GPT-5.6 Ships as Three Models That Differ Only in Price

Introduction

On July 9, 2026, OpenAI made GPT-5.6 generally available across ChatGPT, Codex and the API. It ships as three named models — Sol, Terra and Luna — and the interesting thing about them is what they have in common.

All three carry a 1,050,000 token context window, a 128,000 token output cap, and a February 16, 2026 knowledge cutoff. Same modalities, same tools, same API surface. What separates them is speed, how much reasoning they will spend, and a 5x price spread: Sol at $5.00 / $30.00 per million tokens, Terra at $2.50 / $15.00, Luna at $1.00 / $6.00.

That is a different product shape from the usual tiering, where the cheap model is also the small-context model. Here the cheap model is not crippled — Luna will take a million-token prompt for a dollar.

Routing, not choosing

Because the tiers share a spec sheet, picking one stops being an architectural decision and becomes a runtime one. OpenAI's developer guidance says so more or less directly: start on Luna for repetitive work, escalate to Terra for multi-step tasks, and reserve Sol for steps where failure is expensive.

For anyone building an AI agent, that is the shape of the system: Luna doing extraction and classification as a subagent, Terra orchestrating, Sol called only when a grader flags something hard. The bare gpt-5.6 alias routes to Sol — convenient for experiments, and the wrong default for a production bill.

Reasoning effort is a second, independent dial, now ranging from none through max. Both it and the tier are worth logging per request, because neither is a constant you can reason about once and bury in a config file.

One caveat that breaks the naive version of this. A cheaper tier is cheaper per token, not per finished task. Launch analyses found Terra spending substantially more output tokens than Sol on the same long coding jobs — enough that the per-token saving can evaporate. Reasoning tokens are billed as output, so a model that thinks longer to reach the same answer costs more despite the lower rate card. Measure cost per solved task, on your own workload, before committing an architecture to a tier.

Programmatic Tool Calling

The most substantial API change is a new way to call tools. Instead of the model emitting one tool call, receiving the result into its context, and emitting the next, GPT-5.6 can write JavaScript that runs in an isolated V8 runtime with no network access, orchestrating several tools and filtering their output before anything reaches the context window.

The problem it solves is real and expensive. A conventional tool loop that queries an API returning 40,000 tokens of JSON pays for all 40,000 tokens as input on every subsequent turn, even if the model needed three fields from it. Filtering inside the sandbox means the context only ever sees the three fields.

A related beta, the ultra setting, runs four agents concurrently on independent subtasks. OpenAI is explicit that it is experimental, and it only pays off when the subtasks genuinely do not depend on each other and you have written sensible synthesis logic downstream.

Benchmarks, including the one OpenAI lost

The headline result is Agents' Last Exam, a benchmark from UC Berkeley's Center for Responsible, Decentralized Intelligence. It is a better test than most: over 1,500 tasks drawn from real professional work across 55 industries, contributed by more than 300 practitioners, and graded by deterministic scripts against expert reference solutions rather than by an AI judge.

OpenAI's published table:

The number in the table is not the number in the headline. OpenAI's prose claims "a new high of 53.6" and a 13.1-point lead over Fable 5, while its own comparison table says 52.7. Only the higher figure reconciles — 53.6 minus 40.5 is exactly 13.1. Berkeley RDI attributes the 53.6 specifically to Codex running Sol at high reasoning effort, a different configuration from the bare-model row. So both figures are real; they measure different things, and OpenAI's page does not say so.

Note the spread in that table. Luna, at a sixteenth of Sol's price, sits 2.4 points behind it and nearly ten points ahead of Anthropic's flagship. If the routing argument above needed evidence, this is it.

Then there is SWE-Bench Pro, the most-cited agentic coding benchmark, where Sol scores 64.6% against Claude Fable 5's 80.3%. A 16-point deficit is not a rounding error.

OpenAI's response was to challenge the benchmark, estimating that roughly 30% of its tasks are broken. That may well be true — SWE-Bench variants have a long history of contaminated and unsolvable instances, and the criticism deserves investigation. But a vendor disputing the scoring of a test it lost is an argument, not a result, and until someone re-runs the suite the published gap is what exists. Sol also trails Claude models on tool-use evaluations.

Sol additionally leads the Artificial Analysis Coding Agent Index at 80, at the time of publication.

What it costs, and one thing that got more expensive

Cached reads stay at a 90% discount. Cache writes now cost 1.25x the uncached input rate — that is new, and it inverts the usual advice. Caching a prefix you will reuse fifty times is still an obvious win; caching one you will reuse once is now a measurable loss. Put stable instructions before request-specific data and set your breakpoints deliberately.

Two older modifiers still apply: prompts above 272K input tokens are billed at a premium, and regional processing adds 10%.

OpenAI is also serving Sol on Cerebras hardware at up to 750 tokens per second, which matters for any inference workload where latency, not cost, is the constraint.

What did not ship

There is no GPT-5.6 Pro. The high-compute option is still gpt-5.5-pro, sitting on a December 2025 cutoff at $30 / $180 per million. If your workload depends on the Pro tier, this release does not reach you.

Fine-tuning is not supported on any GPT-5.6 model.

GPT-5.5 is not retired, and this is the detail worth acting on: it costs exactly the same as Sol — $5 and $30 — with a knowledge cutoff eleven weeks earlier. There is no reason to start new work on it, and the migration is a model-ID change.

Conclusion

GPT-5.6 is a pricing and plumbing release more than a capability one. The three-tier structure with a shared spec sheet makes cost a routing problem you can actually solve, Programmatic Tool Calling attacks context bloat at its source, and Luna at $1 per million input tokens with a full 1M window is the cheapest serious long-context option OpenAI has shipped.

It is not a clean sweep. Claude still leads on SWE-Bench Pro by 16 points and on tool use, there is no Pro tier, and the cache-write surcharge quietly makes some existing prompt architectures worse. Read together with Google's Gemini 3.6 Flash release twelve days later — also cheaper, also more token-efficient, also not the benchmark leader — the pattern across both vendors is the same: the frontier is being sold on cost per completed task rather than on top-line scores.

Sources

Specifications and pricing here are taken from OpenAI's developer documentation. The benchmark figures originate in OpenAI's launch post, which does not serve to our fetcher; we have taken them from the launch analyses cited below rather than from the page itself, and flagged where those sources disagree.

• GPT-5.6 Sol model reference — context, limits, tools, rate limits
• GPT-5.6 Luna model reference
• OpenAI API pricing — all three tiers, cached and long-prompt rates
• GPT-5.6 announcement — OpenAI's general-availability post, the origin of the benchmark claims
• The new GPT-5.6 family: Luna, Terra, Sol — independent launch analysis
• Agents' Last Exam and the 52.7 / 53.6 discrepancy — benchmark methodology and Berkeley RDI's attribution
• Sol, Terra, Luna: benchmarks and which tier to use — cost-per-task analysis
• GPT-5.6 — our model page, with the full family table
• ChatGPT — where the tiers land for end users