# OpenAI Previews GPT-5.6 Sol in Codex, Timed Against Anthropic

Leaked text strings and an official preview point to a Sol, Terra, and Luna family, with a top-tier "Sol Ultra" version leading command-line coding benchmarks.

- Published: 2026-07-06T10:57:47.986Z
- Canonical: https://polylog.news/ai/2026-07-06/openai-previews-gpt-5-6-sol-in-codex-timed-against-anthropic
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Hacker News / X (Sottiaux)](https://x.com/thsottiaux/status/2073933490513752151), [OpenAI](https://openai.com/index/previewing-gpt-5-6-sol/), [TestingCatalog](https://www.testingcatalog.com/openai-might-be-preparing-gpt-5-6-for-next-weeks-release/)

OpenAI has begun a limited [preview of GPT-5.6](https://openai.com/index/previewing-gpt-5-6-sol/), a family it describes as Sol (flagship), Terra (balanced), and Luna (fast and low cost). It is available first through the application programming interface (API) and Codex to selected partners. A post that appeared on Hacker News noted that [GPT-5.6 Sol Ultra will be in Codex](https://x.com/thsottiaux/status/2073933490513752151), and application text strings had earlier exposed the "Sol Ultra" label before any formal announcement.

On Terminal-Bench 2.1, reporting puts Sol Ultra at 91.9% and plain Sol at 88.8%, ahead of GPT-5.5 at 88.0% and Claude Fable 5 at 83.4%. The rollout timing, days before an expected Anthropic release window, has been interpreted by trade press as a [deliberate competitive move](https://www.testingcatalog.com/openai-might-be-preparing-gpt-5-6-for-next-weeks-release/) on the coding tasks where the two labs compete most directly.

Caveats apply. The benchmark figures circulating ahead of general availability come from leaks and vendor-adjacent posts, not from a published system card with reproducible methodology. Terminal-Bench compares scaffolds as much as models, so rankings across labs shift with harness choices.

## What this means

Both leading labs are now releasing coding-optimized flagship models within days of each other and routing them first through developer tools such as Codex, which confirms that agentic coding is the main area of competition. For teams building on these APIs, the pace of model change is now measured in weeks, which raises the value of evaluation harnesses that let teams swap models without rewriting their agents.

## What to watch

- Publication of a GPT-5.6 system card with reproducible benchmark methodology, which would move the numbers from leak to verified.
- Whether Sol Ultra's pricing undercuts or matches Anthropic's, the factor that decides which model teams use by default for agent workloads.