GPT 5.6 Launching SOON! (What to Expect)
Video summary
Companion notes
GPT-5.6 launches Thursday as a three-model family — Soul, Tara, Luna
The lineup
OpenAI is dropping GPT-5.6 publicly on Thursday, and it ships as three models, not one: Soul (flagship reasoning and long-horizon agents), Tara (balanced daily driver), and Luna (fast and cheap, for scale-heavy background work). Pricing per million tokens: Soul at $5 input / $30 output, Tara at half that, Luna at $1 input / $6 output. Creators running Codex-style stacks should think Opus/Sonnet/Haiku — pick by role, not by default.
How it actually feels
Early tester Theo calls it "very good" and "incredibly capable," saying it "fixed all the problems I had with GPT-5.5" and that "5.6 soul is world-leading in computer use. It made me use it 100 times more." He notes it runs agent loops for a full day even without /goal, and that sub-agent orchestration and iOS/UI-UX work feel natural. Skirano, testing for months, called it the "best model he's ever used" and said front-end design issues are finally fixed. Creator Ron hedges that, on plan-mode reasoning, "Fable on plan mode is still really good" — so Soul is the orchestrator brain, not an automatic replacement.
Reasoning modes: max vs ultra
Two new effort tiers: max (more thinking time) and ultra (sub-agents under the hood for decomposition and parallel work). Use max only on the first planning prompt; per-token max or X-high on every subsequent step is wasted spend. Use ultra when a single prompt covers many components and you want the model to fan out.
Pricing and routing by domain
Build a tiered pipeline: Luna for intake/routing, Tara for execution, Soul only for ambiguous, long-horizon, or first-prompt planning on ultra. Assume cyber, bio, exploit, or malware-adjacent work gets heavier safeguards, slower outputs, and possible fallback to GPT-5.5 — separate defensive debugging agents from anything that could read as offensive security work.
Watch for the Fable 5 squeeze
Anthropic extended the Claude Fable 5 paid-plan deadline to Sunday, the same week as the OpenAI drop. If GPT-5.6 lands as a true flagship, expect Fable 5 to stay bundled in paid plans — OpenAI likely cannot afford to pull the replacement while Codex is still ramping.
Watch on YouTube
Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=BRThrYEiccA
