How to Switch to GPT-5.6 (Sol, Terra, Luna): Model IDs, the New max/ultra Reasoning, and the One Breaking Change

Read time: ~7 minutes. Key facts:

  • GPT-5.6 reached general availability on July 9, 2026 across ChatGPT, Codex, ChatGPT Work, and the OpenAI API.
  • It’s a three-model family: Sol (flagship), Terra (balanced), Luna (fast).
  • API model IDs: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna; the bare gpt-5.6 alias routes to Sol.
  • Sol holds GPT-5.5’s pricing — $5 / $30 per 1M input/output tokens — so switching the flagship is a same-price upgrade.
  • New max and ultra reasoning-effort settings, and one breaking change: prompt caching now uses prompt_cache_options.ttl; the old prompt_cache_retention field is deprecated for GPT-5.6+.

Sourcing note: the release, model IDs, pricing, reasoning-effort settings, and the prompt-caching change are drawn from OpenAI’s GPT-5.6 preview/GA announcement and Model Release Notes, cross-checked against the benchmark figures in our earlier GPT-5.6 Sol news brief (which cites the official blog). Where a newer API-surface detail is only reported by third parties, this guide flags it — confirm those against the live API docs. Links at the bottom.

If you build on OpenAI models, GPT-5.6 is the version you switch to now: it went generally available on July 9, 2026, the flagship keeps the same price as GPT-5.5, and for most codebases the migration is a one-line model-ID change. The nuance is which of the three models to pick and one caching field that changed. Here’s the whole switch.

1. What actually shipped

GPT-5.6 is a family of three, not a single model (from OpenAI’s announcement):

ModelTierInput ($/1M)Output ($/1M)Best for
SolFlagship$5.00$30.00Complex coding, reasoning, agentic work
TerraBalanced$2.50$15.00Daily work — GPT-5.5-class quality at half the price
LunaFast$1.00$6.00High-volume tasks, summarization, automation

The headline for switchers: Sol’s $5/$30 exactly matches GPT-5.5’s flagship pricing — you get the newer model with no price increase. And Terra gives you roughly GPT-5.5-class quality at half the flagship cost, which is the real cost-cutting move for everyday workloads.

On benchmarks, Sol reported 88.8% on Terminal-Bench 2.1 (standard) and 91.9% in Ultra mode, above Claude Mythos 5 (88.0%) and GPT-5.5 (83.4%). (Source: OpenAI official blog, via our GPT-5.6 Sol brief.)

2. The one-line switch (API)

For most API code, switching is just changing the model string:

from openai import OpenAI
client = OpenAI()

resp = client.responses.create(
    model="gpt-5.6-sol",           # was e.g. "gpt-5.5"
    input="Refactor this function and explain the change.",
)
print(resp.output_text)

The IDs you can pass:

  • gpt-5.6-sol — the flagship, explicit.
  • gpt-5.6-terra — balanced, half the flagship price.
  • gpt-5.6-luna — fast/cheap for high volume.
  • gpt-5.6 — the bare alias. It routes to Sol (the most capable and most expensive tier).

Production tip: don’t hardcode the bare gpt-5.6 alias in production. It resolves to Sol, so you can silently run every request at flagship price. Pin an explicit ID (gpt-5.6-terra for most workloads) so cost is predictable and a future alias change can’t move you.

3. Pick your tier — the actual switch decision

The switch isn’t “5.5 → 5.6,” it’s “5.5 → which 5.6.” A simple rule:

  • Were you on the GPT-5.5 flagship for hard problems? → Sol. Same price, better model. Straight upgrade.
  • Do you run high volume where cost dominates (classification, summarization, extraction, background jobs)? → Luna ($1/$6) or Terra ($2.50/$15). Terra is the sweet spot for “GPT-5.5-quality daily work, cheaper.”
  • Agentic / long-horizon coding? → Sol, and read §4 on reasoning effort — that’s where Sol pulls ahead.

Start by swapping your flagship calls to Sol (zero price change), then move your cost-sensitive routes down to Terra/Luna and measure quality. That two-step keeps quality steady while you capture the savings.

4. New: max and ultra reasoning effort

GPT-5.6 adds reasoning-effort levels above what GPT-5.5 exposed — this is the main new capability to adopt when you switch, not just a rename.

  • max effort — the “think longer” setting, giving the model the most time to reason on a single hard task. Per OpenAI, max can be toggled on by anyone with 5.6 access in ChatGPT Work or Codex.
  • ultra — a higher tier that (per OpenAI and multiple reports) coordinates several agents in parallel for the hardest problems, at correspondingly higher token cost. Availability is gated: Pro and Enterprise in ChatGPT Work, and Plus and above in Codex.

In the API, reasoning effort is set on the request (the family uses a reasoning object with an effort level). Reach for max on genuinely hard, single-thread problems where more deliberation pays off, and reserve ultra for the few tasks worth multi-agent cost. For everyday calls, leave effort at its normal level — max/ultra spend materially more tokens.

Confirm the exact reasoning.effort enum values and any ultra-mode request shape against the live API reference before wiring them into production — the exact parameter surface is still settling post-GA.

5. The one breaking change: prompt caching fields

This is the migration detail that will actually break code if you ignore it. GPT-5.6 changes how prompt caching is configured:

  • For GPT-5.6 and later, use prompt_cache_options.ttl to set the minimum cache lifetime.
  • The older prompt_cache_retention field is deprecated for these models.

If your GPT-5.5 integration set prompt_cache_retention, update it to the new prompt_cache_options.ttl when you move to GPT-5.6, or you’ll lose your intended caching behavior (and pay for re-processing you meant to cache). GPT-5.6 also moves toward more explicit, predictable caching with declared cache breakpoints — worth revisiting your caching strategy at the same time rather than porting the old one verbatim.

One more API-surface note, reported at GA (confirm in docs before relying on it): GPT-5.6 introduces programmatic tool calling where tool orchestration runs inside the model turn, described as Responses-API-only (not Chat Completions). If you use it, that’s a reason to be on the Responses API rather than Chat Completions — which is also OpenAI’s recommended surface for the newer capabilities.

6. Switching in ChatGPT and Codex (no code)

Not everyone switches via the API:

  • ChatGPT — GPT-5.6 rolled out on July 9. Free/Go tiers get Terra; Plus/Pro/Enterprise get all three. Pick the model from the model selector; choose Sol for hard tasks, Terra for daily use.
  • ChatGPT Work — Pro/Enterprise get Sol (and the gated ultra).
  • Codex — GPT-5.6 is in the model picker; Plus and above get Sol plus the Ultra beta. If you code in Codex, switching is just selecting gpt-5.6-sol and, for the hardest tasks, enabling the higher reasoning tier.

7. Should you switch?

Yes, for most people — and the reasons are unusually clean:

  • Flagship is a same-price upgrade. Sol costs exactly what GPT-5.5 did ($5/$30), with better benchmarks. There’s little reason to stay on 5.5 for flagship work.
  • You can cut cost without losing much. Terra ($2.50/$15) covers a lot of “daily driver” work at half the flagship price — a genuine savings lever GPT-5.5 didn’t give you as cleanly.
  • The migration is small. A model-ID change plus the one caching-field update (prompt_cache_retention → prompt_cache_options.ttl) covers most integrations.

The honest caveats: pin an explicit model ID (don’t ride the bare gpt-5.6 alias into flagship pricing), budget for max/ultra if you enable them (they spend more tokens), and confirm the newest API-surface bits (reasoning enum, programmatic tool calling) against the live docs, since the post-GA surface is still settling.

The takeaway

Switching to GPT-5.6 (GA July 9, 2026) is, for most codebases, a one-line model-ID change: pick gpt-5.6-sol for a same-price flagship upgrade, gpt-5.6-terra to cut daily-workload cost in half, or gpt-5.6-luna for high-volume jobs — and avoid the bare gpt-5.6 alias in production since it routes to Sol. Adopt the new max/ultra reasoning effort where hard problems justify the tokens, and handle the one breaking change — swap prompt_cache_retention for prompt_cache_options.ttl. Do the flagship swap first (zero price change), then push cost-sensitive routes down to Terra/Luna and measure.

For migrating between other frontier models, see how to switch to Claude Sonnet 5 and how to switch to Claude Opus 4.8; and if you run long agentic sessions, why long-running coding agents drift.

Sources

  • Previewing GPT-5.6 Sol — OpenAI — the GPT-5.6 family (Sol/Terra/Luna), flagship Sol pricing ($5/$30, flat vs GPT-5.5), Terminal-Bench 2.1 benchmarks (Sol 88.8% standard / 91.9% Ultra), the new max reasoning effort and gated ultra mode
  • OpenAI Model Release Notes — Help Center — July 9, 2026 GA across ChatGPT, Codex, ChatGPT Work, and API; model IDs gpt-5.6-sol / gpt-5.6-terra / gpt-5.6-luna and the gpt-5.6 alias routing to Sol; prompt caching change (prompt_cache_options.ttl sets minimum cache lifetime; prompt_cache_retention deprecated for GPT-5.6+); reasoning-effort availability (max for any 5.6 access; ultra gated to Pro/Enterprise in Work and Plus+ in Codex)
  • GPT-5.6 Sol news brief — 7minai — our earlier coverage of the June 27 preview with the official benchmark and pricing figures
  • Model IDs, pricing, and the caching change are confirmed across OpenAI’s announcement and release notes. Newer API-surface details (exact reasoning.effort enum, ultra request shape, programmatic tool calling being Responses-API-only) are reported post-GA and should be confirmed against the live OpenAI API reference before production use. Verified July 11, 2026.