OpenAI GPT-5.5 is now selectable in the Dawn model catalogue. Pick it on any agent, reasoning support is enabled by default, but a request answered by GPT-5.5 uses three times the credits of the same request on GPT-5.4 — a cost worth knowing about up front, although easily offset by the quality of answers.
Nothing about your existing agents changes. GPT-5.5 simply joins the list alongside the rest of the OpenAI lineup, and the per-agent model picker is where you reach for it on the work that justifies the higher reasoning depth and cost.
Where GPT-5.5 Earns Its Keep
The model is a fit for the steps where Dawn is already doing the most thinking. Architectural reviews where the agent has to weigh competing approaches across product, engineering, and delivery constraints. Multi-step tool sequences where the agent has to plan, call, observe, and adjust — the kind of run where each turn’s quality compounds into the next. Long-running explore runs where the agent is gathering context across multiple connected sources before composing a response however may become costly and are usually best left to the mini series of models.
It is also a fit for outputs whose shape matters as much as their content: structured plans with explicit dependencies, decision documents with clear tradeoffs, code reviews that have to hold the whole change in mind. These are the cases where a faster model produces a fluent answer that misses a constraint two paragraphs up; GPT-5.5’s reasoning support makes those misses noticeably rarer, at the cost of taking longer and using more tokens per response.
The corollary is that it is not a fit for everything. A short DM in Slack asking Dawn to summarise a Linear ticket does not need it. A connected-source query that comes back with three lines of JSON and an explanation does not need it. Pointing GPT-5.5 at those uses tokens without buying anything; the cheaper models in the catalogue answer them just as well.
The pattern most teams settle into is a split: a smaller, faster model as the workspace default, and a deliberate choice of GPT-5.5 on the agents whose job involves the kind of work above. The catalogue is built to support that split rather than force a single workspace-wide choice.
Selecting It for an Agent
GPT-5.5 appears in the OpenAI section of the model catalogue alongside the rest of the OpenAI lineup. Open the agent you want to upgrade, change its model, save — the next message that agent answers uses the new model. There is no migration step, no separate enable toggle, and no prompt rewrite required. The agent’s existing system prompt, skills, memory pages, and connected sources all carry over unchanged.
For workspaces using bring-your-own-key, GPT-5.5 picks up the same OpenAI key the rest of the catalogue uses; no additional configuration is needed beyond the OpenAI account having access to the model. Workspaces using Dawn’s hosted OpenAI provider see the new entry appear automatically.
A workspace can run several agents on different models at the same time. The triage agent answering quick questions in a shared channel can stay on a faster model; the engineering-review agent that operates on pull requests can move to GPT-5.5; the customer-comms agent drafting outage updates can sit somewhere in between. Each agent makes its own choice, and the cost reporting follows the choice rather than the workspace.
Reasoning Support, And Why It Is On
GPT-5.5 is one of the OpenAI models with explicit reasoning support — an internal “think before answering” step where the model works through the problem before composing its response. In Dawn, reasoning is on by default for GPT-5.5, because the agents most likely to be pointed at this model are the ones whose work most benefits from it.
The visible effect is small: the response takes a few seconds longer to begin, but the answer is more often right the first time. The invisible effect is larger: on the long tool-using runs Dawn agents do, reasoning between tool calls means the agent picks the right next call more often, which compounds across a long run. The places this shows up most are explore runs, multi-step plan execution, and any agent with several tools in its repertoire.
Reasoning is not free, in either latency or credits. The next section spells out what that means in practice.
Three Times the Credits of GPT-5.4
Dawn’s cost reporting works in workspace credits, which is the same unit the rest of the catalogue is measured in. A request answered by GPT-5.5 consumes roughly three times the credits of the same request answered by GPT-5.4. That ratio reflects both the model’s higher per-token price and the additional tokens the reasoning step uses internally before composing the response.
The practical implication is the obvious one: GPT-5.5 is meaningfully more expensive per response, so picking it on agents that handle high-volume short-form work will cost noticeably more than the same agents on GPT-5.4. The flip side is the one this whole feature is built around — on the work where reasoning is what makes the answer correct, paying three times the credits for an answer that lands the first time is cheaper than paying one credit for a wrong answer plus another round of human or agent time to fix it.
Cost shows up in the same per-agent and per-channel views the rest of the workspace already uses. The GPT-5.5 line appears alongside the others, with no separate “advanced model” view to learn — the higher rate just means the line is taller for the same number of requests.
Adopting It Without Losing the Plot
The friction with model upgrades is rarely technical. It is that the team has to decide where to use the new model, validate that the responses are actually better in their workflows, and keep a hand on the cost. The catalogue and per-agent selection support that loop directly.
The pattern that has worked best in early use is to pick one or two agents whose work most plausibly benefits — usually a planning or review agent — and move them to GPT-5.5 first. Run them on real work for a few days. Compare the responses against the previous model on a handful of representative tasks. With the 3x credit cost in mind, the bar for “worth keeping” is not just better answers — it is enough better that the higher per-request cost pays for itself in fewer revisions, fewer follow-up runs, or fewer human hours. If the quality clears that bar, expand to other agents whose work has a similar shape; if it does not, the rollback is the same one-line model selection that promoted them in the first place.
What Comes With It
The release is exactly what the title says: GPT-5.5 in the catalogue, reasoning support enabled, billed at three times the credit rate of GPT-5.4. The rest of Dawn is unchanged. Existing agents keep their existing models. Existing connected sources, skills, and memory pages keep working with the new model the same way they worked with the old. There is no breaking change to migrate around and no feature flag to flip.
If your team has been waiting to evaluate GPT-5.5 in real day-to-day use, the path is now: open the agent you want to try it on, change the model, and use the agent on work where reasoning depth is the bottleneck. Report back when the responses tell you whether it was worth the move.