OpenAI moved the frontier model fight onto cost-per-result territory on July 9, announcing in its newsroom the general availability of GPT-5.6, a three-tier family led by flagship Sol alongside Terra for everyday work and Luna as the budget option. The models rolled out across ChatGPT, Codex and the API within 24 hours, priced at $5 input and $30 output per million tokens for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna.
The pitch is efficiency as much as raw capability. OpenAI says Sol sets a new high of 53.6 on Agents' Last Exam, a benchmark of long-running professional workflows across 55 fields, beating Anthropic's Claude Fable 5 by 13.1 points, and claims Terra and Luna outperform that rival "at around one-sixteenth the cost." On the independent Artificial Analysis Coding Agent Index, Sol posts a record 80 while using less than half the output tokens of its closest competitor. The company's framing is blunt: "more successful work for the same spend, or comparable results at a lower total cost."
Two structural changes matter more than the benchmark table. First, a new ultra setting coordinates four agents in parallel by default, with 16-agent configurations tested, trading token spend for faster results on demanding tasks; developers get the same pattern through a multi-agent beta in the Responses API. Second, Programmatic Tool Calling lets the model write and run small programs that orchestrate tools and filter intermediate data instead of routing every tool response back through the model. Early customers report the gains are real: Notion says agents on Terra match GPT-5.5 at half the cost, and legal software firm Clio measured 38 percent fewer prompt tokens on document analysis.
There is a harder edge too. GPT-5.6 nearly doubles its predecessor's exploit-development scores, so OpenAI is gating its strongest cyber capability behind a Trusted Access program, requiring hardware-backed passkeys by September 1, and says its new safeguards block roughly ten times more potentially harmful activity, backed by some 700,000 GPU hours of automated red teaming.
For the Gulf, the economics are the story. OpenAI's compute buildout in Abu Dhabi with G42 under the Stargate UAE program means this generation of models will increasingly run on regional infrastructure, and the tiered pricing changes what is viable here: Luna at $1 per million input tokens makes Arabic-language customer service, document processing and agentic back-office automation affordable at bank and telco scale, not just at pilot scale.
The practical read: builders should benchmark Terra and Luna against whatever they run today, because the cost-per-task math has moved. Enterprises buying AI transformation programs should reprice vendor quotes built on last quarter's token economics. And rivals selling on capability alone should worry; OpenAI just made efficiency the battleground, and per-dollar performance is a metric procurement teams understand.