Ask on Auto
Leave the composer on its default Auto setting and ask INARY as usual.
11 · Let AI pick the AI
AI Adaptive Orchestrator
Spend your time on the work, not on keeping up with AI.
When your business runs on AI, your time belongs to the actual work. But AI models and settings change every week, and if you don't keep up, you overpay. That's no way to run a work tool. AI Adaptive Orchestrator fixes this: leave the composer on Auto, and AI picks a cost-effective model and reasoning effort for each request. Everyone at the company uses AI every day, so cost matters. In our October 2026 test, Auto matched the accuracy of Claude Fable 5.1 at its API default for 92% less cost.
What is AI Adaptive Orchestrator?
AI Adaptive Orchestrator is how INARY lets AI choose the AI. For every request, it looks at the question's difficulty, length, attachments, and work surface, then selects one model and reasoning effort with cost-performance in mind. ChatGPT, Claude, Gemini, and Grok answer differently, and cost very differently, depending on the model and reasoning level, and their prices and performance keep changing. Your team shouldn't have to track that. For teams comparing ChatGPT for business with the alternatives, INARY gives you ChatGPT and the others in one account, and you can still choose the vendor, model, and effort yourself.
Same 440 questions, six AI models at their API defaults
We ran six AI models at each vendor's API default setting, plus INARY Auto, on the same 440 questions under the same conditions.
92% lower cost
vs. Claude Fable 5.1 (default: high)
Comparable accuracy (INARY 85.00%, Claude Fable 5.1 84.09%)
86% lower cost
vs. GPT-6 Astra (default: medium), which OpenAI calls its most capable model
INARY scored about 2 points lower (INARY 85.00%, GPT-6 Astra 87.05%)
27% lower cost
vs. GPT-6.1 Sol (default: medium), OpenAI's cost-efficient model
Accuracy gap within the margin of error (INARY 85.00%, GPT-6.1 Sol 85.91%)
| AI (API default) | Accuracy | Cost (440 questions) | INARY Auto cost |
|---|---|---|---|
| GPT-6 Astra · Medium | 87.05% | $3.77 | 86% less |
| Claude Opus 5.5 · Medium | 86.14% | $4.44 | 88% less |
| GPT-6.1 Sol · Medium | 85.91% | $0.72 | 27% less |
| INARY Auto | 85.00% | $0.53 | — |
| Claude Fable 5.1 · High | 84.09% | $6.82 | 92% less |
| Gemini 3.8 Flash · Medium | 82.05% | $1.59 | 67% less |
| Claude Sonnet 5.5 · High | 79.77% | $2.29 | 77% less |
Cost is the input and output tokens used to generate answers, multiplied by each vendor's published API prices (as of October 7, 2026). INARY Auto's cost does not include the step that picks the AI (about $0.03 for 440 questions). See Test conditions below.
WHAT CHANGES
How AI Adaptive Orchestrator changes the work
- 01
Ask without thinking about which AI, model, or reasoning effort to use. INARY picks for every request.
- 02
Keep AI costs in check even when everyone uses it. In our October 2026 test, Auto cost 27–92% less than six AI models at their vendors' API defaults.
- 03
Override Auto whenever you need to by choosing ChatGPT, Claude, Gemini, or Grok, then the exact model and one of its available reasoning levels.
Picked for cost-performance
INARY selects one available model and reasoning effort for the request, weighing cost against quality.
Keep human oversight
The answer shows the model and effort that ran, and administrators can review organizational usage and billing.
Test conditions
- Questions: 440 in total: SimpleQA (100), MMLU-Pro (100), GPQA Diamond (100), and MMMLU (140: 10 questions × 14 languages). Every setting received the same questions and the same prompts.
- No web search, tools, or INARY system prompt. Results as of October 7, 2026.
- "API default" is the reasoning effort each vendor's API uses when none is specified: medium for GPT-6.1 Sol and Claude Opus 5.5, high for Claude Sonnet 5.5 and Claude Fable 5.1, and medium for Gemini 3.8 Flash (per each vendor's official documentation). OpenAI does not document a default for GPT-6 Astra; we used medium, which the API returned when called without an effort level in INARY's test.
- Scoring: multiple-choice questions (MMLU-Pro, GPQA, MMMLU) were scored by matching the answer letter. SimpleQA was graded by AI, with the same grader, GPT-6 Luna (low), for every setting.
- Questions the specified AI did not answer were counted as incorrect, and their cost was excluded (1–5 questions each for Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5.5, and Gemini 3.8 Flash).
- Gemini 3.8 Flash is priced at its rate through December 31, 2026 ($0.75 input / $3.75 output per million tokens). From January 1, 2027, the rate rises to $1.50 / $7.50.
- All 440 questions test knowledge. Other kinds of work, such as writing or reading documents, were not part of this test.
- Grok 4.7 was excluded because it did not finish all 440 questions, and Gemini 3.1 Pro because it is a preview release.
- Because these are public benchmarks, the questions may appear in the AI models' training data.