OpenAI Slips 'Ultra' Mode Into Codex — But Under the Hood It's Just Smart Prompting

Codex source code reveals GPT-5.6 Sol Ultra is a prompt-level wrapper around max effort plus subagent orchestration, not a new backend capability. The move mirrors Anthropic's ultracode dynamic workflows as the coding-agent arms race intensifies.

OpenAI's coding agent Codex is getting GPT-5.6 Sol Ultra support, but a look at the open-source code tells a different story than the branding suggests. According to analysis by developers who combed through the Codex repository, "Ultra" is not a new model tier or backend capability — it's a client-side alias for the existing max effort level, plus a single instruction directing the model to use subagents proactively.

The discovery, first surfaced by developer Tibo Sottiaux on X and confirmed by HN commenters with direct links to the Codex source, shows that Ultra mode lives entirely at the prompt layer. The relevant code paths in multi_agents.rs, client.rs, and multi_agent_mode_instructions.rs map Ultra to "xhigh" effort with a subagent activation string tacked on.

The pattern mirrors Anthropic's approach with Claude Code. Anthropic shipped "ultracode" — now called dynamic workflows — which similarly wraps the model in a harness that creates deterministic JavaScript orchestration scripts to manage subagents. Google reportedly has an equivalent internal concept called "deterministic workflows." The naming may be silly — Ultra, Ultracode, dynamic workflows — but the underlying technique is real: rather than waiting for models to get smarter at multi-step tasks on their own, both companies are building scaffolding that programs the coordination explicitly.

One HN commenter at a large US corporation reported that their org already has access to Sol Ultra but is now being told via weekly emails to switch to cheaper models whenever possible — a sharp pivot from the token-usage leaderboards management was touting two months ago. The tension between the promise of frontier models and the sticker shock of API billing is starting to reshape enterprise deployment patterns.

The naming chaos hasn't gone unnoticed. "Bruh when did understanding chatbots become like following Pokémon?" one HN user wrote. "Tf is sol? Tf is ultra? Tf is codex?" Fair questions. Sol means "Sun" (paired with Terra and Luna in OpenAI's celestial naming scheme), Ultra describes the agent harness, and Codex is the coding agent product. But as another commenter put it, branding shouldn't require a glossary.

Sources: Tibo on X, HN Discussion, Codex multi_agents.rs (GitHub), Codex client.rs (GitHub), multi_agent_mode_instructions.rs (GitHub), Claude Code Dynamic Workflows