# Alleged OpenAI codenames Zinc and Magnesium surface in Design Arena > Source: > Published: 2026-07-28 02:24:31+00:00 Two disabled model entries labeled "Zinc" and "Magnesium" appeared in Design Arena on July 28th, according to model watcher [Lentils (@Lentils80)](https://x.com/lentils80/status/2081906160278352358?s=46), raising the possibility that OpenAI is already testing revisions to its GPT-5.6 family. Lentils described both entries as GPT models and suggested Zinc may be an update to GPT-5.6 Sol while Magnesium may be tied to GPT-5.6 Terra. The post did not include test results, model identifiers or evidence connecting the codenames directly to OpenAI. Neither model was enabled for public battles at the time of the sighting. The entries remain unconfirmed. [OpenAI's GPT-5.6 launch materials](https://openai.com/index/gpt-5-6/) identify three public tiers: Sol, the flagship; Terra, a lower-cost model positioned against GPT-5.5; and Luna, the fastest and least expensive option. The materials do not identify Zinc or Magnesium. OpenAI did, however, describe Sol, Terra and Luna as "durable capability tiers" that can advance on separate schedules. That structure gives OpenAI room to update the models behind those names without waiting for a full GPT-5.7 release. If the Design Arena entries are authentic, they could represent new post-training versions, internal evaluation builds or routing aliases intended to remain under the existing Sol and Terra product tiers. A disabled entry provides only an early signal. It does not establish that either model will ship, retain its codename or become a standalone product. ### A startup benchmark becomes a model-release signal The sighting places [Design Arena](https://www.designarena.ai/) in an increasingly influential position between frontier model labs and users. [Grace Li](https://www.linkedin.com/in/grace-li-721a4017b/) and [Kamryn Ohly](https://www.linkedin.com/in/kamrynohly/), Harvard classmates who previously worked at Apple, founded the San Francisco operation in 2025 and took it through [Y Combinator's Summer 2025 batch](https://www.ycombinator.com/companies/design-arena). Li and Ohly built Design Arena around a problem conventional AI benchmarks largely ignore: whether generated websites, images, slides and other creative outputs actually look good to people. Users submit a prompt, receive anonymous outputs from multiple models and rank them through [head-to-head comparisons](/article/qwen-3-7-max-benchmarks-pressure-closed-models). Those votes feed a Bradley-Terry ranking system, according to [Design Arena's published methodology](https://www.designarena.ai/about). That model pool creates a natural staging ground for labs testing unreleased systems. Model identities remain hidden during voting, and Design Arena says four active models plus a backup are sampled for each session. A model can therefore be evaluated without its commercial name appearing in front of voters. A configuration record surfacing before activation may expose an internal label even when the underlying model remains unavailable. Design Arena is the flagship product of [The Intelligence Company](https://www.intelligence.ai/), the public brand adopted by Arcada Labs. The Intelligence Company says Design Arena has surpassed 3 million users across more than 190 countries. Y Combinator's launch page recorded a much earlier milestone of 47,000 users across 136 countries during its first four weeks. The benchmark already has a direct connection to the GPT-5.6 rollout. In a [July 15th analysis](https://notes.designarena.ai/how-openais-sol-finally-learned-design-taste/), Design Arena ranked GPT-5.6 Sol first in its non-agentic web design category, 18 positions above GPT-5.5. Design Arena attributed the improvement partly to Sol suppressing common AI-generated design patterns such as purple gradients, oversized hero text and repetitive bento-box layouts. ### Claude Opus 5 adds pressure on design and price Lentils suggested the alleged models could be a response to Anthropic's [Claude Opus 5](https://www.anthropic.com/news/claude-opus-5), released on July 24th. The four-day gap makes the theory chronologically possible, though the listings provide no evidence of a direct connection. Anthropic positioned Opus 5 as a daily-use model with stronger coding, computer-use and visual-output performance. Anthropic priced it at $5 per million input tokens and $25 per million output tokens. OpenAI charges the same $5 input rate for GPT-5.6 Sol and $30 for output, while Terra costs $2.50 for input and $15 for output. OpenAI released GPT-5.6 for general availability on July 9th, less than three weeks before the Zinc and Magnesium sighting. An update this soon would fit the faster model cadence that the Sol, Terra and Luna tiering system was designed to support. It would also show how quickly frontier labs are being pushed to refine models after launch as competitors attack the same combination of coding ability, visual judgment, speed and cost. For now, Zinc and Magnesium are labels inside a third-party benchmark rather than announced OpenAI products. Their significance depends on whether they enter Design Arena's active pool and produce enough public comparisons to reveal what, if anything, changed from Sol and Terra.