Creative Signals · PlatformData: Jun 9–Aug 31, 2026Published Sep 2026
Agent articles describe what AI agents could do. This one measures what they already ask for. In August 2026, software—MCP clients, the public API, and the platform’s agent—requested 4.35% of the 100.9 million attributed image, video, and audio generations on Magnific and Freepik. That is 1 in 23 generations, 3.80% of the calls, 7.52% of the accounts that generate, and 5.38% of credits at list price.
The account share is the fragile one: 5.49% on the widest denominator, 9.04% among accounts with 5 or more generations. In June, a partial month from the 9th, the share was 1.41%, and it is not spread evenly: in August software asked for 18.82% of audio generations against 4.22% of image. Nor is it a separate audience: 84.67% of MCP accounts also pressed generate in the app that month.
Software share of AI generations
4.35%
1 in 23 attributed image, video, and audio generations, August 2026. MCP 3.21%, the API 0.80%, the agent 0.35%. Without the API's 10 largest accounts, 0.51% and a 4.08% headline. June, a partial month from the 9th, was 1.41% and July 2.86%
of AI generations come from software
Audio, the automated modality
18.82%
Nearly 1 in 5 in August, and 18.76% of audio calls, against 4.22% of image and 4.01% of video. MCP carries 17.82 of the 18.82 points
of audio generations come from software
Auto on MCP and the API
0%
“Auto” is 4.87% of the app's image in August and exactly 0 on MCP and on the API. The agent is the only software origin using it, at 0.73%
of MCP and API image generations let the platform pick the model — “Auto” is the setting that hands the choice of model to the platform. Origin is not autonomy: an MCP client has a person typing, and the agent runs what someone asked for.
Requests per account-day
9 vs 11
In August the median API account is lighter than a person in the app, but its tail is 7.5 times thicker: 0.994% of its account-days pass 500 requests against 0.133%
requests per active account-day, the API against the app
Section 01
AI agents and other software tripled their share of AI generations since June
4 origins ask for the work, 3 of them software. In August 2026 they requested 4.35% of attributed image, video, and audio generations, 3.80% of the calls, 5.38% of the credits, and touched 7.52% of the accounts that generate, a share that falls to 5.49% on the widest denominator, rises to 9.04% with 5 or more generations, and 12.38% with 20 or more.
MCP carries most of it, 3.21 of the 4.35 points, or 2.66% on the call base, ahead of the public API at 0.80%, all 3,523 accounts that call it, and the agent at 0.35%. The API figure is the concentrated one: its 10 largest accounts make 35.52% of its generations, and without them its share is 0.51% and the headline 4.08%. Swap the denominator and it barely moves: 4.33% with the unattributed generations inside, 4.56% with the image tools added, 4.47% with both, a band 0.23 points wide, 0.12 on either side of its midpoint. The direction is the finding. Monthly, 1.41% in June, a partial month from the 9th, then 2.86% and 4.35%: that rise is a floor. Weekly, 1.36% in W25 to 4.98% in W35: a share multiple of 3.66 against 2.60 in generations, with the attributed base down 28.9%.
Four lines over the 11 complete ISO weeks from W25 to W35 of 2026: the software share of the platform's attributed image, video and audio generations rising from 1.36% to 4.98%, with MCP, the public API and the platform's agent as thin lines under it
| Modality (August 2026) | Attributed generations | A person in the app | MCP | API | Agent | Software, generations | Software, calls | Software, credits |
|---|---|---|---|---|---|---|---|---|
| Image | 90,400,970 | 95.78% | 3.04% | 0.81% | 0.37% | 4.22% | 3.68% | 4.48% |
| Video | 9,457,083 | 95.99% | 3.20% | 0.63% | 0.18% | 4.01% | 3.23% | 5.82% |
| Audio | 1,084,855 | 81.18% | 17.82% | 0.89% | 0.12% | 18.82% | 18.76% | 14.64% |
| All three | 100,942,908 | 95.65% | 3.21% | 0.80% | 0.35% | 4.35% | 3.80% | 5.38% |
Section 02
Software automates sound and cutouts, and barely touches the canvas
What software asks for is not a smaller version of what a person asks for. Among the tools with volume in August, the most automated are the deterministic jobs.
Sound effects run at 23.27% software, music at 20.86%, voice at 19.39%, and remove background at 15.07%, far above the 4.35% average. The interactive surfaces are the counterweight. Talk to image is 0.00% software in August and across the window, a zero that rests on one column of the records and would hide a surface mislabeling. Spaces is 0.75% across its four modalities, and the agent is the partial exception at 14.02% of its volume. The API is the clearest editor: 14.89% of its 945,378 generations are image tools against 3.53% for the app. When software generates an image, 94.10% of it is the plain generator in aggregate: 95.04% on MCP, 94.15% on the API, 86.22% on the agent, whose 11.80% in Spaces explains the gap. A person splits the same work between the generator, the canvas, and Talk to image: 67.36%, 20.01%, and 10.75%.
Eighteen horizontal bars, the software share of each tool's August 2026 attributed generations, from the sound effects generator at 23.27% down to Talk to image at 0.00%, with a dashed line at the 4.35% platform average
Section 03
The typical API account is lighter than a person in the app
A call does not ask for a bigger batch. Generations per call sit near 1 in all four origins: 1.263 on MCP, 1.104 on the agent, 1.039 in the app, 1.016 on the API.
MCP’s 1.263 is one call that really does charge more than one generation, not the same generation charged twice. Its 404,224 multi-generation calls carry distinct transactions and 93.58% resolve inside a second. The volume comes from a few accounts, not a different kind of user. The median API account makes 9 requests per active day against 11 for the app, nobody excluded. Its mean of 48.03 is 1.815 times the app’s, but that multiple does not survive its own concentration: 1.197 times without the API’s 10 largest accounts, 0.768 without the 100 largest. 68.72% of API generations sit in bursts of 20 or more, 55.64% without its 10 largest accounts and 44.63% without the 100 largest, against 41.85% for the app. What holds is the tail: 0.994% of API account-days pass 500 requests against 0.133% in the app, 7.5 times thicker. The agent is the opposite shape, with 43.96% of its bursts a single shot.
Grouped horizontal bars for six populations, each with the median, the 90th and the 99th percentile of requests per active account-day in August 2026: a person in the app at 11, the API at 9 with all its accounts, the API without its 10 and its 100 largest accounts, MCP and the platform's agent, each labelled with the share of its account-days above 500 requests
Window
June 9, 2026 to August 31, 2026 (UTC), the period from which MCP and agent traffic can be attributed. Clean months: July and August. June is partial and a ramp-up, and the report says so next to every June figure. Complete ISO weeks W25–W35, each with 7 days and 2 weekend days. MCP client families are published from July, because client coverage is 32.94% in June, 97.47% in July, 97.85% in August. There is no interface-versus-API series before June: the API, its contracted flag, and its wallet pattern all appear in April 2026. Monthly and weekly volume is not published as a trend: generations follow platform traffic, and this report measures shares, ratios, medians, and indexes.
Population
Charged generations recorded in the platform’s credit records (Magnific and Freepik combined). One record is one charged generation. Families come from the record that charged the generation, never from the surface it was labeled with: image, video, and audio make the publishable denominator, image tools (upscaler, remove background, expand, reimagine, icons, SVG, 3D) are published apart with their own caveats, and text steps inside agents and workflows, voice cloning setups, and an unclassified remainder are left out and declared. Excluding text costs the agent more than anyone, 10.88% of its charged generations. August 2026: 100,942,908 attributed generations · 96,589,909 calls · 758,826 accounts · 19,790,014,540 credits, the three modalities partitioning the month exactly. Window total: 318,880,806 attributed generations, 2.87% of them software. Origins: a person in the app (web, iOS, Android, desktop), the platform’s agent, the public API, and MCP clients. Generations with no origin are counted, declared, and kept out of every share. Calls by origin add up to the month’s total exactly, and no request crosses origins. The audio side of the same wave is measured from the voice angle in The typical AI voice clip is 12 seconds long, whose population and window are different, so its surface shares are not comparable with these.
Definitions
A generation is one charged record. A call is one request, which can charge more than one generation, so every share is published on both bases. Software means MCP clients, the public API, and the platform’s agent together. Origin is not autonomy: an MCP client has a person typing and the agent runs what someone asked for, so this report says who or what sent the request, never who did the thinking. An account-day is one account on one UTC day. A burst is consecutive generations by the same account on the same tool less than 10min apart, built without partitioning by origin and assigned to the origin of its first generation.
Credits are list price, never currency, and the credit share is published in two valuations, 5.38% as charged and 6.46% with unlimited generations valued at 0. Model names are the public ones, read from the charge each generation carries: Nano Banana 2, Nano Banana, Seedream 5 Pro, Seedream 4.5, GPT Image 2, Auto, Flux 2 Max, Flux 2 Pro, Flux Kontext, Recraft V4, Z-Image, Ideogram, Mystic, Freepik Fast, Flux Dev, Kling 2.5, Kling 2.6, Kling 3.0, Kling Omni 3, Seedance 2.0, Seedance 2.0 Mini, Seedance 2.5, Hailuo 2.3, and Veo. Nano Banana Pro has no generations in August. All quantiles are exact by value and cumulative count, not approximations. Full definitions in the Report Glossary.
Limitations
- 1. June is partial, from the 9th, so the 3.09 times rise from June to August is a floor: the 8 days the window drops hold 38,861,188 attributed generations with only 0.545% software, and the multiple over the full calendar month would be 3.71 times, which is not comparable because MCP and the agent are unattributed before the 9th. The weekly rise is 3.66 times in share and 2.60 times in generations, and both are published together. No daily maximum is published: weekend share runs 5.76% against 4.06% on weekdays because attributed volume falls to 44.5%, which is composition, not adoption.
- 2. The account share is the fragile figure. 7.52% of accounts becomes 5.49% once the image tools' unattributed residue enters the denominator, because it adds 439,617 accounts with almost no app activity. Conditioned on real use it moves the other way: 9.04% among accounts with 5 or more generations and 12.38% among those with 20 or more, so 7.52% is a floor as much as a ceiling.
- 3. API volume is concentrated: 10 accounts make 35.52% and 100 make 62.46% of its generations. Without its 10 largest accounts the API share falls from 0.80% to 0.51% and the headline from 4.35% to 4.08%. Part of August's API volume is concentrated in a single account, which is why the daily series is declared and not published, and no account, client, or organization is named or described. “The API” is also two populations: every API figure in this report covers all 3,523 accounts that call it, and the subset on a signed contract is 187 accounts and 29.57% of API generations.
- 4. Internal traffic is inside the published base. The filter the glossary inherited is inert, because no generation carries the label it looks for. Real internal traffic, identified by the product it is charged to, is 0.12% of the publishable base and 19.69% software, 4.5 times the platform rate, so it biases the headline upward. The bound is measured and small: excluding it moves 4.354150% to 4.336658%, 0.017 points. The frozen base is published with the bound declared.
- 5. Image tools are never added to creation without their caveats: 29.53% of that modality has no origin, 99.78% of that residue is background removal at about 2.4 credits per generation against 185 in the attributed ones, and 2.54% of the modality is downloads rather than generations. Its software share is also fragile, 9.95% falling to 7.01% with its own residue inside, while audio only moves from 18.82% to 18.25%.
- 6. 2 zeros rest on one column. Talk to image and the video modifier are 0.00% software in August and across the window, but no charge record names a conversation surface, so a mislabeled surface would be invisible. Bursts are built without partitioning by origin, and 8.05% of the agent's bursts mix origins, so its exact burst figures depend on the cut while the direction holds. Unattributed generations sit outside the burst population, so a burst crossing them splits in two.
- 7. Duration coverage is uneven, so seconds are reported per charged generation. The API's audio has a duration on 20.56% of its August generations, and unattributed audio on 15.21%, which makes the API's 20.6 audio hours a minimum. 97.5% of software audio hours are MCP with coverage above 99.92%, so the audio headline does not move. Exact-match model rules leave sibling versions of Kling Omni 3, Kling 2.6, and Hailuo 2.3 in the remainder, so those version shares are lower bounds, while family totals are complete, and Veo is present in all four origins.
Reproducibility
A second pass with its own queries recomputed the four bases, four hostile denominators, the monthly and weekly series, the modality split, the account overlap and concentration, the 37 tool and modality pairs above 5,000 generations, requests per account-day and burst sizes with three concentration scenarios, credits in three valuations, the MCP client families, and the full-base model tables before publication. It also ran a double-charge test that came back clean across every origin. The first round failed the Data Room with 4 blocking findings and 29 corrections. The re-audit passed at 93/100 with 10 precision corrections, all of them marked in the source.
Cite & data
Cite this report
Citation
Magnific (2026). 1 in 23 AI generations is now requested by software. Creative Signals, September 2026. magnific.com/blog/ai-agents-vs-people
Short form for press
“In August 2026, software—MCP clients, the public API, and the platform's own agent—requested 4.35% of the 100.9 million attributed image, video, and audio generations on Magnific and Freepik: 1 in 23 generations, 3.80% of the calls, 7.52% of the accounts that generate, 9.04% of those with 5 or more generations in the month, and 5.38% of the credits at list price, 6.46% with unlimited plans valued at 0. The share was 1.41% in June, a partial month measured from the 9th, and 2.86% in July, and it ran from 1.36% to 4.98% across 11 weeks. Audio is the automated modality at 18.82%, against 4.22% of image and 4.01% of video. Chat assistants requested 74.47% of MCP generations, coding tools 16.26%, and 84.67% of MCP accounts also generated in the app the same month. Software names its model: the setting that lets the platform choose is 4.87% of the app's image generations and exactly 0 on MCP and on the API. And the typical API account is lighter than the typical person in the app, 9 requests per active day against 11, with a tail 7.5 times thicker.”
