GPT-5.6 ships as three models, and that makes model choice a deployment decision

OpenAI has split its newest generation into three models rather than shipping one general-purpose release: Sol for advanced professional work, Terra for a balance of capability and cost, and Luna for cost-sensitive, high-volume tasks, generally available since July 9, 2026 ^[1][2][3]. That structure changes the nature of the decision in front of engineering teams, because there is no longer a single "newest model" to point production traffic at ^[8].

The commercial signal arrived fast. A pricing update on July 30, 2026 cut Luna by around 80% and Terra by around 20%, according to a dev.to write-up of OpenAI's announcements ^[5]. That is 21 days between general availability and a repricing of two of the three tiers ^[13]. Arithmetically, an 80% reduction means a fixed monthly API budget buys roughly five times the Luna volume it bought at launch, while the Terra cut buys about 1.25 times ^[14]. If your unit economics were the reason a workload stayed on a cheaper or older model, that calculation has moved, and it moved after you would have finished your launch-week evaluation.

The second half of this is retirement. OpenAI has signalled the phase-out of GPT-4o and related GPT-4.x variants as customers move to GPT-5.x and GPT-5.6 ^[6]. Deprecation is the part that generates real work: the source notes a retirement path can affect application behaviour, cost forecasting, internal approval processes and the technical effort to update production systems ^[9]. None of that is optional in the way a version bump is.

What is missing from the public picture matters as much as what is in it. The material carries no benchmark data, no context-window specifications and no detailed API pricing, which means the tier names are marketing segmentation and not a substitute for measurement ^[7]. OpenAI frames the generation as a higher-performance foundation for ChatGPT and for API use cases involving agents and coding ^[4], but the source is explicit that this does not establish identical results for every agent or coding implementation ^[11]. Agentic and coding workloads are exactly where behavioural drift between models shows up as broken tool calls and silent quality regressions, so treat tier selection as an experiment with a control.

The practical review the source recommends is unglamorous and correct: inventory production dependencies tied to older GPT-4 models, segment workloads so advanced, balanced and high-volume tasks can be tested against Sol, Terra and Luna respectively, revisit cost controls in light of the July reductions, and put change management in front of any migration that touches customer-facing or business-critical paths ^[10]. The strategic read is portfolio simplification across ChatGPT and API surfaces, which may ease procurement over time but raises the value of knowing exactly which model each of your systems calls ^[12].

Watch whether the next repricing moves Sol, which so far has not been discounted ^[5], and whether OpenAI publishes firm retirement dates for the GPT-4.x line rather than a signalled phase-out ^[6]. Also watch your own routing logic: a fivefold shift in effective Luna volume ^[14] is enough to justify re-running a build-versus-buy comparison that looked settled in June.