OpenAI is presenting model selection and token efficiency as central to building lower-cost AI agents. In a post from OpenAI Developers, the company describes examples from startups using GPT-5.6 models with different reasoning levels, tool-calling methods and routing strategies.
For document extraction, real estate startup Hypha AI reportedly retained 98% of GPT-5.5’s accuracy with GPT-5.6 Luna at one-eighteenth the cost. Financial research company Rogo AI used programmatic tool calling to retrieve filings and analyze data, matching its evaluation quality while using 21% fewer input tokens.
OpenAI also highlights a result involving GPT-5.6 Sol. With retained reasoning and compaction, the model improved from 13.3% to 38.3% on ARC-AGI-3 while using roughly six times fewer output tokens. Those figures come from OpenAI’s own examples and do not establish how the models perform across production workloads.
Routing models by task
Several responses to the post focused on routing routine work to less expensive models and reserving larger models for tasks requiring more judgment. One commenter described using a smaller model for extraction and escalating only ambiguous requests, while another recommended sending routine file edits to Luna and complex software architecture to Sol.
That approach aligns with OpenAI’s examples, although the company’s post does not provide production error rates, latency figures or cost per completed task. Those measures could produce a different comparison from token costs alone, particularly for agents that require retries or lengthy reasoning.
A user discussing a customer-service stack also claimed that GPT-5.6’s prompt caching makes cached Luna input 90% cheaper than standard input. The saving would depend on stable prompts and repeated workloads, so it would not apply equally to every deployment.
Performance comes with a price question
An accompanying “Artificial Analysis Coding Index” chart plots API cost against index score for GPT-5.6 Sol, Terra and Luna alongside GPT-5.5, Claude Fable 5, Claude Opus 4.8 and Gemini 3.1 Pro Preview. The plotted Sol series reaches an index score of about 80 at an API cost of roughly $2,400, while Luna reaches about 77 below $1,000.
Those figures prompted questions about whether peak coding performance is worth a substantially higher price. One response also criticized comparisons that give competing models a single plotted point rather than a full cost curve, arguing that a model maker’s own chart may favor its products. Another commenter questioned why Grok was absent from the comparison.
The post also drew reports that GPT-5.6 Sol has a 1-million-token context window in Codex. Separate usage data dated Aug. 12 lists 1,573 uses of codex-auto-review, compared with 190 for GPT-5.6 Luna and 23 for GPT-5.6 Sol, though the figures do not indicate task complexity or total cost.
OpenAI’s examples support a narrower conclusion: agent economics can improve when model choice, reasoning effort, tool calls and prompt reuse are tuned to the task. Whether GPT-5.6 delivers lower costs for a particular product will depend on completion rates, latency, retries and the share of requests that actually require Sol-level performance.
Source: OpenAI Developers on X
