OpenAI shipped GPT-5.6. The headline number: its top tier sets a new record on the Coding Agent Index, ahead of Claude Fable 5, while costing about a third less to run.
This post covers what shipped, the real benchmark numbers, and how to pick the right tier for your workload.
The News
GPT-5.6 ships as three tiers, not one model:
- Sol — the flagship, built for frontier reasoning work
- Terra — a balanced mid-tier for everyday tasks
- Luna — the fastest and cheapest, for high-volume jobs
The benchmark numbers, from Artificial Analysis:
- Coding Agent Index: Sol (max reasoning) scores 80 points — a new state of the art, 2.8 points above Claude Fable 5. It gets there using less than half the output tokens, less than half the time, and about a third less cost.
- Intelligence Index: Sol (max) scores 59, just 1 point behind Fable 5’s 60, at roughly a third of the cost.
- Token efficiency: Sol (max) uses about 15,000 output tokens per Intelligence Index task, down from GPT-5.5’s 16,000.
So Sol does not beat Fable 5 on every axis. It wins clearly on coding tasks. On general intelligence it is close behind, at a much lower price.
Pricing by Tier
| Tier | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Sol | $5 | $30 |
| Terra | $2.50 | $15 |
| Luna | $1 | $6 |
Luna costs about 80% less than Sol. Cached input tokens (repeated context, like a long agent conversation) get a 90% discount on the input price.
What It Means for Developers
1. Coding agents get cheaper without losing quality. If you run an AI coding agent for reviews, bug fixes, or feature work, Sol now beats the previous best (Fable 5) at that job, for less money.
2. Most tasks do not need the flagship. Terra and Luna exist because not every call needs frontier reasoning. A simple classification, a short rewrite, or a routine agent step can run on Luna at a fraction of Sol’s cost.
3. Cache-friendly agent loops save even more. If your agent re-sends similar context on every step (common in long coding sessions), the 90% cache-read discount compounds fast.
How to Use It
Pick the model ID for the tier you need:
import openai
client = openai.OpenAI()
# Sol: frontier reasoning, e.g. a hard bug in unfamiliar code
resp = client.chat.completions.create(
model="gpt-5.6-sol",
messages=[{"role": "user", "content": "Find the race condition in this async function."}],
)
# Terra: everyday coding tasks
resp = client.chat.completions.create(
model="gpt-5.6-terra",
messages=[{"role": "user", "content": "Write a unit test for this function."}],
)
# Luna: fast, high-volume, simple tasks
resp = client.chat.completions.create(
model="gpt-5.6-luna",
messages=[{"role": "user", "content": "Classify this log line as error, warning, or info."}],
)
print(resp.choices[0].message.content)
The bare alias gpt-5.6 routes to Sol, so model="gpt-5.6" and model="gpt-5.6-sol" behave the same today. Use the explicit tier name so your code does not silently change behavior if the alias target ever moves.
Quick Summary
| Question | Answer |
|---|---|
| Best for coding tasks | Sol — new state of the art, beats Fable 5 |
| Best for general intelligence | Close behind Fable 5, at ~1/3 the cost |
| Cheapest tier | Luna — 80% cheaper than Sol |
| Cache discount | 90% off input on cache reads |
| Pick Sol when | Hard reasoning, coding-agent work |
| Pick Terra when | Everyday coding tasks, balanced cost |
| Pick Luna when | High-volume, simple, fast jobs |
FAQ
Is GPT-5.6 better than Claude Fable 5? On coding tasks, yes — Sol sets a new benchmark record there. On general intelligence, Fable 5 is still slightly ahead, but Sol closes most of the gap at about a third of the price.
Which tier should I default to? Start with Terra for everyday coding work. Reach for Sol only when a task needs frontier reasoning. Use Luna for high-volume, low-complexity calls where speed and cost matter more than depth.
Does the pricing include caching? Cached input tokens (context you have sent before) cost 90% less than fresh input tokens. Cache writes cost 1.25x the input price. This matters most for long-running coding agents that resend context on every step.