GPT-5.6 review in short: The flagship Sol scores 88.8% on Terminal-Bench 2.1 (Claude Fable 5: 83.1%) and 80 on the Artificial Analysis Coding Agent Index, at half Fable 5’s input price ($5 vs. $10 per million tokens at launch). Fable 5 still leads on SWE-Bench Pro (80% vs. 64.6%).
Our verdict: Sol is the more economical agent model; Fable 5 has the higher ceiling. Since September, GPT-6 has taken over for new projects.
Latest Changes at a Glance
GPT-6 Astra launched in early September 2026 at $10 input / $50 output per million tokens (CloudZero).
GPT-6 Sol ($2 / 10)andLuna(0.10 / $0.50) followed on September 22. They replaced GPT-5.6 Sol and Luna at about half the price (Yotta Labs). There is no Terra tier in GPT-6.
GPT-5.6 got cheaper too: Luna fell 80% and Terra 20% on July 30, and Sol was cut by more than 20% on August 21 for three months (OpenAI).
The benchmarks below are launch figures from July 2026. For the successor, read our GPT-6 Sol and Luna assessment.
What Is GPT-5.6 (ChatGPT 5.6)?
GPT-5.6 is the model family. OpenAI released it on July 9, 2026 across ChatGPT, Codex, and the API. It is commonly known as “ChatGPT 5.6”. Instead of one model, there are three tiers:
Sol for the hardest coding, research, and agent work
Terra for everyday and high-volume enterprise use
Luna for speed and low cost
The theme is “performance per dollar”: more useful output from every token used. It follows GPT-5.5, and OpenAI says all three tiers beat it on most tests.
GPT-5.6 Benchmarks: How Good Is It?
Benchmark | Sol | Terra | Luna | Claude Fable 5 |
Terminal-Bench 2.1 | 88.8% | 87.4% | 84.7% | 83.1% |
Coding Agent Index | 80 | 77.4 | 74.6 | 77.2 |
Intelligence Index v4.1 | 58.9 | 55.0 | 51.2 | 59.9 |
SWE-Bench Pro | 64.6% | 63.4% | 62.7% | 80% |
Agents’ Last Exam | 52.7% | 50.4% | 50.3% | 40.5% |
BrowseComp | 90.4% | 87.5% | 83.3% | 84.4% |
ExploitBench | 73.5% | 52.9% | 33.2% | – |
Launch figures, July 2026, mostly vendor-reported (OpenAI).
Coding: Sol’s Ultra mode (four agents) reaches 91.9% on Terminal-Bench. OpenAI says Sol beats Fable 5 on the Coding Agent Index with under half the output tokens.
Agents: Sol’s Agents’ Last Exam score appears as 52.7% in OpenAI’s table and 53.6% in its text, and the page doesn’t state the reasoning setting behind each. Treat it as roughly 53. On BrowseComp, standard Sol scores 90.4% and Ultra 92.2%.
Cybersecurity: 73.5% on ExploitBench against 47.9% for GPT-5.5. Full capability sits behind the Trusted Access for Cyber program. See our OpenAI Daybreak breakdown.
Reading the table: Fable 5 leads SWE-Bench Pro and the Intelligence Index. Sol’s edge is efficiency, since it gets close to or ahead on most tests with a fraction of the tokens.
GPT-5.6 Pricing: What Does the API Cost?
Prices per million tokens (input/output):
Model | At launch (July 9) | Current |
Sol | $5 / $30 | Cut by 20%+ on Aug 21 for three months; one source lists $4 / $20 |
Terra | $2.50 / $15 | $2 / $12 |
Luna | $1 / $6 | $0.20 / $1.20 |
These come from OpenAI and CloudZero. Check OpenAI’s pricing page before you set a budget, because the Sol discount is temporary, and prompts over 272K tokens are billed at a higher rate.
Prompt caching is now predictable: explicit cache breakpoints, a 30-minute minimum cache life, writes at 1.25x the input price, and a 90% discount on reads. Agents that resend a long system prompt save the most. Read the better prompt caching in GPT-6 article for better understanding.
Is GPT-5.6 Free?
There is no free API tier. In ChatGPT, the Free and Go users get Terra in ChatGPT Work and Codex. Plus, Pro, Business, and Enterprise users can pick Sol, Terra, or Luna and set the reasoning effort.
GPT-5.6 Sol: The Flagship in Detail
Sol is for anything difficult, like large codebases, multi-step research, and security audits. For bulk work, Terra or Luna is usually enough. For developers, the new features are:
Programmatic Tool Calling: The model writes small programs that coordinate tool calls and filter big outputs, saving tokens. It works with Zero Data Retention.
Ultra mode: Four agents in parallel by default; in beta in the Responses API. More tokens, faster results.
Reasoning effort: Standard, high, max, and ultra.
ChatGPT Work runs on GPT-5.6. It pulls context from Slack, Notion, Google Drive, and Microsoft 365 and builds spreadsheets, decks, and documents. Codex now lives in a unified ChatGPT desktop app. Anthropic’s counterpart is covered in our Claude Cowork deep dive.
GPT-5.6 vs. Claude Fable 5 vs. GLM 5.2 vs. DeepSeek V4
Particulars | GPT-5.6 Sol | Claude Fable 5 | GLM 5.2 | DeepSeek V4 Pro |
Release | Jul 9, 2026 | Jun 9, 2026 | Jun 2026 | Apr 24, 2026 |
License | Closed / API | Closed / API | Open weights (MIT) | Open weights (MIT) |
API price per 1M (in / out) | $5 / $30 at launch | $10 / $50 | Low cost | $0.435 / $0.87 |
Intelligence Index v4.1 | 58.9 | 59.9 | about 51 | about 44 |
Self-hosting | No | No | Yes | Yes |
Fable 5: Highest ceiling for complex software work, at double the input price. Sol matches or beats it on many agent tasks with fewer tokens. See our Claude Fable 5 guide and Claude Opus 5 review.
GLM 5.2: The strongest open-weight challenger, level with Luna on the Intelligence Index but weaker on the hardest reasoning. Read our GLM 5.2 article.
DeepSeek V4: The cost leader for self-hosting and bulk work, behind Sol on long agent tasks. See our DeepSeek V4 write-up.
Which Model Should You Choose?
Pick Sol for frontier agent work in OpenAI’s ecosystem, Terra or Luna for high volume, Fable 5 for the highest ceiling, and GLM 5.2 or DeepSeek V4 for self-hosting and data residency. Not sure which fits? Our AI consulting team tests models on your data, and our AI agency builds the solution. For budgeting, read our article on what AI consulting costs in Germany.
GPT-5.6 or GPT-6: What Should You Use Now?
Start with GPT-6 Sol or Luna if you’re starting fresh. OpenAI says GPT-6 Sol makes about half as many factual mistakes as GPT-5.6 Sol, and both cost roughly half as much. GPT-5.6 stays relevant for existing Terra workflows and for comparing summer 2026 results. For the cyber side, see our GPT-6 Astra analysis.
The Verdict: Is GPT-5.6 Worth It?
GPT-5.6 doesn’t win every benchmark. Fable 5 still edges it out on some of the hardest coding tests. What it changes is the cost equation: “good enough, much cheaper and faster” beats “slightly better at twice the price” for most business workloads. Today, new users should test GPT-6 Sol first and keep GPT-5.6 as the reference point.
GPT-5.6 FAQ
Is ChatGPT 5.6 the same as GPT-5.6?
Yes. “ChatGPT 5.6” is the everyday name for GPT-5.6, OpenAI’s model family from July 9, 2026, used in ChatGPT Work, Codex, and the API.
How much does GPT-5.6 cost?
At launch, Sol cost $5/$30, Terra $2.50/$15, and Luna $1/$6 per million tokens. Terra and Luna have since been cut, and Sol is discounted for a limited time.
What is GPT-5.6 Sol?
The most capable tier, tuned for coding, research, and agent tasks. It scores 88.8% on Terminal-Bench 2.1 and 80 on the Coding Agent Index.
How many parameters does GPT-5.6 have?
OpenAI hasn’t published a parameter count, so there is no reliable figure.
Is GPT-5.6 better than Claude Fable 5?
It depends on the task. Fable 5 leads SWE-Bench Pro and the Intelligence Index, while Sol costs half as much on input and matches or beats it on many agent tests.
Is GPT-5.6 available in Germany?
Yes. GPT-5.6 is available in Germany, subject to your ChatGPT plan or API access and the availability of the specific model in your product. Germany is within the European Economic Area (EEA), which OpenAI lists among supported European regions.
What is Ultra mode in GPT-5.6?
The highest setting. It runs four agents in parallel by default for tougher tasks, using more tokens.