Skip to content
Breaking
July 9, 20269 min read

GPT-5.6: OpenAI's Sol, Terra, and Luna Three-Tier Family

OpenAI made GPT-5.6 generally available on July 9, 2026 as a three-tier family: Luna ($1/$6), Terra ($2.50/$15), and flagship Sol ($5/$30), all with a 1.05M-token context. We break down the verified Terminal-Bench 2.1, Agents' Last Exam, and SWE-Bench Pro numbers, the new agent-first API (programmatic tool calling, subagents, cache breakpoints), the pricing ladder, and where each tier lands in our live rankings.

3
Model tiers
Luna / Terra / Sol
1.05M
Context window
tokens, all tiers
91.9%
Terminal-Bench 2.1
Sol Ultra, top score
$1 / $6
Luna pricing
per 1M in / out
Overview

What OpenAI shipped

On July 9, 2026, OpenAI made GPT-5.6 generally available across the API, Codex, and ChatGPT, after previewing it on June 26. Rather than a single model, GPT-5.6 is a three-tier family with a shared 1.05M-token context window and a February 16, 2026 knowledge cutoff. The tiers trade cost for capability: Luna is the fast, cheap option; Terra is the balanced everyday model; and Sol is the flagship for the hardest work. All three landed the same day in GitHub Copilot as well.

The headline structural change is not the raw scores but the delivery: OpenAI positions Terra as matching GPT-5.5-class quality at roughly half the price, and both Terra and Luna reportedly match Anthropic's Claude Fable 5 on the Agents' Last Exam at around one-sixteenth of the cost. GPT-5.6 also ships new API machinery - programmatic tool calling, subagents, and explicit prompt-cache control - aimed squarely at long-running agent workloads.

The family

Three tiers: Luna, Terra, Sol

TierPositioningInput / 1MOutput / 1MTerminal-Bench 2.1
GPT-5.6 LunaFast and affordable$1.00$6.0082.5%
GPT-5.6 TerraBalanced, everyday work$2.50$15.0084.3%
GPT-5.6 SolFlagship, hardest tasks$5.00$30.0088.8%

Each tier also has a -pro variant in the API, and Sol adds an Ultra mode that scores 91.9% on Terminal-Bench 2.1 by leaning on subagents rather than a single pass. Pricing is identical across a tier and its pro variant; the difference is reasoning depth and orchestration, not per-token cost.

Benchmarks

Where the numbers actually land

On Terminal-Bench 2.1 (agentic terminal tasks), the family stacks up cleanly against the field:

GPT-5.6 Sol Ultra91.9%
GPT-5.6 Sol88.8%
Claude Mythos 588.0%
GPT-5.6 Terra84.3%
GPT-5.583.4%
GPT-5.6 Luna82.5%
Claude Opus 4.878.9%

On the Agents' Last Exam (long-running professional workflows across 55 fields), Sol scores 53.6, which OpenAI reports as 13.1 points ahead of Claude Fable 5, with Terra and Luna also edging out Fable 5 at a fraction of the cost.

The honest counterpoint: on SWE-Bench Pro, Sol scores 64.6% versus Claude Fable 5's 80% - a clear loss. OpenAI's own audit noted that roughly 30% of SWE-Bench Pro tasks are broken, so the gap is worth reading with that caveat, but GPT-5.6 does not top every board. It leads on agentic terminal and long-horizon workflow tasks and trails Anthropic on this particular software-engineering set.

Economics

Pricing and the caching story

GPT-5.6 Luna
$1.00 / $6.00
per 1M input / output tokens
GPT-5.6 Terra
$2.50 / $15.00
per 1M input / output tokens
GPT-5.6 Sol
$5.00 / $30.00
per 1M input / output tokens

The three-tier ladder is the whole point: Terra costs half of Sol and, per OpenAI, holds GPT-5.5-class quality, while Luna is cheaper still for high-volume or latency-sensitive work. Cached input reads keep the 90% discount, cached writes cost 1.25x, and GA added a 30-minute minimum cache life plus explicit cache breakpoints for fine-grained control.

Simon Willison's cost probe makes the spread concrete: the same image-generation task ranged from 0.71 cents on Luna with no reasoning to 48.55 cents on Sol at max reasoning effort - a ~68x span within one model family, which is exactly the knob the three tiers plus reasoning modes are meant to give you.

What's new

Agent-first API machinery

Programmatic tool calling

The Responses API can orchestrate tools with JavaScript instead of round-tripping every call through the model, cutting latency and token overhead on multi-tool agents.

Subagents and Ultra mode

Sol Ultra spins up subagents to parallelize complex work, going beyond a single agent pass. That is how Ultra reaches 91.9% on Terminal-Bench 2.1 versus 88.8% for a single Sol run.

Max reasoning effort

A new max reasoning tier gives Sol the most time to think on the hardest problems, trading latency and cost for depth.

Cerebras speed path

GPT-5.6 Sol runs on Cerebras at up to ~750 tokens per second, aimed at interactive and agentic use where wall-clock latency matters.

Rounding out the release: explicit prompt-cache breakpoints and image detail options (including detail: original for full-resolution image inputs). The through-line is that GPT-5.6 is tuned less for one-shot chat and more for long-running, tool-heavy agents.

Live data

Where GPT-5.6 sits in our rankings

Our leaderboard rescinds and rescores every model hourly. Here is where the GPT-5.6 family currently lands in the coding category:

Compare the family directly against the current leaders on the live model leaderboard, or line Sol up against Anthropic's flagship on the GPT-5.6 Sol vs Claude Fable 5 comparison.

Guidance

Which tier should you use

Reach for Luna

High-volume, latency-sensitive, or cost-capped work: classification, extraction, routing, first-pass drafting, and agent steps where you call the model thousands of times. At $1/$6 it is the cheapest tier and still clears 82.5% on Terminal-Bench 2.1.

Default to Terra

The everyday workhorse. OpenAI positions it at GPT-5.5-class quality for roughly half the price, so most production agents and coding assistants should start here and only escalate to Sol when a task genuinely needs it.

Escalate to Sol (or Sol Ultra)

The hardest agentic and long-horizon work: multi-step research, complex refactors, and tasks worth paying max reasoning for. Use Ultra mode when parallel subagents help, keeping in mind Anthropic's Fable 5 still leads on SWE-Bench Pro.

Sources

Primary sources

Pricing, context, and benchmark figures reflect OpenAI's GPT-5.6 GA announcement (July 9, 2026) and the coverage cited above. Live ranking positions are computed hourly from this site's pipeline and will move as benchmark coverage of the new models fills in.

Frequently Asked Questions

GPT-5.6 is a three-tier model family OpenAI made generally available on July 9, 2026 (previewed June 26). Instead of one model, it ships as Luna (fast and affordable, $1/$6 per million tokens), Terra (balanced everyday model, $2.50/$15), and Sol (flagship for the hardest work, $5/$30). All three share a 1.05M-token context window, a 128K max output, and a February 16, 2026 knowledge cutoff, and are available through the OpenAI API, Codex, ChatGPT, and GitHub Copilot.

Per million tokens: Luna is $1 input / $6 output, Terra is $2.50 / $15, and Sol is $5 / $30. Cached input reads keep a 90% discount, cached writes cost 1.25x, and GA added a 30-minute minimum cache life plus explicit cache breakpoints. OpenAI positions Terra at GPT-5.5-class quality for roughly half the price, so the family is designed to let you match spend to task difficulty.

It splits the decision. On Terminal-Bench 2.1, GPT-5.6 Sol scores 88.8% and Sol Ultra reaches 91.9%, ahead of Claude Mythos 5 at 88.0% and Claude Opus 4.8 at 78.9%. On the Agents' Last Exam, Sol scores 53.6, which OpenAI reports as 13.1 points ahead of Claude Fable 5. But on SWE-Bench Pro, Sol scores 64.6% versus Fable 5's 80%, a clear loss - though OpenAI's audit noted roughly 30% of SWE-Bench Pro tasks are broken. GPT-5.6 leads on agentic terminal and long-horizon workflow tasks; Fable 5 leads on that software-engineering set.

Sol is the flagship single-pass model. Sol Ultra uses subagents to parallelize complex work and reaches 91.9% on Terminal-Bench 2.1 versus 88.8% for a single Sol run. Each tier also has a -pro variant in the API. Pricing is identical within a tier and its pro variant - the difference is reasoning depth and orchestration, plus a new max reasoning effort mode that gives Sol more time to think on the hardest problems.

GPT-5.6 is tuned for agents. The Responses API adds programmatic tool calling (orchestrating tools with JavaScript instead of round-tripping each call through the model), multi-agent subagents, explicit prompt-cache breakpoints, and image detail options including full-resolution inputs. GPT-5.6 Sol also runs on Cerebras at up to about 750 tokens per second for latency-sensitive use.

Use Luna for high-volume, latency-sensitive, or cost-capped work like classification, extraction, and routing (it still clears 82.5% on Terminal-Bench 2.1). Default to Terra for most production agents and coding assistants, since it targets GPT-5.5-class quality at about half the cost of Sol. Escalate to Sol or Sol Ultra for the hardest agentic and long-horizon work where max reasoning and subagents earn their cost.

想自己探索数据吗?

GPT-5.6: OpenAI's Sol, Terra, and Luna Three-Tier Family | LM Market Cap