GPT-5.6 Sol, Terra, and Luna: OpenAI Preview Launch Explained
OpenAI July 9: GPT-5.6 Sol, Terra, Luna rolling out in ChatGPT, Codex, and API. ALE 53.6, AA Coding Agent Index 80.0, Ultra mode, Terminal-Bench 91.9%, pricing.
Update — July 9, 2026 (OpenAI rollout):@OpenAI posted that Sol, Terra, and Luna are starting to roll out now in ChatGPT, Codex, and the API — with new GA benchmarks:
Claim
Score / detail
Agents' Last Exam
Sol 53.6 — +13.1 vs Fable 5 (adaptive)
AA Coding Agent Index
Sol 80.0 — +2.8 vs Fable; <½ tokens, ~⅓ cost
Ultra mode
Parallel multi-agent setting ships at launch
Efficiency
Terra/Luna beat Fable on ALE at ~1/16 cost (OpenAI claim)
July 1, 2026:Fable 5 live globally (status hub). GPT-5.6 was limited preview until this July 9 rollout.
On June 26, 2026, OpenAI officially previewed the GPT-5.6 family — not a single model, but three tiers: (flagship), (balanced everyday work), and (fast, affordable volume). The announcement landed the same week as and — a snapshot of frontier AI where capability, cost, and Washington oversight move in the same news cycle.
This is the end of the leak era for GPT-5.6. We now have official benchmarks, API pricing, Ultra mode with subagents, and OpenAI's clearest statement yet on why most people cannot use it today.
TL;DR — GPT-5.6 at a glance
Model
Role
vs GPT-5.5
API price (in / out per 1M)
GPT-5.6 Sol
Flagship
Step-function better
$5 / $30
GPT-5.6 Terra
Balanced daily work
Competitive at 2× lower cost
$2.50 / $15
GPT-5.6 Luna
High-volume
Strong capability, lowest cost
$1 / $6
Topic
Detail
Availability
Rolling out in ChatGPT, Codex, API (July 9, 2026)
GA benchmarks
ALE 53.6 · AA Coding Index 80.0 · TB 91.9% (Ultra)
Government
U.S. requested vetted-partner start; OpenAI opposes long-term default
Terminal-Bench 2.1
Sol Ultra 91.9% · Sol 88.8% · Mythos 5 84.3% · Fable 5 83.4%
New modes
Max reasoning effort · Ultra mode (subagents beyond single agent)
Cyber
Does not cross Cyber Critical threshold; layered safeguard stack
Speed
Cerebras Sol up to 750 tps in July (select customers)
Caching
Explicit breakpoints · 30-min minimum cache life · writes 1.25× · reads 90% off
GPT-5.6 Sol, Terra, and Luna — the June 26 preview, Terminal-Bench scores, and what limited access means in practice.
Sol, Terra, Luna — OpenAI's new naming system
OpenAI introduced a split that mirrors how enterprises already think about model routing:
Generation number (5.6) — training generation and capability baseline
Tier name (Sol / Terra / Luna) — durable product lanes that can advance independently
Sol is the new flagship — OpenAI calls it a step function better than GPT-5.5, not an incremental polish. Terra targets the economic middle: performance competitive with GPT-5.5 at half the token bill. Luna is the volume tier — strong capability at OpenAI's lowest cost in this family.
For developers already abstracting models behind routers, this is explicit permission to map hard agent tasks → Sol, production defaults → Terra, and classification / bulk → Luna without waiting for a new generation number every month.
Availability: rolling out July 9, 2026
OpenAI's July 9 thread superseded the June preview-only framing:
Limited preview ran June 26 through early July for trusted API and Codex partners
July 9 rollout — Sol, Terra, and Luna starting to appear in ChatGPT, Codex, and the API
Tier waves — not every account sees all three models on day one; check your picker
Government context — preview stagger was U.S.-requested; OpenAI opposes making that permanent
Practical read: If your account does not show GPT-5.6 yet, GPT-5.5 remains the fallback. International users: Will GPT-5.6 only be available in the USA? — rollout is global per OpenAI's thread, but verify your region.
Terminal-Bench 2.1 — where Sol leads the frontier
OpenAI's headline benchmark for coding agents is Terminal-Bench 2.1 — command-line workflows needing planning, iteration, and tool coordination (the same class of work Codex and terminal agents target).
Model
Terminal-Bench 2.1
GPT-5.6 Sol Ultra
91.9%
GPT-5.6 Sol
88.8%
GPT-5.5
88.0%
Claude Fable 5
83.4%
Claude Mythos 5
84.3%
GPT-5.6 Terra
82.5%
GPT-5.6 Luna
84.3%
Claude Opus 4.8
78.9%
Gemini 3.1 Pro Preview
70.7%
explainx.ai read: On this specific agentic terminal benchmark, GPT-5.6 Sol Ultra is the new state of the art — and base Sol (88.8%) still edges GPT-5.5 (88.0%) and Fable 5 (83.4%). That does not automatically mean Sol wins every coding leaderboard (SWE-Bench Pro gaps may persist on other harnesses), but it validates months of Codex log leaks pointing at agentic reliability gains.
Ultra mode — OpenAI's new configuration that uses subagents beyond a single agent — sits at the top at 91.9%. That connects directly to multi-agent orchestration themes in loop engineering and token economics: more agents, more tokens, more capability on long horizons.
OpenAI also added max reasoning effort on Sol — giving the model more time to reason deeply on hard tasks.
Biology and cybersecurity — beyond coding
GeneBench v1 (genomics / quant biology)
GPT-5.6 Sol shows broad improvements in long-horizon biology workflows. On GeneBench v1, OpenAI reports stronger results than GPT-5.5 while using fewer tokens — efficiency plus capability, not just brute-force compute.
Update — July 1, 2026: OpenAI shipped GeneBench-Pro — 129 harder synthetic computational biology problems. GPT-5.6 Sol hits 28.7% (31.5% Pro mode).
Update — July 11, 2026: OpenAI launched the $50K Bio Bounty Program the same day as GPT-5.6 GA — universal biology jailbreak testing on GPT-5.6 and GPT-5.5 (through July 27).
Cyber: capability up, safeguards up
OpenAI positions Sol as most capable yet for cybersecurity, shifting the performance-efficiency frontier on long-horizon security tasks including vulnerability research.
Benchmark
Finding
ExploitBench²
Sol competitive with Mythos Preview at roughly ⅓ the output tokens
ExploitGym
Sol, Terra, and Luna all show strong cyber improvements as reasoning increases
Preparedness Framework
Sol does not cross Cyber Critical — no autonomous full-chain exploit in Chromium/Firefox tests under stated conditions
OpenAI's framing: Sol is better at helping defenders find and fix vulnerabilities than at reliably executing end-to-end attacks. That mirrors Washington's Mythos-class concern while arguing for defender access with safeguards.
Safeguard stack (preview):
Model-trained refusals for prohibited cyber assistance (including jailbreak attempts)
Real-time cyber and biology misuse classifiers — generation can pause for larger-model review
Account-level review across conversations for persistent abuse vs legitimate dual-use work
700,000+ A100-equivalent GPU hours on automated red teaming for universal jailbreaks
Continued third-party human red teaming during preview
Expect blocks, refusals, and latency during preview — OpenAI says that friction is intentional while they tune false positives on legitimate security work.
Pricing and caching — Terra undercuts GPT-5.5 economics
Official API pricing per 1M tokens:
Model
Input
Output
GPT-5.6 Sol
$5.00
$30.00
GPT-5.6 Terra
$2.50
$15.00
GPT-5.6 Luna
$1.00
$6.00
Sol at GPT-5.5 list price is notable — flagship upgrade without a headline price hike on the top tier. Terra at half of GPT-5.5 is the enterprise story: route "good enough" workloads down a tier and cut spend without leaving the 5.6 generation.
Prompt caching (5.6+):
Explicit cache breakpoints
30-minute minimum cache lifetime
Cache writes billed at 1.25× uncached input rate
Cache reads keep 90% discount
For agent loops that reuse system prompts and tool schemas, predictable caching matters as much as raw per-token rates — especially when Ultra mode multiplies subagent calls.
Cerebras: 750 tokens per second in July
OpenAI is launching GPT-5.6 Sol on Cerebras at up to 750 tokens per second in July 2026, initially for select customers as capacity expands. That is a latency play for interactive agents — separate from the benchmark story but relevant for production UX on Sol.
How this changes the Fable 5 comparison
Claude Fable 5 is live globally since July 1 (status). OpenAI's July 9 GA thread adds new comparison axes beyond Terminal-Bench:
Benchmark
GPT-5.6 Sol
Fable 5 (implied)
Agents' Last Exam
53.6
~40.5 (adaptive)
AA Coding Agent Index
80.0
~77.2
Terminal-Bench 2.1 (Ultra)
91.9%
83.4%
Terra/Luna beat Fable on ALE at ~1/16 cost per OpenAI — the volume-tier economics story. SWE-Bench Pro may still favor Fable until OpenAI publishes Sol scores there.
Access beats benchmarks in 2026: Both frontier tiers are now reachable — route by task and harness, not vendor loyalty alone.
What people are asking on X
"Is GPT-5.6 out or not?"
Rolling out as of July 9, 2026 — ChatGPT, Codex, and API. Tier waves may lag; check your account.
"Did OpenAI confirm the government story?"
Yes — in their own blog. They previewed with the U.S. government, started limited at government's request, and explicitly said they do not want this to become permanent process.
"Is Sol worth it if Terra matches GPT-5.5 cheaper?"
Sol for hardest agentic work, Ultra/subagents, and max reasoning. Terra when GPT-5.5-class quality suffices at half the bill. Luna for high-volume pipelines where 84.3% Terminal-Bench-class behavior is enough.
"Does this mean Fable 5 is obsolete?"
On ALE, AA Coding Index, and Terminal-Bench, Sol leads per OpenAI's GA claims. Fable's SWE-Bench Pro lead is not fully refuted until OpenAI publishes Sol scores there — and Fable remains the default for Claude Code stacks.
"Will cyber safeguards block my red team?"
Maybe during preview. OpenAI warns of false positives on dual-use security work and asks for feedback to reduce unnecessary blocks before wider release.
What to do now
If you are…
Action
On GPT-5.5 in production
Stay put until your account tier gets GA; benchmark Terra for cost routing
Building Codex agents
Watch partner preview reports; design for Sol / Terra / Luna routing now
Comparing to Fable 5
Update Terminal-Bench assumptions; read full comparison post
Pricing and June 26 preview context retained. July 9 rollout benchmarks (ALE, AA Coding Index) are vendor-reported — verify availability on openai.com and your account tier.
