Across 152 live Oasis agent templates, Claude models appear in 96 configurations, GPT models in 48, and 8 leave the model field unset because another coding-agent foundation owns that choice. The raw split favors Claude. It does not show that Claude is universally better.
The more useful answer sits one level down. Smaller configured tiers handle bounded, repeatable work. Higher-capability models appear when the agent must research an open question, review consequential work, or keep a long task on track. Reasoning effort and turn budget move with the job too.
That is a better way to frame ChatGPT versus Claude for agents: separate AI agents from AI assistants, choose a workload, then choose the model tier and runtime settings that fit it.
The short answer
- Claude Sonnet is the most common configured model. It covers the widest range of medium-effort work and many of the longest coding workflows.
- Claude Haiku is the clearest efficiency choice for narrow operational tasks such as triage, routing, and scheduling.
- GPT Terra is the OpenAI workhorse for structured data and operations workflows.
- GPT Sol is concentrated in research-heavy agents, usually with high reasoning and web access.
- GPT Luna handles short, low-effort digests and monitoring jobs.
Read these as configuration evidence. Output quality, latency, and total run cost fall outside the dataset.
What we analyzed
The source is the first-party export behind the live Oasis template library. Each row records the configured model, foundation, reasoning effort, maximum turns, memory setting, browsing setting, required apps, and workflow description.
I grouped exact model labels and compared them with the other configuration fields. The analysis answers a deployment question: where did each model tier look appropriate enough to ship as the default?
| Configured model | Live templates | Reasoning pattern | Configured turn range |
|---|---|---|---|
| Claude Sonnet 4.6 | 63 | 50 medium, 12 high, 1 low | 20 to 60 turns |
| Claude Haiku 4.5 | 33 | 33 low | 20 to 35 turns |
| GPT-5.6 Terra | 29 | 26 medium, 2 low, 1 high | 25 to 40 turns |
| GPT-5.6 Sol | 14 | 13 high, 1 medium | 35 to 45 turns |
| GPT-5.6 Luna | 5 | 5 low | 20 to 25 turns |
| Model unset | 8 | 8 medium | 40 to 55 turns |
There is an important confounder. 92 templates use the Claude Managed Agent foundation, 48 use OpenAI Agents, and 4 use Claude Code. The remaining templates use other coding-agent foundations and do not set a model in this dataset. Family totals therefore reflect platform architecture as well as model preference.
The comparison becomes more informative within each family.
Where Claude Sonnet is the default
Claude Sonnet 4.6 appears in 63 templates, more than any other configured model. 50 use medium reasoning. The remaining set leans high, with 12 high-effort templates and only 1 low-effort template.
The medium-effort group spans drafting, support, sales operations, content production, and cross-system coordination. The Content Repurposer, for example, gets up to 30 turns, can browse, and adapts one source across several channels. The work needs judgment and context, but it has a defined output.
High effort appears where an error is harder to undo or the agent must sustain a longer plan. The GitHub PR Reviewer uses high reasoning and up to 40 turns to inspect failure paths and authorization problems. The Framework Migration Runner reaches 60 turns, works in reviewable batches, and stops when the test suite fails.
That pattern supports a practical reading: Sonnet is the Claude default when the workflow combines judgment with several tool or review steps. It is not evidence for the old article's claim that Claude writes more naturally. The template data does not measure prose quality.
There is also a maintenance issue. Anthropic's documentation lists Sonnet 4.6 as an active legacy model and recommends moving to Sonnet 5. These templates are a snapshot of deployed defaults, so model lifecycle review belongs in the operating process.
Where Claude Haiku is enough
Every one of the 33 Claude Haiku 4.5 templates uses low reasoning. 31 have browsing turned off, and their turn budgets run from 20 to 35.
The work is usually constrained by a clear input and an explicit action boundary. The Google Calendar Scheduler reads a scheduling thread, proposes available slots, and leaves the invite as a draft. It uses low reasoning and up to 25 turns. The Outlook Inbox Manager triages messages and drafts replies without sending them, using low reasoning and up to 30 turns.
Those defaults line up with Anthropic's efficiency-first guidance, which positions Haiku for high-volume, straightforward work and recommends upgrading only when an evaluation exposes a capability gap.
The important qualifier is scope. Haiku is chosen where instructions, connected data, and approval boundaries keep the task narrow. The same family switches to Sonnet when the agent must judge a pull request, synthesize a long document, or coordinate a more consequential workflow. For project-based work rather than an autonomous workflow, see how to use Claude Projects.
How the GPT tiers divide the work
The OpenAI configurations show an even sharper tiering pattern.
Terra for structured operational work
GPT-5.6 Terra appears in 29 templates, with 26 at medium reasoning. Turn budgets stay between 25 and 40.
Many of these agents turn structured source material into a checked output: spreadsheet cleanup, invoice extraction, reporting, analytics, list enrichment, and finance operations. The Expense Processor uses medium reasoning and up to 35 turns to match receipts, transactions, policy, and account categories. The Data Analyst uses high reasoning and up to 40 turns because ambiguous definitions and unsafe joins can produce plausible but wrong answers.
OpenAI describes Terra as the GPT tier that balances intelligence and cost. The shipped configurations use it in that middle position rather than spending the highest reasoning budget on every data workflow.
Sol for open-ended research
GPT-5.6 Sol appears in 14 templates. 13 use high reasoning, 12 can browse, and every turn budget falls between 35 and 45.
That cluster includes market briefs, competitive research, candidate sourcing, meeting preparation, and research synthesis. The Research Assistant searches the web, opens sources, handles disagreement, and can run for up to 45 turns. The Meeting Prep Brief combines calendar context, internal history, and current web research with high reasoning and up to 35 turns.
The shared need is evidence gathering under ambiguity. These agents have to decide which sources matter and how to reconcile them before producing an answer. That explains the stronger model, high effort, browsing access, and longer loop more convincingly than a generic claim that GPT is better at research.
Luna for short recurring outputs
All 5 GPT-5.6 Luna templates use low reasoning and no more than 25 turns. The Morning Briefing is representative: it reads connected calendar and email context, produces a daily digest, and stops after at most 20 turns.
OpenAI positions Luna for cost-sensitive, high-volume workloads. In this library, it is the narrowest tier rather than the default for research or complex analysis.
Model choice is only part of the configuration
Across the library, 85 templates use medium reasoning, 41 use low reasoning, and 26 use high reasoning. Turn budgets range from 20 to 60, with 30 turns used by 58 templates.
Those settings change the practical comparison:
- A cheaper model with low effort and a short loop can be the right default for classification, routing, drafting, or scheduled summaries when inputs and stopping conditions are clear.
- Medium effort fits repeatable work that still needs interpretation, reconciliation, or several tool calls.
- High effort and a longer loop belong where the agent must gather evidence, inspect risky changes, or recover from intermediate failures.
Browsing and memory should also follow the workflow. The library enables browsing in 42 templates and leaves it off in 110. Memory is on in 122 and off in 30. Web research needs fresh sources. A transaction processor working from connected systems may gain little from open-web access. A recurring assistant can benefit from learned preferences, while a one-off research synthesis may be cleaner without them.
A practical selection rule
Start with the failure you need to prevent. These practical AI agent examples can help identify the workflow before you choose a model tier.
If the task has a stable schema, bounded source data, and a human approval point, begin with the efficiency tier: Haiku on a Claude foundation or Luna on an OpenAI foundation. Move up when your evaluation shows missed edge cases or weak judgment.
For structured work that still requires interpretation across tools, Sonnet and Terra are the workhorses in this dataset. Raise reasoning effort before assuming that a new model family will solve the problem.
For open-ended web research, the shipped OpenAI pattern is Sol with high reasoning, browsing, and a larger turn budget. For long-running code review or migrations, the Claude pattern is Sonnet with high reasoning and explicit stop conditions.
Then evaluate the full agent under real workflow conditions. Use the same source material, tools, approval rules, and pass criteria for each candidate. Anthropic's model-selection guidance also recommends testing with your actual prompts and data. Record failures by type so you can tell whether the model, instructions, tool access, or turn limit caused them.
So, ChatGPT or Claude for agents?
The production configuration does not support a universal winner. Claude is more common in this library, but the foundation split accounts for much of that lead. Inside each stack, the stronger finding is consistent tiering: Haiku and Luna for bounded low-effort work, Sonnet and Terra for the operational middle, and Sol for research-heavy workflows.
Choose the agent architecture first. Match model capability, reasoning effort, and turn budget to the cost of failure. Then keep evaluating, because a sensible default can become a legacy model.