Ringg’s agents resolve up to 65% of routine customer inquiries without a human agent, according to OpenAI’s writeup on the deployment. That number is the headline. The more useful detail for anyone building voice or chat agents is how Ringg gets there: not one model doing everything, but a routing layer that assigns each task to a different OpenAI model.
The problem Ringg started from
Ringg builds a voice and chat agent platform for large consumer businesses in India. Per the OpenAI account, those companies faced rising call volume and fragmented manual systems — agents juggling separate tools to sell insurance or book appointments. Scaling meant hiring, which raised the cost and complexity of every interaction.
That is a familiar wall. Adding headcount to absorb volume does not fix the underlying workflow; it just makes each conversation more expensive.
Four models, four jobs
Ringg’s production stack splits work by task rather than picking one model for everything:
- GPT-4.1 handles most real-time voice and chat traffic.
- GPT-5.6 Luna stays in the stack for requests where its performance, latency, or price-performance profile fits better.
- GPT-5.6 Terra handles post-call analysis — summaries and sentiment classification.
- GPT-5.6 Sol supports evaluation, prompt improvement, and model-as-judge workflows.
When a customer speaks or types, Ringg’s orchestration system combines the input with agent instructions, conversation history, customer data, knowledge-base context, and available tools. The routing layer picks the model and configuration, and the output flows back through orchestration to whichever channel the customer used.
This is the pattern worth stealing. If you are choosing a single model for an agent that does real-time conversation and offline analysis, you are probably overpaying on one side or underperforming on the other. The earlier post on tiering models and caching context covers the same split from the infrastructure side.
Where the cost reduction actually came from
Migrating suitable real-time workloads from GPT-4.1 to GPT-5.6 reduced model costs by roughly 90%, per the OpenAI account. The same source notes that moving certain real-time workloads to GPT-5.6 Luna produced that approximate 90% reduction.
Two things are worth separating here. One is the model swap itself. The other is the routing decision about which workloads qualify for the cheaper path. The second is the harder engineering problem, and it is the one that determines whether the savings survive contact with production quality bars.
Ringg also manages context explicitly: when a conversation approaches roughly 80,000 tokens, the system creates a structured summary so the interaction can continue without resending the full history. That is a cost lever as much as a quality lever.
Evals before rollout, then traffic shifting
Ringg tests models against historical conversations and simulated customer flows before deployment. In one case, it compared GPT-5.6 Terra against Gemini 2.5 Flash for post-call analysis and moved summaries and sentiment classification to Terra. The supplied account reports up to 97% accuracy on common regional languages, including conversations that mix English with local-language phrases.
Models that pass offline testing get a small share of production traffic first. In production, the router monitors latency and endpoint health across regions and shifts traffic when an endpoint goes down or crosses a latency threshold.
What the customer numbers show
The supplied account lists several deployments. Policybazaar connects more than 57,000 customer requests, with 67% of calls handled without human intervention and average response time falling from 8–12 minutes to under 60 seconds. Practo reports an 85% first-call resolution rate, response times below three seconds, and operating costs down 70% versus its previous human-led workflow, with more than 1,000 appointment bookings daily. Groww resolves 72% of inbound IPO, futures, and options queries through self-service.
These are vendor-reported figures from a single platform’s customer base, so treat them as directional rather than a benchmark you can assume transfers to your domain.
The part that is still in progress
Ringg is developing browser agents using OpenAI’s computer-use capabilities for onboarding, KYC, IT troubleshooting, incident support, and claims processing. It is also building a context layer meant to carry information across channels, so a request that starts on voice, continues on WhatsApp, and finishes in a browser does not force the customer to repeat details.
That cross-channel context problem is unsolved in most stacks. If you are building agents today, the routing and summarization patterns above are available now; persistent cross-channel memory is the piece to watch.
The practical takeaway: before adding a second model, map your tasks by latency and quality requirement. The 90% cost reduction came from knowing which workloads could move, not from the move itself.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.