OpenAI’s Three-Tier Bet Goes Public
GPT-5.6 is not one model but a family of three, each priced and tuned for a different job. According to OpenAI’s preview announcement, Sol is the flagship tier, tuned for maximum reasoning effort and positioned as the strongest option for coding, research, and complex agentic work. Terra sits in the middle, built for high-volume enterprise tasks such as document analysis and customer support. Luna is the fastest and cheapest tier, meant for everyday drafting and summarization. The rollout completed within 24 hours across ChatGPT, ChatGPT Work, Codex, and the API, according to TestingCatalog’s launch coverage, a much faster distribution pattern than the staged rollouts OpenAI used for earlier GPT generations.
Pricing scales cleanly with capability. Per Simon Willison’s technical breakdown, Sol runs $5 per million input tokens and $30 per million output tokens; Terra costs $2.50/$15; Luna costs $1/$6. All three share a 1 million token context window and a 128,000 token maximum output, with a knowledge cutoff of February 16, 2026. On the Agents’ Last Exam benchmark, Sol scored 53.6 — 13.1 points ahead of Claude Fable 5’s adaptive-reasoning result — while the cheaper tiers matched or beat that rival at roughly one-sixteenth the cost, per Willison’s analysis. That combination of frontier benchmark performance and aggressively tiered pricing is the headline story most coverage focused on. The feature buried underneath it — ultra mode — is the one that actually changes how enterprises need to architect and budget for agentic work.
Ultra Mode: When One Prompt Becomes Four Workers
Ultra mode is not a fourth model tier; it is a coordination layer available on top of Sol. Instead of a single reasoning pass, ultra mode decomposes a task and spins up parallel subagents that work on different components simultaneously before a final synthesis step recombines their output. MarkTechPost’s coverage of the release describes this as embedding agent coordination directly into the core API through a new multi-agent tool, rather than shipping it as a bolt-on feature. By default, ultra runs four agents in parallel — a deliberate design choice that trades token efficiency for stronger results and faster wall-clock completion on demanding tasks, according to TestingCatalog’s reporting.
The performance gain is measurable: ultra mode lifted Terminal-Bench 2.1 results from 88.8% to 91.9%, a jump that would normally require a new base model rather than a routing change. OpenAI has also demonstrated the coordination layer scaling well past its four-agent default — Neowin’s rollout report notes that a specialized “multiagent v2” configuration coordinated up to 64 concurrent subagents in an internal research exercise, though that configuration has not shipped to customers. What did ship is more modest but still consequential: ultra mode is available in ChatGPT Work for Pro and Enterprise subscribers, in Codex for Plus-tier and higher accounts, and in the Responses API as a beta feature. Alongside ultra mode, OpenAI introduced programmatic tool calling — model-written JavaScript executed in an isolated V8 runtime with no network access — which early enterprise customers report cutting token usage by 38% to 63.5% on tool-heavy workflows, per MarkTechPost. That efficiency gain partially offsets ultra mode’s token cost, but only for teams that adopt both features together deliberately.
Cache economics shifted at the same time. Cache writes now bill at 1.25x the uncached rate, while cache reads keep a 90% discount and a 30-minute minimum lifespan, per Willison’s breakdown. For teams running ultra mode’s parallel subagents against a shared context, that write penalty compounds quickly — each subagent that populates its own cache slice pays the surcharge, even before any of the four workers produces usable output.
Advertisement
What Enterprise AI Leaders Should Do About Ultra Mode
Ultra mode’s four-way parallelism is easy to enable and easy to under-budget for. Before flipping it on for production workloads, engineering and finance leaders need to treat it as an architectural decision, not a checkbox.
1. Rebuild token budgets around a 4x multiplier, not a per-token rate card
Standard cost models assume one reasoning pass per request. Ultra mode’s default of four parallel subagents means a single ultra-mode call can consume roughly four times the tokens of an equivalent non-ultra Sol call, even before accounting for the synthesis step that recombines subagent output. Finance teams that budget AI spend off Sol’s advertised $5/$30 per-million rate without an ultra-mode multiplier will be blindsided by their first monthly invoice. Model the 88.8%-to-91.9% Terminal-Bench gain against the token cost directly — for tasks where the accuracy jump doesn’t change a downstream business decision, the marginal ultra-mode spend is pure overhead, not investment.
2. Gate ultra mode by task complexity, not by user tier
Because ultra mode ships restricted to Pro/Enterprise ChatGPT Work seats and Plus-and-above Codex accounts, the natural failure mode is enabling it for every user who happens to hold that subscription tier rather than for the specific tasks that justify the cost. Build an internal policy that routes only long-horizon, high-stakes tasks — complex refactors, multi-document synthesis, research aggregation — through ultra mode, and defaults everything else to standard Sol or Terra. A blanket “give power users ultra mode” policy multiplies the token bill across routine requests that never needed four parallel workers in the first place.
3. Instrument subagent-level observability before scaling adoption
A single ultra-mode response now contains the combined work of up to four independent reasoning threads plus a synthesis pass, which means a wrong or hallucinated answer could originate in any one of five places. Standard single-call logging won’t isolate which subagent introduced an error or where the synthesis step dropped a contradiction between two subagents’ outputs. Before rolling ultra mode into any workflow with real business consequences — contract review, financial analysis, customer-facing responses — instrument logging at the subagent level, not just the final output, so failures are traceable back to their source.
4. Renegotiate caching strategy around the new write penalty
The shift to a 1.25x surcharge on cache writes changes the math for any architecture that relies on shared context across parallel calls. Teams that previously treated aggressive caching as a free efficiency win now need to model whether ultra mode’s four concurrent subagents populating separate cache slices creates enough write overhead to erase the 90% read-side discount. For high-volume agentic pipelines, this may mean consolidating context into fewer, larger cache writes rather than letting each subagent cache independently — a design choice that has to happen at the orchestration layer, not left to default API behavior.
The Structural Lesson
Ultra mode is a preview of where the frontier AI competition is actually heading next. The benchmark race between labs — Sol’s 53.6 on Agents’ Last Exam versus Claude Fable 5’s adaptive-reasoning score — still generates the headlines, but the more consequential shift is that “intelligence” is increasingly being delivered as coordinated agent work rather than a single smarter model pass. That has a direct economic implication: pricing power is moving from the per-token rate card to the orchestration layer, where decisions about how many agents to spin up, how long to cache their shared context, and when to trigger synthesis determine the real cost of a task far more than the base model’s list price does.
For enterprises, this means vendor evaluation criteria need to expand. A model comparison sheet that only tracks benchmark scores and per-token pricing will miss the variable that actually drives total cost of ownership under an agentic workload: how aggressively a vendor’s orchestration defaults to parallelism, and how transparently that default is exposed and controllable. OpenAI made ultra mode’s four-agent default visible and documented; the next wave of competitive pressure will likely come from labs that either expose finer-grained control over agent count, or bury it entirely inside an opaque “smart routing” layer that is much harder to budget against.
Frequently Asked Questions
What is GPT-5.6’s ultra mode?
Ultra mode is a coordination feature on top of OpenAI’s Sol model that decomposes a task and runs four parallel subagents by default, each working on a different part of the problem before a synthesis step combines their output. It lifted Terminal-Bench 2.1 scores from 88.8% to 91.9%, but at a proportionally higher token cost than a standard single-pass request.
How much more does ultra mode cost than standard GPT-5.6 usage?
OpenAI has not published a separate ultra-mode rate card — it bills against Sol’s standard $5 input / $30 output per-million-token pricing, but because ultra mode runs roughly four parallel subagents plus a synthesis pass, a single ultra-mode request can consume several times the tokens of an equivalent non-ultra Sol call. Enterprises need to model this multiplier explicitly rather than assume list-price billing.
Is ultra mode available to all GPT-5.6 users?
No. Ultra mode is restricted to Pro and Enterprise subscribers in ChatGPT Work, Plus-tier and higher accounts in Codex, and is offered as a beta feature in the Responses API. It is not available on Terra or Luna, and OpenAI has demonstrated — but not shipped — a more powerful 64-subagent research configuration beyond the four-agent default.
