Case study ·
Why Michelin switched off GitHub Copilot's Auto mode
Original title, in French: “Pourquoi Michelin a coupé le mode Auto de GitHub Copilot”
- Publication
- Silicon.fr (France)
- Author
- Clément Bohic
- Published on
- 21 September 2026 · read in the page structured data · updated on 22 September 2026
Summary
Michelin details an experiment conducted on around 700 requests to evaluate GitHub Copilot's Auto mode, which automatically selects an AI model using a classifier based on four criteria. The company found that this classifier analyses words rather than the actual context of tasks, leading to a model choice that is sometimes random and to large cost differences (from 0.23 to 3.24 credits for the same request). Michelin also shows that the 10% discount promised by GitHub does not always offset the additional cost of Auto mode, notably in chat sessions. The company considers that a cheaper model (GPT Luna) delivers equivalent results for single tasks, and describes Auto mode as a "safety net" rather than a source of savings. An experimental orchestrator, HydraFusion, proves significantly more expensive.
Summary based on the full text of the article.
Read the article on Silicon.fr →Figures cited
-
0.23 credit
Cost of a variable renaming request with MAI-Flash
Michelin test on GitHub Copilot
« Avec le premier, il en a coûté 0,23 crédit. »
-
3.24 credit
Cost of a variable renaming request with GPT-5.6 Terra
Michelin test on GitHub Copilot
« Avec le second, 3,24. »
-
11.9 %
Average Auto mode discount with the same model (Claude Sonnet 5)
6 requests tested by Michelin
« sur 6 requêtes, 11,9 % moins cher en moyenne »
-
13.7 credit
Average cost of a chat session opened with a difficult question
Michelin test, 7-turn sessions
« pour un coût moyen de 13,7 crédits »
-
0.78 credit
Average cost of a chat session opened with a simple question
Michelin test, 7-turn sessions
« pour un coût moyen de 0,78 crédit »
-
28 credit
Additional cost of HydraFusion compared with Auto mode for a cascading task
Michelin test, HydraFusion vs Auto mode
« Il en a coûté 16 et 28 crédits, là où le mode Auto en aurait consommé 3. »
Classification
- Country
- France
- Company size
- Large companies
- Industries
- Industry and manufacturing · Software vendors, IT services companies, telecoms
- Companies
- GitHub (supplier), Michelin (deploys AI)
- Keywords
- github copilot, generative artificial intelligence, AI costs, software development, model routing, AI optimisation
Entry written by Marsanne Studio from the cited article, with the help of an AI model (claude-sonnet-5). The rights to the article and its illustrations belong to Silicon.fr and its author. Entry translated from French by Claude Sonnet 5.5.