Stack Overflow Developer Survey 2026
LLM vendors: Used in agents by gender
In the Non-binary group, 21% selected “Alibaba Qwen”; in the Woman group, 7% did. Of the 12 answers tested, “Alibaba Qwen” differs most by gender, a negligible association: χ²(4, N = 6802) = 19.31, Holm-adjusted p = .008, V = .05. Respondents could select more than one answer, so the percentages add to more than 100%. These are unweighted respondents, not population estimates.
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.The table
The table
Question asked: Which LLM model vendors for AI tools have you used for work in the past year? Which would you like to use next year? Select all that apply.
Respondents could select more than one answer. Shown by gender, as the percentage of each group's respondents who selected each answer, with the number of respondents beneath. A respondent is counted in every row they selected, so a column adds to more than 100%.
| LLM vendors: Used in agents | Man (A) | Non-binary (B) | Prefer not to say (C) | Prefer to self-describe: (D) | Woman (E) | All |
|---|---|---|---|---|---|---|
| Anthropic | 79.1%4,864 | 80.5%70 | 76.4%155 | 66.7%32 | 73.6%231 | 78.7%5,352 |
| OpenAI | 56.0%3,442 | 56.3%49 | 52.7%107 | 39.6%19 | 52.2%164 | 55.6%3,781 |
| Google AI | 29.3%1,805 | 26.4%23 | 35.0%71 | 29.2%14 | 25.2%79 | 29.3%1,992 |
| Microsoft | 14.8%910 | 10.3%9 | 16.3%33 | 6.3%3 | 18.8%59 | 14.9%1,014 |
| DeepSeek | 10.9%671 | 17.2%15 | 14.8%30 | 10.4%5 | 8.9%28 | 11.0%749 |
| Alibaba Qwen | 8.4%517 | 20.7%AE18 | 10.8%22 | 12.5%6 | 7.3%23 | 8.6%586 |
| Mistral AI | 4.9%302 | 3.4%3 | 4.4%9 | 8.3%4 | 3.8%12 | 4.9%330 |
| Moonshot AI | 4.1%255 | 6.9%6 | 5.4%11 | 8.3%4 | 4.5%14 | 4.3%290 |
| Other, please specify | 3.8%234 | 4.6%4 | 3.4%7 | 10.4%5 | 3.2%10 | 3.8%260 |
| xAI | 3.6%223 | 1.1%1 | 5.9%12 | 8.3%4 | 2.2%7 | 3.6%247 |
| Zhipu AI | 2.7%163 | 4.6%4 | 5.9%a12 | 6.3%3 | 3.5%11 | 2.8%193 |
| Meta AI | 2.5%152 | 6.9%a6 | 4.4%9 | 4.2%2 | 5.4%A17 | 2.7%186 |
| Databricks (DBRX) | 1.2%73 | 1.1%1 | 1.5%3 | 4.2%2 | 2.2%7 | 1.3%86 |
| Snowflake | 1.1%68 | 1.1%1 | 1.5%3 | 4.2%2 | 1.6%5 | 1.2%79 |
| Cohere | 0.9%56 | 1.1%1 | 1.5%3 | 4.2%2 | 2.2%7 | 1.0%69 |
| Voyage AI (MongoDB) | 0.6%39 | 1.1%1 | 1.0%2 | 4.2%A2 | 2.5%A8 | 0.8%52 |
| Cascade | 0.7%40 | 1.1%1 | 1.0%2 | 4.2%A2 | 1.9%a6 | 0.7%51 |
| Black Forest Labs | 0.6%35 | 1.1%1 | 1.0%2 | 8.3%AC4 | 1.9%A6 | 0.7%48 |
| 01.AI | 0.5%29 | 1.1%1 | 1.0%2 | 4.2%A2 | 1.9%A6 | 0.6%40 |
| Luel | 0.5%28 | 1.1%1 | 1.0%2 | 6.3%A3 | 1.9%A6 | 0.6%40 |
| Stability AI | 0.5%29 | 1.1%1 | 1.0%2 | 4.2%A2 | 1.9%A6 | 0.6%40 |
| Guide Labs | 0.4%26 | 1.1%1 | 1.0%2 | 4.2%A2 | 1.9%A6 | 0.5%37 |
| Reka | 0.4%22 | 1.1%1 | 1.0%2 | 6.3%A3 | 2.2%A7 | 0.5%35 |
| Respondents | 6,150 | 87 | 203 | 48 | 314 | 6,802 |
A letter beside a percentage names the columns it is significantly higher than (uppercase 95%, lowercase 90%; pairwise column tests, Bonferroni-corrected). 24,101 of 30,903 respondents have no answer to one of the two questions (not asked or not answered) and are left out.
Analysis
Results
Chi-square test of “Alibaba Qwen”
χ² = 19.31
df = 4, Holm-adjusted p = .008
Effect size
Cramér's V = 0.053
a negligible association
The strongest answer, APA-style: χ²(4, N = 6802) = 19.31, Holm-adjusted p = .008, V = .05. Each answer is tested on its own, as the respondents who selected it against those who did not, across the 5 groups (N = 6,802). The p-values are Holm-adjusted for the 12 answers tested; 2 of them differ by gender at the 0.05 level. 11 answers are not tested: too few respondents selected them for the test to be reliable (an expected count below 1, or more than 20% of the expected counts below 5).
| Answer | Selected | χ² (df = 4) | Adjusted p | Cramér's V |
|---|---|---|---|---|
| Anthropic | 78.7% | 10.46 | .300 | 0.039 |
| OpenAI | 55.6% | 7.47 | .578 | 0.033 |
| Google AI | 29.3% | 6.11 | .725 | 0.030 |
| Microsoft | 14.9% | 8.35 | .558 | 0.035 |
| DeepSeek | 11.0% | 7.87 | .578 | 0.034 |
| Alibaba Qwen | 8.6% | 19.31 | .008 | 0.053 |
| Mistral AI | 4.9% | 2.48 | .728 | 0.019 |
| Moonshot AI | 4.3% | 4.33 | .728 | 0.025 |
| Other, please specify | 3.8% | 6.25 | .725 | 0.030 |
| xAI | 3.6% | 9.34 | .424 | 0.037 |
| Zhipu AI | 2.8% | 11.25 | .239 | 0.041 |
| Meta AI | 2.7% | 18.31 | .012 | 0.052 |
| Databricks (DBRX) | 1.3% | Not tested: expected counts too small | ||
| Snowflake | 1.2% | Not tested: expected counts too small | ||
| Cohere | 1.0% | Not tested: expected counts too small | ||
| Voyage AI (MongoDB) | 0.8% | Not tested: expected counts too small | ||
| Cascade | 0.7% | Not tested: expected counts too small | ||
| Black Forest Labs | 0.7% | Not tested: expected counts too small | ||
| 01.AI | 0.6% | Not tested: expected counts too small | ||
| Luel | 0.6% | Not tested: expected counts too small | ||
| Stability AI | 0.6% | Not tested: expected counts too small | ||
| Guide Labs | 0.5% | Not tested: expected counts too small | ||
| Reka | 0.5% | Not tested: expected counts too small | ||
How to read it
No single chi-square test covers this table. A respondent appears in every row they selected, so the rows are not separate groups of people, and a test of the whole table would count the same person several times. Each answer is tested on its own instead: did the share who selected it differ between the groups? Testing 12 answers gives 12 chances of a false alarm, so the p-values are Holm-adjusted to keep the chance of any false alarm in this table at 0.05.
For “Alibaba Qwen” the test rejects independence at the 0.05 level (Holm-adjusted p = .008): differences this large between the groups are unlikely if selecting it did not depend on gender. Cramér's V of 0.053 is in the negligible range, which is trivially weak in practical terms, even where a sample this large can detect it.
With 6,802 respondents, even small differences reach statistical significance. Read the percentages and the effect size before the p-value.
Limits
- The survey has no weights. These tables are unweighted counts of the people who answered, so they describe the respondents, not developers in general.
- The test treats respondents as a simple random sample. They are people who chose to answer a survey Stack Overflow promoted, not a random sample of developers, so a result holds for these respondents and may not hold beyond them.
- The adjustment covers the answers of this question by this group only, not the other tables in the library. The headline answer is the strongest of those tested, so its effect size is larger than a typical answer's.
- A difference between groups is not evidence that group membership causes the answer.
Next steps
Work with this table
Open the test of “Alibaba Qwen” in the calculator
The link fills in the respondents who selected this answer and those who did not, by group. The whole table does not open there: its rows overlap, and the calculator's test needs rows that are separate people.
Open this question in the full file
The workspace opens all 30,903 respondents in your browser with this question selected. The survey has no weights, so the table there is unweighted too. Change the group, add a filter or export from there.
Open in the workspace →Run this on your own data — free, no signup
Upload a CSV, Excel, SPSS, Stata, SAS, R, Parquet or JSON file. Everything runs in your browser; your file never leaves your device.
Open the workspace →The same question by other groups
Other questions that differ by gender
- AI tools or agents currently used at work by gender
- First actions when trying to solve a problem by gender
- How AI has changed their Stack Overflow use by gender
- Tasks delegated to AI agents by gender
- What prevents taking part in technical communities by gender
- How they learned to code in the past year by gender
Source
Stack Overflow. 2026 Developer Survey, https://survey.stackoverflow.co/2026, licensed under the ODbL 1.0. Tabulated by crosstabs.com. Variables LLMUsedInAgents (LLM vendors: Used in agents) and Gender. Stack Overflow licenses the survey data under the Open Database License (ODbL) 1.0 and its contents under the Database Contents License (DbCL) 1.0: credit Stack Overflow, keep the licence notice, and share any database made from the data under the ODbL. These tables and their downloads are offered on those terms.