Exactly one month after Claude Fable 5's debut, OpenAI plays its own hand of three. On July 9, 2026, the GPT-5.6 family arrived — three models named Sol, Terra, and Luna — alongside ChatGPT Work, an agent built to complete entire workflows, not just answer questions.
In short: GPT-5.6 is OpenAI's new family of models, available in three versions — Sol (top-tier), Terra (mid-tier), Luna (fast and affordable) — on ChatGPT, Codex, and the API. It's the direct answer to Anthropic's Claude Fable 5 and Mythos 5, but the benchmarks say neither contender truly dominates.
What Sol, Terra, and Luna Are
Sol, Terra, and Luna are the three versions of OpenAI's GPT-5.6 family, differentiated by power and price: Sol is the flagship model, Terra the mid-tier for everyday tasks, Luna the fast, low-cost option.
All three share the same base specs:
- A 1-million-token context window, with output up to 128,000 tokens
- Knowledge current as of February 16, 2026
- Availability on ChatGPT, Codex, and the API, with Sol reserved for Plus, Pro, Business, and Enterprise users
API prices per million tokens follow the scale of the three tiers:
| Model | Positioning | Input (1M tokens) | Output (1M tokens) |
|---|---|---|---|
| Sol | flagship | $5 | $30 |
| Terra | mid-tier | $2.50 | $15 |
| Luna | fast/affordable | $1 | $6 |
The three-tier strategy mirrors the one Anthropic has long used with its model families: a top-tier model for complex tasks, a balanced model for everyday work, and a lightweight one for high volume.
The Answer to Anthropic — With a Forced One-Month Delay
The timing isn't a coincidence. Anthropic had launched Claude Fable 5 and Mythos 5 on June 9, 2026, but three days later the US Department of Commerce imposed emergency export controls, forcing the company to suspend both models globally for 19 days. Only on July 1, after the controls were lifted, did Fable 5 become available worldwide again.
OpenAI chose the opposite path: a rollout coordinated with Washington. According to reports, the company had given a small group of American partners an early look at GPT-5.6 as far back as late June, at the government's request, and the public launch came after testing and meetings with administration officials. The reason for the caution is the same one that had overtaken Anthropic: these models have an unprecedented ability to find software vulnerabilities, a double-edged capability that worries cybersecurity experts.
The Benchmarks: No Clear Winner
The "overtaking" narrative only holds up halfway. The numbers say the contest between GPT-5.6 and Anthropic's models is the most evenly matched in years:
- Agents' Last Exam: Sol sets a new high of 53.6, beating Claude Fable 5 by 13.1 points
- Artificial Analysis Coding Agent Index: Sol leads at 80, 2.8 points above Fable 5
- SWE-Bench Pro (coding on real repositories): here Anthropic wins clearly — 80.3% versus 64.6% for Sol
- Artificial Analysis Intelligence Index v4.1: Fable 5 stays ahead, 59.9 versus 58.9
A curious detail: alongside the launch, OpenAI published an analysis questioning the validity of SWE-Bench Pro — precisely the benchmark where it loses — claiming that about 30% of the tasks are compromised. Simon Willison, one of the industry's most respected observers, noted that Sol has yet to demonstrate it beats Fable on complex coding tasks.
ChatGPT Work and What's New for Developers
Beyond the models, the launch brings ChatGPT Work, an agentic platform that integrates ChatGPT and Codex to automate complete professional tasks — from research to producing the final output.
On the API front, the most interesting developments for developers:
- Programmatic Tool Calling — the model writes JavaScript code to orchestrate its own tools, executed in an isolated V8 runtime with no network access
- Parallel subagents — the multi-agent mode pushes Terminal-Bench 2.1 from 88.8% to 91.9%
- Explicit breakpoints for the prompt cache, for better cost control
- Image processing without resizing with the "detail: original" option
Why This Launch Matters
For the first time since 2023, competition between the industry's two most important labs is playing out on equal footing and on three fronts at once: raw model power, price per token, and — above all — agents, meaning the ability to complete entire jobs instead of answering individual questions.
For those using these tools every day, the practical takeaway is simple: choosing a model is no longer a matter of brand loyalty, but of the specific task. Complex coding on real codebases? The benchmarks still favor Anthropic. Long agentic workflows and costs to optimize? The GPT-5.6 family is more aggressive than ever.
Full availability of GPT-5.6 was reached within 24 hours of the July 9 launch. The matter of export controls on frontier models, however, remains an open issue: both labs now coordinate their rollouts with the US government, a precedent that will shape upcoming launches.