OpenAI released the GPT-5.6 family on July 9, and the numbers tell a story that is easy to miss if you only glance at the benchmark table. The headline is not that GPT-5.6 Sol beats Claude Fable 5 on Agents’ Last Exam by 13.1 points. It is that the medium-reasoning variant beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. The story of GPT-5.6 is efficiency, not raw supremacy.
The family has three tiers. Sol is the flagship, designed for maximum intelligence. Terra is the balanced workhorse. Luna is the cost-optimized entry point. OpenAI frames the release around a single metric: performance per dollar. That framing is worth taking seriously.
On the Artificial Analysis Coding Agent Index, GPT-5.6 Sol with max reasoning scores 80, 2.8 points above Fable 5. It uses less than half the output tokens, takes less than half the time, and costs about one-third less. Terra performs just above Fable 5. Luna outperforms Opus 4.8. Each does so in roughly one-third of the time, with about half as many output tokens, and at approximately one-quarter the estimated cost.
These are not marginal gains. They are structural shifts in the economics of deploying frontier models at scale.
The agent multiplier
The most architecturally interesting feature in GPT-5.6 is ultra, a capability setting that coordinates multiple agents in parallel by default. Ultra uses four agents on demanding tasks, and OpenAI’s charts show that scaling to 16 agents on evaluations like BrowseComp and SEC-Bench Pro shifts the score-latency frontier upward and to the left. More agents means stronger results in less time, not just more compute spent on the same reasoning path.
This is a bet on parallelism as the primary lever for capability improvement at inference time. It mirrors what the research community has been exploring with test-time compute scaling, but OpenAI has productized it. Developers can build ultra-like experiences using the multi-agent beta in the Responses API. The implication is that the next frontier of model capability will come less from bigger pretraining runs and more from smarter orchestration of multiple model instances working in concert.
The partner quotes in the release bear this out. Simon Last, co-founder at Notion, says Terra and Luna “punch well above their price” and that many agents running GPT-5.5 perform just as well on Terra for half the cost and 16% fewer tokens. Scott Wu at Cognition calls GPT-5.6 a top-tier model that “combines strong coding-agent performance with very strong cost efficiency.” These are not generic endorsements. They are specific claims about the economics of production deployments.
Programmatic Tool Calling
OpenAI also introduced Programmatic Tool Calling in the Responses API. The mechanism lets the model write and run lightweight programs that coordinate tools, process intermediate results, and choose the next action without passing every tool response back through the model. Alex Wang at Rogo reports that it matched quality while using 24% fewer output tokens and completing tasks 28% faster on financial research workflows.
This is a direct response to the token waste problem that plagues agentic systems. Every round trip through the model costs tokens and latency. By filtering intermediate data inside a lightweight program, GPT-5.6 reduces the number of model calls needed for tool-heavy tasks. Angel Faus at Clio says Programmatic Tool Calling cuts prompt tokens by 38% with no quality loss on multi-step document analysis.
The pattern is consistent across the release: OpenAI is optimizing for the total cost of getting work done, not just for benchmark scores.
The knowledge work surface
GPT-5.6 also improves on tasks that do not map neatly to coding benchmarks. It sets new state-of-the-art results on BrowseComp at 92.2% and OSWorld 2.0 at 62.6%. On OSWorld, it surpasses Opus 4.8 while using 85% fewer output tokens. The model can inspect and refine rendered interfaces, not just generate code. It follows complex reference formats in presentations and spreadsheets more faithfully than its predecessor.
These capabilities matter for the enterprise use cases that OpenAI is clearly targeting. The partner list includes Ramp, Shopify, Cisco, Clio, and Balyasny Asset Management. The quotes emphasize reliability, cost efficiency, and the ability to carry context across long sessions. Ian Tracey at Ramp says GPT-5.6 “felt less like a chat assistant and more like an end-to-end technical operator.” Shane Moran at Shopify notes that it “consistently produced accurate line-linked GitHub references where prior models often missed.”
What it means for builders
The practical takeaway for anyone building on top of these models is that the cost floor just dropped. Luna nearly matches GPT-5.5’s peak performance at less than half the estimated cost. Terra surpasses it at a lower cost. The efficiency gains are large enough to change product economics for applications that rely on frequent model calls.
The agent parallelism in ultra opens up a design space that was previously the domain of custom orchestration layers. Developers can now treat multi-agent coordination as a model-level primitive rather than something they build themselves. The Programmatic Tool Calling feature similarly reduces the engineering overhead of building tool-using agents.
OpenAI has not released the full system card or the safety evaluation report yet. The deployment safety page is linked but the PDF may still be forthcoming. The release does claim “our most robust safeguards to date” with layered protections trained into the model combined with real-time checks and monitoring calibrated to trust and risk. The proof will be in the independent evaluations.
The broader pattern is clear. Frontier model releases are no longer about who can post the highest single number on a benchmark. They are about who can deliver the most useful work per dollar. GPT-5.6 makes that case more strongly than any OpenAI release to date. The question for competitors is whether they can match the efficiency curve, not just the capability ceiling.