OpenAI rolls out GPT-5.6 with multi-agent parallelism

If you have run multi-step coding agents lately, you know the pain: the model burns through tens of thousands of tokens wandering down rabbit holes, and the output is still half-baked.

OpenAI has officially moved its GPT-5.6 family from limited preview to general availability. The lineup splits into three distinct tiers: the flagship Sol, the balanced Terra for everyday tasks, and Luna, aimed strictly at cost efficiency. Alongside the models comes "ultra," a heavy-duty execution mode that coordinates four parallel agents out of the box to tackle massive workloads faster.

OpenAI followed the launch with aggressive price cuts. On July 30, it slashed Luna pricing by 80% and Terra by 20%. Then on August 21, it dropped Sol API and credit rates by over 20% for a three-month window.

Token efficiency over raw bragging rights

Rather than chasing empty benchmark highs, GPT-5.6 focuses squarely on performance per dollar.

On Agents' Last Exam, which evaluates long-horizon professional workflows across 55 fields, GPT-5.6 Sol hit a record score of 53.6, beating Claude Fable 5 by 13.1 points. Even dialled down to medium reasoning, Sol topped Fable 5 by 11.4 points at roughly one-fourth the estimated cost. The smaller models punch well above their weight too: both Terra and Luna outperform Fable 5 while running at about one-sixteenth of the cost.

Coding shows a similar story. Sol with max reasoning scored 80 on the Artificial Analysis Coding Agent Index, edging past Fable 5 while burning less than half the output tokens, cutting latency in half, and trimming cost by about one-third. Terra landed just above Fable 5, and Luna surpassed Opus 4.8.

A big chunk of those savings comes from Programmatic Tool Calling in the Responses API. Instead of routing raw tool output back through the neural net at every turn, GPT-5.6 writes and executes lightweight scripts to filter data on the fly. It monitors progress locally and only passes critical context back to the model, drastically shrinking token round-trips.

When a problem is simply too big for a single thread, ultra kicks in. By default, it orchestrates four parallel agents to explore alternatives and execute subtasks simultaneously. In evaluations like BrowseComp and SEC-Bench Pro, scaling to 16 agents pushed scores higher while significantly lowering overall latency.

Visual judgment and locked-down cyber capabilities

GPT-5.6 also takes a noticeable leap in understanding visual layouts and design rules.

Instead of just spitting out presentation slides or web templates that look slightly off, it infers entire design systems—including font hierarchies, spacing, and master slide conventions. Leveraging stronger computer-use skills, the model renders the interface, inspects it for visual bugs, and polishes the result before returning it to the user. On OSWorld 2.0, Sol posted a 62.6% score, overtaking Opus 4.8 while consuming 85% fewer output tokens.

In cybersecurity, Sol demonstrated a massive jump in capability. It achieved 73.5% on ExploitBench compared to GPT-5.5's 47.9%, and jumped to a 24.9% success rate on ExploitGym within a two-hour window.

Because of those offensive capabilities, OpenAI is keeping the strongest defensive tooling behind Daybreak's Trusted Access for Cyber program. To keep access to these cyber-focused models after September 1, individual users must activate advanced account security using hardware-backed passkeys, such as Yubico keys, or lose their elevated access.

Why it matters

For developers and teams building automated software pipelines, GPT-5.6 changes the economic math. By allowing models to write code that filters tool responses before hitting the context window, it tackles the biggest barrier to agentic AI: runaway token bills and unbearable latency. The fact that Luna beats previous-generation flagships at a fraction of the cost means small teams can now run robust agents without enterprise-tier budgets.

We think OpenAI's strategic shift toward token frugality is exactly the right move. Competing purely on raw model intelligence yields diminishing returns if running the agent bankrupts the customer. Packaging parallel execution natively into ultra also simplifies what used to require complex multi-agent orchestration frameworks.

That said, ultra's brute-force parallel approach carries its own risks: running four or sixteen agents simultaneously can easily cause surprise billing if left unmonitored. Furthermore, the mandatory requirement of hardware passkeys for cyber capabilities indicates that as frontier models acquire genuine destructive potential, the era of open, unrestricted access to the most capable tiers is quietly closing.