GPT-5.6 Goes GA: Built for Fan-Out, Billed for Cache Writes

OpenAI moved GPT-5.6 out of limited preview on July 9 and made it generally available across ChatGPT, Codex and the API in three tiers: Sol, Terra and Luna. GitHub started rolling the same models out in Copilot the same day. As someone who runs several coding agents a day, I care less about the benchmarks than two quieter changes: the release is built for running many agents at once, and it changes how prompt caching is billed.

Three Tiers, Three Price Points#

Launch pricing per 1M tokens:

• Sol, the flagship: $5 input / $30 output.
• Terra, a lower-cost model that OpenAI calls competitive with GPT-5.5: $2.50 / $15.
• Luna, the fastest and most affordable: $1 / $6.

OpenAI says the number marks the generation, while Sol, Terra and Luna are durable tiers that can advance on their own cadence. If model IDs are hard-coded across your scripts, put them behind one task-to-model mapping now.

Built for Fan-Out#

Several features point the same way:

• max reasoning gives the model more time than xhigh to explore alternatives, run checks and revise.
• ultra coordinates four agents in parallel by default, trading higher token use for stronger results and faster time-to-result. In Codex it is available on Plus and higher.
• The Responses API adds a multi-agent beta that runs concurrent subagents in a single request, and Programmatic Tool Calling, where the model writes and runs lightweight in-memory programs that coordinate tools and filter intermediate results. OpenAI says the latter is Zero Data Retention compatible.

I would try Programmatic Tool Calling first: large tool outputs passing back through the model on every step are expensive. Fan-out itself is a shape I already use in Quorate, which runs several reviewers over one diff. Parallel agents multiply whatever per-call waste you already have, so the next section matters more under ultra.

The Caching Change Will Show Up on Your Bill#

GPT-5.6 adds explicit cache breakpoints and a 30-minute minimum cache life. It also adds a price: cache writes cost 1.25x the uncached input rate, while reads keep the 90% discount. The prompt caching guide lists no additional cache-write charge for GPT-5.5. A prefix written and never read used to cost nothing extra. Now it costs 25% more than not caching.

With reuse, caching still wins: one write plus one read is 1.35x the plain input cost instead of 2x, and one write plus nine reads is 2.15x instead of 10x. It loses on content nobody reads again. In the guide's own implicit-mode example, a first 12,000-token request reports 12,000 cache-write tokens. For a one-shot call with a unique payload, that is a 25% surcharge on the input it writes to cache.

The fix is explicit-only mode. You mark the end of the stable prefix, and content after the last breakpoint is billed at the uncached rate with no write charge:

{
  "model": "gpt-5.6-sol",
  "prompt_cache_options": { "mode": "explicit" },
  "input": [
    { "role": "developer", "content": [{
      "type": "input_text",
      "text": "Stable rubric and examples...",
      "prompt_cache_breakpoint": { "mode": "explicit" } }] },
    { "role": "user", "content": "The unique item to evaluate..." }
  ]
}

Two more details from the guide: the minimum cacheable prefix on GPT-5.6 is 1,024 tokens, and prompt_cache_key is no longer needed for routing. It is now positioned for separate per-customer cache accounting, which also helps prevent cache-hit probing across users. In a multi-tenant product, that is a security decision, not only a billing one.

Compare Cost per Task, Not per Token#

OpenAI's headline claim: Sol with max reasoning scores 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Claude Fable 5, with less than half the output tokens, less than half the time and about one-third lower cost.

These are vendor numbers, but the framing is right: per-token price is a poor comparison when output length and retries differ this much. The metric I trust is cost per accepted change, counting input, output, cache writes, retries and human review time. A cheaper model that needs a second pass is not cheaper.

Copilot Is Off by Default for Organizations#

In Copilot, Sol is on Pro+, Max, Business and Enterprise, and Terra and Luna are also on Pro. Business and Enterprise admins must enable the GPT-5.6 policy, which is off by default, and usage is billed at provider list pricing. Decide the default tier before you flip it on; Sol at list price is expensive for routine work.

What I'm Doing About It#

For any harness that calls GPT-5.6, this is my list:

• Move dynamic content (dates, run IDs, user names) below the stable prefix.
• Log cache_write_tokens next to cached_tokens on every call, so writes are visible.
• Use explicit-only mode for one-shot review and judge calls, and implicit mode for long threads that only append.
• Run a small set of real tasks through Sol, Terra and the current default model, and compare cost per accepted patch.
• Keep ultra off for routine work until I know what four parallel agents cost on real workloads.

One note for security review agents: OpenAI says Sol's cyber safeguards block roughly ten times more potentially harmful activity than previous models. Test your prompts before depending on them.