GPT-5.6 Sol Ultra Hit Codex. What It Costs You

TL;DR

– GPT-5.6 Sol Ultra landed at 91.9% on Terminal-Bench 2.1. That’s above plain Sol at 88.8% and GPT-5.5 at 88.0%. The win comes from spinning up parallel subagents.
– Token pricing ranges from $5 in / $30 out per million for Sol down to $1 in / $6 out for Luna, the budget tier.
– OpenAI has shipped Sol to partners. Wider rollout? “The coming weeks.”
– Independent evaluator METR said the model isn’t really beyond state of the art for software and R&D tasks.
– Cerebras support is expected to deliver 750 tokens/second for selected customers.

GPT-5.6 Sol Ultra is live inside Codex. That’s OpenAI’s agentic coding platform, and Sol Ultra just posted a 91.9% on Terminal-Bench 2.1. A test built around command-line workflows that need planning, tool coordination, and iteration. Previous gen GPT-5.5 sat at 88.0%. Plain Sol grabbed 88.8%. Ultra’s edge?

It spawns subagents to split the workload.

But here’s the catch.

Each subagent burns its own tokens. At $30 per million output tokens for Sol, that adds up brutally.

How Ultra Changes the Architecture

Ultra mode isn’t a different model.

Can’t stress that enough.

It’s a reasoning configuration layered on top of GPT-5.6 Sol. OpenAI added two new reasoning tiers with this release — `max`, which just gives Sol more thinking time. Then `ultra`, which deploys subagents to parallelize complex work.

Regular Sol? One agent, sequential, grinds through a problem step by step.

Sol Ultra delegates.

Chunks the problem. Sends pieces out to other agents simultaneously. For multi-file refactoring, debugging chains that span systems, or cross-service coordination. That parallelism genuinely helps.

But look at the actual numbers. Plain Sol hit 88.8% on Terminal-Bench 2.1. Ultra reached 91.9%. That’s a 3.1-point gap. Meanwhile, the jump from GPT-5.5 to plain Sol? Less than a single point.

OpenAI is throwing serious engineering at single-digit benchmark improvements.

Tbh those gains are real though.

Sol beat publicly launched Claude models and Gemini 3.1 Pro on Terminal-Bench 2.1, including heavyweights like Claude Opus 4.8, Claude Fable 5, and Claude Mythos 5. But OpenAI’s own system card admits METR found the model not meaningfully past the state of the art on software and R&D. The wins exist.

They’re thin. And thin margins on a benchmark don’t tell you what happens when your invoice arrives.

Breaking Down the Token Bill

Three pricing tiers for the GPT-5.6 family, all per million tokens:

– Sol: $5 input / $30 output
– Terra: $2.50 input / $15 output
– Luna: $1 input / $6 output

Sol sits at the top.

Terra handles the middle. Balanced, cheaper.

Luna’s the budget pick.

Fastest and least expensive.

Those are standard Sol rates though. Ultra mode multiplies the problem. Every subagent Ultra dispatches runs its own token consumption. Input and output, separately. Say Ultra launches three subagents for one task. You’re now paying for four agents on that single task, not one. At $30 per million output tokens, anything beyond well-scoped, high-value work gets expensive fast. Honestly, you need to understand the cost before you hand something to Ultra, not when the bill shows up.

Side note: OpenAI burned over 700,000 A100-equivalent GPU hours trying to jailbreak this model during testing.

Didn’t work. That’s a staggering amount of compute just for red-teaming.

Tells you something about what’s packed into this generation.

Terra vs. Ultra — Where Should You Build?

Running a small dev shop or solo practice? The question isn’t whether Sol Ultra is the strongest coding model out there.

On Terminal-Bench 2.1, yeah, it is. The real question: does that marginal lift justify the cost multiplier over cheaper tiers?

Here’s my take.

Default to Terra for everyday coding. You’re paying $2.50 input and $15 output per million tokens. That’s GPT-5.6-class reasoning at literally half of Sol’s price. Save Sol and Sol Ultra for the genuinely hard problems where parallel subagent dispatch earns its keep. Multi-file refactors. Deep debugging chains. Cross-service integration work. Those jobs justify the premium.

Speed matters here too. OpenAI says Sol will hit Cerebras at up to 750 tokens per second for select customers. When your coding agent iterates through multi-step workflows, that kind of throughput crushes wait times. Less time waiting on inference means the effective cost per finished task drops. Even if the per-token price doesn’t budge.

Shipping With Ultra — Practical Guardrails

OpenAI has shipped GPT-5.6 Sol to trusted partners. Broader availability through ChatGPT, Codex. And the API is coming in “the coming weeks.” If you ship code for clients, three things should be on your radar.

Start with scope.

Sol Ultra in Codex means your coding agent can parallelize what used to be sequential work. Test it on tightly scoped tasks where parallelism helps. Multi-file refactoring, structured debugging. Don’t throw open-ended exploration at it. Subagent token costs spiral when the task has no boundary. And you won’t catch it until the usage report lands.

Token tracking from day one. Not optional.

Ultra is the most expensive reasoning configuration OpenAI offers. If you bill clients for AI-assisted development and aren’t measuring per-task token consumption, you’re eating the delta between what you charge and what OpenAI charges you.

That gap widens every time Ultra spins up another subagent.

And the safety angle.

Don’t dismiss it. METR said the model isn’t significantly beyond state of the art on software and R&D. That’s not a reason to skip the model. It’s a reason to set hard boundaries before you let Ultra-mode agents touch anything important. Filesystem scope. Deployment permissions. External API access. Lock those down before the first subagent runs, not after something breaks.

Already working inside Codex? Put Sol Ultra on your hardest coding tasks this month. Price-sensitive? Build on Terra and Luna first, escalate to Ultra only when cheaper tiers can’t handle the job. The gap between Sol and Terra on capability is narrower than the gap on price.

Your wallet’s gonna notice before your code does.