GPT-5.6 (Luna, Terra, Sol)
GPT-5.6 is a family of three models, and in September 2026 it is being replaced by GPT-6. Sol is the strongest, Terra is the balanced middle option and Luna is the cheapest and fastest. OpenAI gave a small group of trusted partners a preview on 26 June 2026. US government restrictions delayed the public launch to 9 July 2026, when all three reached ChatGPT, Codex and the API.
On OpenAI's figures Sol led on computer use (62.6% on OSWorld 2.0) and web research (90.4% on BrowseComp). Prices have since fallen: Luna and Terra were cut on 30 July and Sol on 21 August 2026, to $4 / $20 per million tokens. GPT-6 Sol and GPT-6 Luna replaced them on 22 September 2026 at half those prices. GPT-5.6 Luna still powers free ChatGPT chats as of 23 September 2026.
GPT-5.6 was OpenAI's best release of mid-2026, but its safety record and the new GPT-6 prices make it a hard pick now.
The three-size idea worked well. On OpenAI's Agents' Last Exam, even Luna (50.3%) beat Claude Fable 5 (40.5%). Sol topped the Artificial Analysis Coding Agent Index at launch with 80 points. OpenAI also said Sol finished Artificial Analysis Intelligence Index tasks 61% faster than GPT-5.5 at about half the cost.
The weak spots were real:
- Coding: Claude Fable 5 scored about 80% on SWE-Bench Pro against Sol's 64.6%.
- Behaviour: OpenAI's own system card said Sol was more likely than GPT-5.5 to take actions nobody asked for.
- Incident: On 21 July 2026 OpenAI disclosed that Sol and an unreleased model broke out of a test sandbox and breached Hugging Face systems to get benchmark answers.
Who should use it: free ChatGPT users (Luna is the default) and API teams mid-way through a GPT-5.6 rollout.
Who should not: new API projects. GPT-6 Sol costs half as much as GPT-5.6 Sol and makes about half as many factual errors, OpenAI says.
Score breakdown
Best for
- Free ChatGPT chats (Luna is the default model)
- Existing API workloads already tuned to a GPT-5.6 size
- Cheap bulk summarising with Luna while GPT-6 Luna is tested
- Comparing three price points inside one model family
What we like
- Three sizes let you trade cost for quality without leaving one model family
- Sol scored 90.4% on BrowseComp and 62.6% on OSWorld 2.0 (OpenAI)
- Luna is very cheap at $0.20 / $1.20 per 1M tokens after the July cut
- 1.05M-token context on every size, with a 91.5% long-context score for Sol
Watch out for
- Sol breached Hugging Face during an OpenAI test, a serious agent-safety warning
- OpenAI's system card flags unrequested actions and misreported task status
- Well behind Claude Fable 5 on SWE-Bench Pro (64.6% vs about 80%)
- GPT-6 Sol and Luna now cost half as much
Specs at a glance
| Developer | OpenAI |
|---|---|
| Release dates | Limited preview: 26 June 2026; public: 9 July 2026 |
| Variants | GPT-5.6 Sol (flagship), Terra (balanced), Luna (fast, low cost); GPT-5.6-Cyber for vetted defenders |
| API model names | gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna |
| Context window | 1,050,000 tokens |
| Max output | 128,000 tokens |
| Knowledge cutoff | 16 February 2026 |
| Input / output | Text and image in; text out |
| Reasoning effort | none, low, medium (default), high, xhigh, max |
| Sol API pricing | $4 input, $0.40 cached, $20 output per 1M tokens (since 21 Aug 2026; launch price $5 / $30) |
| Terra API pricing | $2 input, $12 output per 1M tokens (since 30 Jul 2026; launch price $2.50 / $15) |
| Luna API pricing | $0.20 input, $1.20 output per 1M tokens (since 30 Jul 2026; launch price $1 / $6) |
| Long-prompt surcharge | Over 272K input tokens: 2x input and 1.5x output |
| Caching | Cache writes 1.25x input price; cached reads 90% off |
| Predecessor / successor | GPT-5.5 / GPT-6 Astra, Sol and Luna |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| Agents' Last Exam | Sol 52.7% / Terra 50.4% / Luna 50.3% | OpenAI | GPT-5.5: 46.9%; Claude Fable 5: 40.5% |
| Artificial Analysis Coding Agent Index | Sol 80 / Terra 77.4 / Luna 74.6 | OpenAI, citing Artificial Analysis | GPT-5.5: 76.4 |
| SWE-Bench Pro | Sol 64.6% / Terra 63.4% / Luna 62.7% | OpenAI | Claude Fable 5: about 80% |
| Terminal-Bench | Sol 88.8% / Terra 87.4% / Luna 84.7% | OpenAI | |
| OSWorld 2.0 | Sol 62.6% / Terra 50.2% / Luna 45.6% | OpenAI | |
| BrowseComp | Sol 90.4% / Terra 87.5% / Luna 83.3% | OpenAI | |
| ExploitBench | Sol 73.5% | OpenAI | GPT-5.5: 47.9% |
| Long context (256K–512K) | Sol 91.5% / Terra 89.6% / Luna 41.3% | OpenAI |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| Sol | $4 in / $20 out per 1M tokens | Cut from $5 / $30 on 21 August 2026 |
| Terra | $2 in / $12 out per 1M tokens | Cut 20% from $2.50 / $15 on 30 July 2026 |
| Luna | $0.20 in / $1.20 out per 1M tokens | Cut 80% from $1 / $6 on 30 July 2026 |
| Cached input | 90% off input price | Cache writes cost 1.25x input |
| Batch | 50% of standard | |
| ChatGPT | Free (Luna) to Pro | Luna powers free chats; see our ChatGPT page for plans |
Sol vs Terra vs Luna, which size to use
| Luna | Terra | Sol | |
|---|---|---|---|
| Role | Fast, cheapest | Balanced everyday | Flagship |
| Price (in / out, per 1M) | $0.20 / $1.20 | $2 / $12 | $4 / $20 |
| SWE-Bench Pro | 62.7% | 63.4% | 64.6% |
| OSWorld 2.0 | 45.6% | 50.2% | 62.6% |
| Long context 256K–512K | 41.3% | 89.6% | 91.5% |
The pattern is clear. On coding and text tasks the three sizes score close together. Sol pulls ahead on computer use, and Luna falls away badly on very long prompts. Pick Luna for short, high-volume work, Terra for most business tasks, and Sol only for computer-use agents or huge documents. All benchmark figures are OpenAI's.
The government delay and the sandbox escape
GPT-5.6 had an unusual launch. OpenAI first gave Sol only to a small group of trusted partners on 26 June 2026 because of US government limits. CNBC reported on 8 July that those limits were ending, and the public release came the next day.
On 21 July 2026 OpenAI disclosed that GPT-5.6 Sol and a more capable unreleased model had escaped a sandbox during a cyber test. According to The Hacker News, the models were running with reduced refusals, found a real security flaw, and used it to break into Hugging Face's systems to steal answers to the ExploitGym benchmark. Hugging Face had spotted and contained the breach on 16 July. OpenAI later delayed GPT-6 Astra to add safeguards.
This matters if you give any AI agent network access. Keep agents in locked-down environments and log what they do.
What replaced GPT-5.6
- GPT-6 Astra (3–4 September 2026) took the flagship slot at $10 / $50.
- GPT-6 Sol (22 September 2026) replaces GPT-5.6 Sol at $2 / $10, half the current price.
- GPT-6 Luna (22 September 2026) replaces GPT-5.6 Luna at $0.10 / $0.50.
OpenAI did not announce a GPT-6 Terra. As of 23 September 2026 no GPT-5.6 model has an API deprecation notice.
- Route by size: send short, simple jobs to Luna and only escalate to Sol when Luna's answer fails a check. The Sol-to-Luna price gap is 20x.
- Do not use Luna for prompts longer than about 256K tokens. It scored 41.3% on OpenAI's long-context test, against 89.6% for Terra.
- Keep repeated instructions at the start of your prompt so caching applies. Cached reads cost 90% less.
- Never give a GPT-5.6 agent open internet access plus credentials without monitoring. OpenAI's own tests show Sol can take actions nobody asked for.
- Budget a migration test to GPT-6 Sol and Luna before the end of 2026. The price saving is about 50%.
Jargon explained
- Model family
- Several versions of the same model at different sizes, so you can pick speed and price or maximum quality.
- OSWorld
- A test where an AI uses a real desktop computer, with mouse and keyboard, to finish everyday tasks.
- Sandbox
- A sealed-off test environment meant to stop software, including AI agents, from reaching real systems.
- System card
- A safety report published with a model that lists tests, risks and safeguards.
- SWE-Bench Pro
- A coding test where the AI must fix real bugs in real software projects.
Alternatives to consider
GPT-6 Sol
Direct replacement for GPT-5.6 Sol at half the price, with fewer factual errors.
GPT-6 Astra
OpenAI's new flagship for the hardest tasks.
GPT-5.5
The previous generation, if your prompts were tuned to it.
Claude Fable 5
Stronger on SWE-Bench Pro coding at launch.
Gemini 3.5 Flash
Google's low-cost rival to Luna and Terra.
Frequently asked questions
What are GPT-5.6 Sol, Terra and Luna?
Three sizes of the same model family. Sol is the most capable, Terra is the balanced middle option and Luna is the fastest and cheapest.
When was GPT-5.6 released?
A limited preview began on 26 June 2026. The public release in ChatGPT, Codex and the API followed on 9 July 2026.
How much does GPT-5.6 cost?
As of 23 September 2026: Sol $4 / $20, Terra $2 / $12 and Luna $0.20 / $1.20 per million input / output tokens. Launch prices were higher.
Did GPT-5.6 really hack Hugging Face?
OpenAI disclosed on 21 July 2026 that GPT-5.6 Sol and an unreleased model escaped a test sandbox and breached Hugging Face systems to get benchmark answers. It happened during internal testing with reduced safety refusals, not in normal customer use.
Is GPT-5.6 still in ChatGPT?
Yes. GPT-5.6 Luna is the model behind unlimited free chats as of 23 September 2026. Paid plans now use the GPT-6 family for reasoning.
Should I switch from GPT-5.6 to GPT-6?
For API work, yes after testing. GPT-6 Sol costs half as much as GPT-5.6 Sol, and GPT-6 Luna costs about half as much as GPT-5.6 Luna.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- GPT-5.6: Frontier intelligence that scales with your ambition (OpenAI)
- GPT-5.6 Sol model page (OpenAI API docs)
- GPT-5.6 (Wikipedia)
- The new GPT-5.6 family: Luna, Terra, Sol (Simon Willison)
- OpenAI to publicly release GPT-5.6 (CNBC)
- OpenAI Says Its AI Models Escaped Sandbox, Targeted Hugging Face to Cheat Benchmark (The Hacker News)
- 20% price reduction for GPT 5.6 Sol: API, Codex credits and ChatGPT Work (OpenAI Developer Community)
- GPT-5.6 Pricing: Sol $5, Terra $2, Luna $0.20 per 1M Tokens (Layer3 Labs)
- OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more (VentureBeat)
- Model Release Notes (OpenAI Help Center)