OpenAI 发布 GPT-5.6（Sol, Terra, Luna）：三层模型家族，Responses API 支持编程式工具调用

OpenAI just moved the GPT-5.6 family to general availability, following a limited preview. The release ships three models rather than one. Sol is the flagship, Terra is the balanced everyday tier, and Luna is the most cost-efficient.

To sum up

GPT-5.6 ships three tiers — Sol, Terra, Luna — priced from $1/$6 to $5/$30 per 1M tokens.
Sol leads the Artificial Analysis Coding Agent Index at 80, 2.8 points above Claude Fable 5.
Programmatic Tool Calling runs model-written JavaScript in an isolated V8 runtime with no network access.
ultra runs four agents in parallel, lifting Terminal-Bench 2.1 from 88.8% to 91.9%.
SWE-Bench Pro remains a gap: Sol’s 64.6% trails Claude Mythos 5’s 80.3% by roughly 15 points.

What is GPT-5.6?

Three models, one generation, priced per 1M tokens. Sol is $5 input and $30 output. Terra is $2.50 and $15. Luna is $1 and $6.

Availability differs by surface:

Chat: Plus, Pro, Business, and Enterprise users access Sol at medium and higher effort. Pro and Enterprise can also select GPT-5.6 Sol Pro.
ChatGPT Work and Codex: Free and Go users access Terra. Paid users choose among all three and set effort per model. max is available to all users with GPT-5.6 access and is toggled in settings.
API: All three tiers are available. Programmatic Tool Calling and a multi-agent beta both live in the Responses API.

Prompt caching also changed. GPT-5.6 supports explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25x the model’s uncached input rate. Cache reads continue to receive the 90% cached-input discount.

Performance

Furthermore, Agents’ Last Exam evaluates long-running professional workflows across 55 fields. OpenAI reports a new high of 53.6 for Sol. It describes this as eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points.

OpenAI’s own eval table lists Sol at 52.7% and Fable 5 at 40.5%. The 13.1-point gap matches 53.6 minus 40.5, so the Fable 5 baseline is consistent across both. Only Sol’s figure differs. OpenAI does not label which reasoning configuration produced 53.6.

On the Artificial Analysis Coding Agent Index v1.1, Sol at max reasoning scores 80. That is 2.8 points above Fable 5. OpenAI reports it does so using less than half the output tokens and less than half the time.

Sol sets new state-of-the-art results on Terminal-Bench 2.1 and DeepSWE. It reaches 92.2% on BrowseComp and 62.6% on OSWorld 2.0. On OSWorld it surpasses Claude Opus 4.8 while using 85% fewer output tokens.

媒体内容 · 前往原文查看

Eval	GPT-5.6 Sol	GPT-5.6 Terra	GPT-5.6 Luna	GPT-5.5	Claude Fable 5	Claude Opus 4.8	Gemini 3.1 Pro Preview
AA Coding Agent Index v1.1	80	77.4	74.6	76.4	77.2	72.5	42.7
AA Intelligence Index v4.1	58.9	55	51.2	54.8	59.9	55.7	46.5
Terminal-Bench 2.1	88.8%	87.4%	84.7%	85.6%	83.1%	78.9%	70.7%
DeepSWE v1.1	72.7%	69.6%	67.2%	67%	69.7%	59%	11.8%
SWE-Bench Pro	64.6%	63.4%	62.7%	59.4%	80%	69.2%	54.2%
Agents’ Last Exam	52.7%	50.4%	50.3%	46.9%	40.5%	45.2%	32.1%
GDPval-AA v2 (Elo)	1,747.8	1,593	1,591.8	1,493.7	1,759.6	1,600.1	962.3
BrowseComp	90.4%	87.5%	83.3%	84.4%	—	84.3%	85.9%
OSWorld 2.0	62.6%	50.2%	45.6%	47.5%	—	54.8%	—
Toolathlon	58%	53.1%	53.4%	55.6%	61.7%	59.9%	48.8%

Source: OpenAI’s published GPT-5.6 eval tables. Sol Ultra reaches 91.9% on Terminal-Bench 2.1 and 92.2% on BrowseComp. Claude Mythos 5 scores 80.3% on SWE-Bench Pro and 88% on Terminal-Bench 2.1. A dash means the score was not reported.

Where GPT-5.6 Does Not Lead

However, four gaps are worth naming:

SWE-Bench Pro: Sol scores 64.6%. Claude Mythos 5 scores 80.3% and Fable 5 scores 80%. That is a roughly 15-point deficit on a widely watched coding eval.
Broad intelligence and knowledge work: Fable 5 leads the Artificial Analysis Intelligence Index v4.1, 59.9 to 58.9. Fable 5 also leads GDPval-AA v2 by about 12 Elo. On HealthBench Professional, Fable 5 scores 60.9% against Sol’s 60.5%.
Tool use: On Toolathlon, Sol scores 58%. Fable 5 reaches 61.7% and Opus 4.8 reaches 59.9%. Luna also edges out Terra here, inverting the tier order.
Long context: Luna drops to 41.3% on OpenAI MRCR v2 8-needle, at both 256K–512K and 512K–1M. Sol scores 73.8% at 512K–1M, slightly below GPT-5.5’s 74%.

Interactive Explainer

Interactive · OpenAI GPT‑5.6 · July 9, 2026

GPT‑5.6 tier, cost and benchmark explorer

Every number below is taken from OpenAI’s published GPT‑5.6 eval tables and price list. Move the controls to see how tier choice changes spend and score.

Three durable capability tiers

The number is the generation. Sol, Terra and Luna are tiers that advance on their own cadence. Select a tier to see its role and rate card.

What a workload actually costs

Set your per-request token volume. Costs use OpenAI’s published per-1M rates. Cached input reads keep the 90% cached-input discount.

Input tokens per request

Output tokens per request

Requests per day

Share of input served from cache

Estimated monthly spend · 30 days

Sol

1.0x

Terra

Luna

Cache billing:

from GPT‑5.6 onward, cache writes bill at 1.25x the uncached input rate. Cache reads keep the 90% discount, with a 30‑minute minimum cache life and explicit cache breakpoints.

Benchmark scores, side by side

Values are OpenAI’s published eval table. A dash in that table means the score was not reported, so the model is omitted here. OpenAI states its latency and cost figures are simulated offline, not measured in production.

GPT‑5.6 family

GPT‑5.5

Anthropic

Google

Ultra: four agents, by default

Ultra coordinates four agents in parallel by default. It trades higher token use for a stronger score and faster time‑to‑result. Toggle to compare against Sol’s one‑agent baseline. OpenAI also charts 16‑agent runs on BrowseComp and SEC‑Bench Pro.

Score · single agent

Terminal-Bench 2.1

BrowseComp

SEC-Bench Pro

Where to get it:

ultra runs in ChatGPT Work for Pro and Enterprise, and in Codex for Plus and higher. In the API, the multi‑agent beta in the Responses API builds ultra‑like flows.

Source: OpenAI, “GPT‑5.6” · verified July 9, 2026

Built by

Strengths and Weaknesses

Strengths

Sets the Artificial Analysis Coding Agent Index at 80, above Fable 5’s 77.2
Programmatic Tool Calling has a documented API contract and named-customer token reductions of 38% to 63.5%
All three tiers clear GPT-5.5 on DeepSWE v1.1, SWE-Bench Pro, and Agents’ Last Exam
A 5x price spread across tiers lets teams route by task difficulty
Explicit cache breakpoints and a 30-minute minimum cache life make caching predictable
Sol reaches 62.6% on OSWorld 2.0 using 85% fewer output tokens than Opus 4.8

At the same time, the release faces a few distinct hurdles:

Weaknesses

SWE-Bench Pro trails Claude Mythos 5 and Fable 5 by roughly 15 points
Fable 5 leads on AA Intelligence Index v4.1, GDPval-AA v2, HealthBench Professional, and Toolathlon
The 53.6 Agents’ Last Exam headline does not appear in OpenAI’s own table, which lists 52.7%
Luna scores 41.3% on MRCR v2 8-needle, and Luna trails GPT-5.5 on Terminal-Bench 2.1
Cyber scores are measured with reduced safeguards; production behavior will differ
Cache writes now cost 1.25x the uncached input rate, a new line item to model
Latency and cost claims are OpenAI’s offline simulations, not measured production numbers