Pre-retirement tool-use observation: o3-mini and o4-mini by OpenAI · V1
Model observation · V1
Why this observation matters
A dated record can preserve how a retiring model behaved in a tool-enabled system. Retaining the surrounding conditions helps teams distinguish a historical reference from an assurance about future availability or behavior.
CEO
A historical reference
The observation preserves a bounded view of model behavior before a vendor lifecycle change.
CFO
Evidence for oversight
Completion, elapsed time and correctness require different evidence; the public chart reports only recorded elapsed time.
CTO
Behavior with tools
Serving route, repeated-run conditions and tool traces help frame a technical regression review.
What we did
We observed two models in a bounded tool-use harness and repeated each model run in a later window. The same task set was used across the four runs. We retained request conditions, serving-provider evidence, execution states, timing and tool traces. This public overview selects run-level aggregates; it omits prompts, responses and task-level records.
Selected observations under these conditions
Final answers under the bounds
Every recorded task reached the harness state for a final answer in the primary and replicate runs.
Teams examining completion signals can use the records to separate reaching an answer from satisfying the task.
The completed state is a control-flow outcome; no independent answer-quality score is established by that state.
Repeated status, changing execution
Primary and replicate runs shared the completed status while tool-call counts, steps and task durations varied.
A repeated terminal state can conceal changes in the work performed by the model and surrounding system.
Only one primary and one replicate per model were recorded; these observations do not estimate long-run stability.
Recorded provider route
The retained result rows identify OpenAI as the serving provider through OpenRouter.
Provider attribution makes the access path explicit when interpreting a historical observation.
The routed model identifiers are recorded aliases, not an independently pinned provider weight snapshot; current availability was not tested.
Selected recorded aggregates
- o3-mini · primary
- 5.14 minutes total
- o3-mini · replicate
- 6.05 minutes total
- o4-mini · primary
- 10.33 minutes total
- o4-mini · replicate
- 7.91 minutes total
Source: retained o3-mini/o4-mini primary and replicate result bundles, October 14–15, 2026 UTC. n=12 task records per run. Each bar shows run-total minutes: the sum of those task durations, including model, tool and network activity. This is not per-task latency. The display axis is 0–12 minutes. Bars retain model/run order and are not a ranking. These are not token-generation rates or answer-quality scores.
Vendor status checked October 28, 2026
OpenAI announced this API deprecation on April 22, 2026. Its current record states that access to the o3-mini and o4-mini API snapshots and aliases was shut down on October 23, 2026. The observation predates that shutdown. Vendor documentation was checked on October 28, 2026; this page is a historical observation, not a statement that either route remains available.
Declared conditions
Recorded October 14–15, 2026 UTC (October 13–15 in US Eastern time). Cloud inference was reached through OpenRouter with routing restricted to OpenAI and provider fallbacks disabled. Reasoning effort was medium, requests were unseeded and bounded, and tool subprocess confinement was declared in the manifests. The 12-task set covered web, computation and local-information tasks. The recorded methodology label is a draft proposal, not an adopted-methodology claim.
Potential uses
CEO / strategy
The records could support a dated review of changes in a model-dependent workflow.
CFO / oversight
The records could support a transition review that considers completion, duration and scope separately.
CTO / engineering
Tool traces and repeated-run differences could inform internal regression-check design.
What's in the data
The retained evidence contains the following record classes. This description does not assert that every artifact is currently released through the API.
- Versioned primary and replicate records with per-task execution state, steps, tool-use traces and timestamps.
- Requested and served model/provider identifiers with route-verification evidence.
- High-level generation conditions, recorded usage metadata and bundle-integrity references.
- Preserved prompts, responses and technical artifacts under their separate access and release boundaries.