What happened
On 9 July 2026 OpenAI began the public rollout of GPT-5.6 — in three variants at once:
- Sol — the flagship for hard work: coding, research, cybersecurity, computer use, design. There is also a Sol Pro mode for the heaviest, longest tasks.
- Terra — the balanced everyday "default": a good mix of quality, speed and price.
- Luna — the fastest and cheapest option for small, high-volume tasks.
Before the public release the model spent about two weeks in a closed preview (from late June) with a limited set of organizations. GPT-5.6 is now rolling out gradually across ChatGPT plans, shipping in the API, and landing in the agentic product ChatGPT Work and in Codex. If Sol has not shown up in your model list yet — access is being opened in waves.
Pricing per 1M tokens (input/output): Sol — $5 / $30, Terra — $2.50 / $15, Luna — $1 / $6. Terra has roughly a 1.05M-token context and up to 128K tokens of output.
One loud detail on its own: Sol delivers up to 750 tokens per second on Cerebras wafer-scale chips. That is roughly 5x faster than typical production models (~150 tok/s). The fast configuration is rolling out in a limited way — partners first, as capacity grows. Estimates of 70–100 WSE-3 wafers circulate in the industry, but OpenAI and Cerebras have never officially confirmed the topology — that is analyst guesswork, not a vendor fact.
Why it matters
- The team default just got twice as cheap. Terra costs exactly half of standard GPT-5.5 ($5 / $30 per 1M tokens) and, by measurements, delivers GPT-5.5-class quality. So anyone who ran everything through GPT-5.5 out of habit gets comparable results for noticeably less.
- Speed becomes a selection criterion. Sol's 750 tok/s is not about a pretty benchmark, it is about agent latency: a chain of dozens of tool calls that takes seconds on ordinary GPUs compresses several times over.
- Fewer tokens spent on tools. The Responses API gained programmatic tool calling — the model writes JavaScript that runs in an isolated V8 sandbox with no network; for a number of customers this cut tokens by 38–63.5%.
- Coding improved. According to Artificial Analysis, Sol tops the Coding Agent Index with a score of 80 and takes 88.8% on Terminal-Bench 2.1 (up to 91.9% in ultra mode with four agents). Terra scores 87.4% on the same Terminal-Bench 2.1 versus 85.6% for GPT-5.5.
Be careful with the hardware numbers: the 750 tok/s figure comes from OpenAI itself, but the Cerebras chip count is an outside estimate that has not been officially confirmed.
What it means for vibe coding
There is now a clear traffic-light system for picking a model per task, and it directly affects your API bill:
- Luna — autocompletions, small edits, cheap high-volume runs.
- Terra — the workhorse: everyday interactive and agentic coding. The logical default if you were sitting on GPT-5.5.
- Sol / Sol Pro — heavy lifting: refactoring large codebases, long agentic runs, hard research.
A practical win for anyone coding with AI: all three models are available in GitHub Copilot from day one (Sol on Pro+, Max, Business, Enterprise; Terra and Luna starting with Pro) and in Codex. You can pick them from VS Code, Visual Studio, JetBrains, Xcode, Copilot CLI and GitHub.com.
The "cheaper + faster" combination hits agentic scenarios hardest: when a single task means 30–40 model calls with tools, Terra's price savings and Sol's speed add up to a genuinely different experience — the "edit → run → fix" loops get shorter and cheaper. If you are building your own coding agents, it is worth reworking your routing: cheap Luna for grunt work, Terra by default, Sol targeted at the hard parts.
Sources
- OpenAI — Previewing GPT-5.6 Sol
- OpenAI — GPT-5.6
- MarkTechPost — OpenAI Releases GPT-5.6 (Sol, Terra, Luna)
- GitHub Changelog — GPT-5.6 Sol, Terra and Luna in GitHub Copilot
- TestingCatalog — OpenAI launches GPT-5.6 Sol, Terra and Luna
- OpenRouter — GPT-5.6 Terra: pricing and benchmarks
- we0.ai — GPT-5.6 Sol on Cerebras: 750 tokens per second