GPT-5.6 Codex (2026): Sol vs Terra vs Luna, Download & Run From Your Phone

To use GPT-5.6 with Codex, update the Codex CLI and select a GPT-5.6 tier — codex --model gpt-5.6-sol for the flagship, or Terra / Luna for cheaper runs. It went generally available in Codex on July 9, 2026. Below: which tier to pick for coding, three real commands, and how to steer a GPT-5.6 Codex agent from your phone with SeaWork while the code stays on your machine.

That’s the whole answer if you just needed the command. The rest of this is the honest version — what the three tiers actually trade off for coding, and why a model this capable at long, unattended work is exactly the kind you want on a leash you can hold from your pocket.

What actually shipped

OpenAI released GPT-5.6 as a three-tier family and made it generally available via the API and Codex on July 9, 2026 (OpenAI; MarkTechPost). The naming is worth understanding because it changes how you choose: the number (5.6) is the generation, and Sol, Terra, and Luna are durable capability tiers that can each advance on their own cadence.

• Sol — the flagship, for the hardest work. It reports 88.8% on Terminal-Bench 2.1, and a Sol Ultra mode reaches 91.9% by spawning coordinating subagents rather than just spending more compute (OpenAI; Vellum). Pricing: $5 / $30 per million tokens (input / output).
• Terra — roughly GPT-5.5-class quality at about half the price. Terminal-Bench 2.1 at 87.4%. Pricing: $2.50 / $15 per million tokens.
• Luna — the fastest and cheapest tier, 84.7% on Terminal-Bench 2.1. Pricing: $1 / $6 per million tokens.

For context, the same reporting puts GPT-5.5 at 85.6% and Claude Mythos 5 at 88.0% on that benchmark (Vellum). Two honest caveats: Terminal-Bench version numbers matter — these figures are on 2.1, so don’t compare them straight against an older run — and the 91.9% Ultra number comes from a subagent-spawning mode, not a single pass. Treat published benchmarks as a rough ranking, not gospel; your repo is the real benchmark.

The practical part: you get GPT-5.6 in Codex the same way you got 5.5 — same key, same quota. Pick the model in the Codex CLI or any client and go (OpenAI). For the subscription and credit side, see Codex pricing and plan limits.

Which tier should you actually use for coding?

This is the real question, and the answer is “not always Sol.” Here’s the first-principles way to think about it, stripped of the marketing.

A coding agent spends most of its tokens on routine mechanical work — reading files, running tests, applying diffs it’s already confident about — punctuated by a few genuinely hard reasoning moments (a subtle bug, an architecture call, a gnarly refactor). Paying flagship prices for the mechanical 90% is waste; underpowering the hard 10% is how you get a confident agent that’s confidently wrong.

So a sane default:

• Reach for Sol when the task is genuinely hard and open-ended: a cross-cutting refactor, a bug that’s resisted you, anything where one wrong turn costs an hour. The Ultra subagent mode is worth it on the biggest tasks where the agent needs to fan work out.
• Reach for Terra for the bulk of everyday coding — feature work, test-writing, migrations you understand. GPT-5.5-class quality at half the price is the sweet spot for tasks you’d trust 5.5 with anyway.
• Reach for Luna for high-volume, low-stakes churn: boilerplate, doc passes, mechanical renames, quick “explain this file” queries where speed and cost beat the last few points of capability.

The trap with any of them is the same trap every capable agent has: they’re most valuable on long, unattended runs, and long unattended runs are exactly when a wrong turn goes unwatched. Which is the whole reason to keep the control surface in your pocket.

Run GPT-5.6 in Codex

Update the Codex CLI, then select a GPT-5.6 tier. Check your exact CLI’s flag against codex --help, since model selection can be a flag or a config entry, but the shape is:

# Update Codex, then run a task on a GPT-5.6 tier
codex --model gpt-5.6-sol "refactor the API layer and keep tests green"

That’s fine when you’re at your desk. The moment you want to start a run and then walk away — which, for a model built to grind on long tasks, is most of the time — you want it on a leash you can hold remotely.

Steer a GPT-5.6 Codex agent from your phone

Here’s the first-principles argument for keeping the agent local and moving only the interface.

Codex on your laptop is already wired into your world: your credentials, your MCP servers, your actual files in your actual repo with your actual environment. That setup is correct and it took you time to build. The clean pattern isn’t to recreate it in a hosted environment — it’s to leave the agent where it already works and move the control surface to your phone.

That’s what SeaWork is built on. The SeaWork daemon runs on your development machine and drives the Codex agents already there. Your phone connects over an end-to-end encrypted relay that only forwards encrypted packets — it can’t read your prompts, diffs, or logs, because it never holds the keys. Your code, logs, and provider credentials stay on your machine. And if you’d rather not depend on any relay at all, the daemon is yours to self-host on your own network.

You don’t install a separate daemon — the desktop app ships with it. Three steps:

1. Install the desktop app. The SeaWork daemon is built in.
2. Pair your phone. Open the mobile app and scan the pairing code — that establishes the encrypted link.
3. Start a GPT-5.6 Codex agent from the phone, the desktop, or the terminal.

The CLI is the clearest way to show what’s happening. These are real commands:

# Start a Codex agent on a task (running GPT-5.6 in your Codex setup)
seawork run --provider codex "refactor the API layer"

# List your agents (running and finished)
seawork ls -a

# Follow an agent's logs live
seawork logs <id> -f

Start something like that, then pick it up on your phone. You’ll see every tool call the GPT-5.6 agent makes as it makes them, read the full conversation, and send a message mid-run to course-correct when it drifts. And when the agent hits an approval prompt — the “allow this command?” gate that would otherwise freeze a Sol run for forty minutes while you’re away from your desk — the request lands on your phone and you approve or reject it in seconds.

One app for Codex and Claude Code

There’s a second reason this matters with GPT-5.6. Tiers change the economics of mixing models — you might run Terra for the bulk of a feature and hand the one hard subtask to Sol, or reach for Claude Code on something else entirely in the same afternoon. The deeper model trade-offs are in GPT-5.6 vs Claude Fable 5. SeaWork runs Codex and Claude Code from a single app, side by side, and runs them in parallel: one agent refactoring the API layer, another writing its tests, a third drafting the migration notes — each in its own workspace, each visible at a glance so you can see what’s running, what’s waiting on you, and what finished.

For anyone splitting implementation, tests, and review across separate worktrees — which is exactly the workflow a subagent-spawning model like Sol Ultra encourages — a single terminal scrollback stops scaling fast. A phone that shows you every agent at once fits that far better.

Where to start

Running GPT-5.6 in Codex is the easy part — it’s the same key and quota, you just pick the tier. The harder, more useful part is not being chained to your desk while a Sol run grinds through a long task. If you want the agent to keep running where your code already lives, controlled from your pocket, across both Codex and Claude Code, that’s the corner we built for.

SeaWork is free. Download SeaWork, read more about Codex on SeaWork, or see how to use Codex from your phone and how to use Claude Code from your phone for the wider survey. If you’re steering a single session, Claude Code remote control covers the first-party path too.

Sources

• Previewing GPT-5.6 Sol — OpenAI
• OpenAI Releases GPT-5.6 (Sol, Terra, Luna) — MarkTechPost
• GPT-5.6 Sol vs Terra vs Luna: benchmarks explained — Vellum