such as TM Forum Open APIs, allowing deterministic BSS/OSS systems to validate eligibility, pricing, authorisation, compliance and business rules before carrying out the transaction.

OTel is a significant step on the road to an open telco AI ecosystem combining specialised models, intelligent gateways and open standards. But its importance lies less in replacing frontier models than in giving operators a broader set of models that can be selected and orchestrated according to capability, cost, security and control.

The “AI-native telco” will not run on a single all-purpose LLM. It will depend on a governed orchestration layer through which specialised and frontier models can work with trusted data, open APIs and deterministic operational systems. For BSS/OSS vendors, the differentiator will not be who has the smartest model, but who can select and integrate the right models safely into the workflows that matter.

Visit the Cerillion AI Hub to explore how model orchestration, agentic AI and governed execution are reshaping BSS/OSS.