Reports

Web Games in 2026: Is Game Development Moving to the Browser?

Web games have gone from the thing you play while waiting for a download to a serious distribution channel with its own platforms, engines and AI-built hits. We went through the usage data, browser support tables, platform announcements and game jam results to answer a narrower question than the headline suggests: which games are moving to the web, and which are not.

By the CreateGame.ai editorial teamOctober 9, 202622 min read

What we mean by "moving to the web"

Three different claims hide under that phrase, and the evidence for each is different. The first is that games are increasingly played in a browser: that is a question about audience, and the portals and social platforms speak to it. The second is that game developers are increasingly building with web technology, whether or not the shipped product runs in a browser (Electron wrappers, web-based tools, JavaScript engines). That is a question about tooling, and npm data and engine releases speak to it. The third is that the web becomes where games are made first and shipped by default, including bigger ones, which is the strong claim and the one we are most skeptical about.

We will keep these separate. A reader building a one-person project and a reader running a 40-person studio need different answers.

The technical floor has risen

WebGPU is now mostly just there

For years "WebGPU games" meant a flag in Chrome Canary. That is over. Chrome and Edge have shipped WebGPU since version 113 (2023). Safari 26 shipped it in September 2025 across macOS, iOS, iPadOS and visionOS, according to WebKit's feature post as quoted in a WebKit bug report (we could not open the WebKit post directly). Firefox has it enabled by default on Windows and, per caniuse notes, on macOS 26 Tahoe with Apple Silicon.

The caniuse data file shows the combined effect. We pulled the WebGPU feature file from the caniuse repository at six points in its history and read the global usage figures:

Two readings of this table. The first is optimistic: the share has more than tripled in under three years, and iOS Safari, which was the missing piece for mobile, now reports full support in the caniuse file. The second is cautious: caniuse's "partial" bucket is mostly Firefox and desktop Safari, where WebGPU is on by default only for specific platforms (Windows for Firefox; macOS 26 or later for Safari, and Apple Silicon for Firefox on Mac, as the notes put it). Chrome on Linux depends on hardware and drivers. And the plain truth about any global usage percentage is that it describes traffic, not your audience. A game aimed at a school laptop fleet or at older Android phones needs its own numbers.

For comparison, WebGL 2 sits at 96.44% and WebAssembly at 96.63% in the same file, which is why almost every production web game in 2026 still ships a WebGL 2 path. Godot's current web export, for example, requires WebGL 2.0 and not WebGPU, per its documentation.

Developers also report that WebGPU is not automatically faster. A Utsubo write-up of Three.js in 2026 says as much and advises profiling before migrating. The main benefit is compute shaders, which WebGL never had: GPU particle simulation, culling, skinning and similar work, which is the work that makes bigger scenes viable.

The engines have moved

Three.js. The library is the center of gravity. Release r171 (November 2024) added the three/webgpu entry point so a project could adopt the new renderer with one import, and kept a WebGL 2 fallback. Since then releases have added a node-based shading language (TSL), storage textures, a docs file for language models (llms.txt, r183 in February 2026), WebXR on the WebGPU renderer and clustered lighting (r185, June 2026), and, in r186 (September 8, 2026), a sun-light addon with cascaded shadows and a Gaussian splatting addon. The repository has about 116,000 stars on GitHub as of today (GitHub API, 2026-10-09).

Babylon.js. Version 9.0 landed in March 2026. Microsoft's announcement (Windows Developer Blog) described clustered lighting available on both WebGPU and WebGL 2, textured area lights, volumetric lighting and a node-based particle editor, per coverage of the release. The GitHub repo has about 26,000 stars.

PlayCanvas. The engine dropped WebGL 1 at 2.0 to make room for WebGPU and has been working on Gaussian splat rendering through the 2.x line; the 2.15 release in January 2026 is an example. Secondary sources disagree on how mature its WebGPU path is (one calls it WebGL 2 first, another calls it the most production-ready), so we would check the release notes before choosing it for that reason alone.

Unity. Unity 6.6, released September 1, 2026, took WebGPU out of experimental: "as of Unity 6000.6, the WebGPU graphics API is no longer experimental," in the words of Unity's web team as relayed by Cinevva (a secondary source that cites Unity's forums; we did not read Unity's post directly). It is off by default, WebGL 2 remains the default API, and device support is described as inconsistent. The same release adds WebAssembly 64 builds that can address up to 16 GB of memory instead of 4 GB, and per-scene progressive asset loading, which addresses two of the classic complaints about Unity web builds, memory ceilings and long first loads.

Godot. Since 4.3 the preferred and default web export is single-threaded, which avoids the cross-origin headers that block ads and third-party scripts, and 4.5 enabled WebAssembly SIMD by default, per the official docs and a summary. The same docs say C# projects in Godot 4 could not be exported to the web at the time of that page. The Godot web export is WebGL 2 only. If your game is 2D, this is a comfortable path; for GPU-heavy 3D, Unity's WebGPU route is further ahead on features, but also heavier.

The pattern: every major engine now treats the browser as a first-class target, and the three that matter for 3D (Three.js, Babylon.js, Unity) all have a production-level WebGPU path with WebGL 2 as the safety net.

What the download numbers do and do not say

Three.js is the headline number in this story, so we looked at it directly. The npm registry's public API gives download counts per package. For the week of September 24 to 30, the figures are:

Read alone, this looks like an explosion in web game development. But we ran the same query for packages that have nothing to do with games, and the controls moved too: React went from about 45.4 million to 214.4 million (4.7x), D3 from 4.9 million to 25.1 million (5.1x), Express from 48.5 million to 163.4 million (3.4x), lodash from 75.6 million to 199.2 million (2.6x) and Chart.js from 6.2 million to 15.3 million (2.5x). The registry as a whole is inflating, and the most obvious candidates are automated installs: CI runs, containers and coding agents that fetch dependencies on every task. npm does not separate those from humans, and we cannot prove what drove the increase.

So we would say: Three.js, Phaser and PlayCanvas grew faster than the registry's general inflation (about 7.4 to 8.5 times against a control range of 2.5 to 5), which is consistent with real extra adoption, but the effect is smaller than the raw numbers suggest. Pixi.js grew about as fast as React. Downloads count installs, not shipped games.

The time series for Three.js shows the same picture in more detail. We used weekly counts for Tuesday-to-Monday windows from the npm API, at roughly quarterly intervals:

Two warnings. The npm daily series has days that return zero, almost certainly missing data, and the week of September 15 to 21, 2026 contains one such day; the API sum for that week is 15.1 million, whereas the Utsubo article, written earlier, reported 12.18 million for the same week. So the real figure for that week is probably a bit higher than 15.1 million, and the sources disagree for reasons we cannot resolve. And the shape of the curve, a flat 2025 and a steep 2026, is the signature of a change in how packages get fetched at least as much as a change in how many people write 3D games. If a reader wants one sentence from this section: Three.js is clearly the dominant JavaScript 3D library and is growing, but the claim that "downloads grew eight times so web gaming grew eight times" does not hold up.

Where the players are

Portals

The web's mass-market games audience lives on a few sites. According to a CB Insights summary of Poki's March 2026 announcement, Poki welcomed 625 million players across desktop and mobile in 2025 and now counts 100 million monthly active players, built by a 65-person team that has taken no outside funding. We could not find Poki's own press release, and one outlet's June 2026 article gave a much lower figure of nearly 50 million, so we treat 100 million as the high end. CrazyGames' own LinkedIn page says it reached 50 million monthly users in 2025 (a company post, again not audited). Compare the older 25 million monthly players in its 2023 tenth-anniversary press materials.

Revenue is where the public record gets thin. Neither portal publishes it. A business-data site lists CrazyGames at $13.6 million for 2026, which is a database estimate and not a company statement, and we would not cite it for anything that matters. What the portals do publish is how they pay developers. According to Poki's SDK page, traffic you bring yourself earns you 100% of the revenue, while traffic that arrives through Poki is split 50/50; Poki also asks for web exclusivity in return for its investment, though developers can publish elsewhere on Steam and mobile. Third-party guides put CrazyGames' ad share at around 60%, which we have not been able to confirm from CrazyGames itself.

The most concrete earnings data we found is one developer's. In a Business Insider profile surfaced via the CB Insights page, the maker of a dress-up game on Poki reports 276 million play sessions from 63.1 million players over its lifetime, and 10 to 15 million sessions a month now. Those are the developer's numbers, not Poki's, and we have no revenue figure to match them. They do show that a single small game can reach audiences larger than most Steam hits.

Our read: the portals are real and large, they have a business model (advertising, with revenue shared with developers), and they are a mature place to ship 2D and light 3D games. They are also gatekeepers. Poki's own terms show the trade: you get distribution and ad infrastructure, and you give up exclusivity and a large cut of portal-sourced revenue.

Social platforms

Beyond portals, three platforms host browser games inside a larger app, and the numbers for all three are murky.

Discord Activities. Discord opened Activities to all developers in 2024. A Discord GDC 2026 announcement, as summarized in a Cinevva guide that we could not trace to Discord itself, says apps and Activities are used by more than 25% of monthly active users; a September 2024 report had already quoted that figure. If you apply 25% to the 200 million monthly users cited in a GDC 2026 summary, you get roughly 50 million (our arithmetic; the base is uncertain because sources between 150 and 200 million circulate, and "apps and Activities" is broader than games). Individual successes are self-reported: one title, Death by AI, is described as peaking around 700,000 daily users.

YouTube Playables. Google has expanded Playables to more than 50 markets including the EU and Nordics, from seven at launch in 2024, per a YouTube partnerships manager quoted by PocketGamer.biz. Google provides SDK samples for plain JavaScript, Flutter web, Godot and Unity. The only 2026 catalog figure we found is from Mediacube, a publisher: its catalog nearly doubled from 788 to 1,477 titles between July 1 and September 4, 2026. That is a vendor blog and unaudited. YouTube has not, as far as we found, published how many people play.

Telegram Mini Apps. The category had its boom and its bust. Hamster Kombat's monthly users fell sharply in late 2024 (a drop from over 100 million to fewer than 44 million in Telegram's own data, per reports quoted by Coin360 and others), and the surrounding clicker genre has not recovered. A vendor blog claims 47 million monthly users for gaming mini apps in January 2026; we have no primary source and show it in the chart below as low confidence.

All three are in-app distribution layers for web code. That is relevant because it means a single HTML5 game, with some SDK glue, can reach audiences on portals, in chat apps and on a video site. The graph of reach is better for web than for any native store. The money is a different story: ad revenue shares, in-app purchase rules and platform policy swings are all set by the host, and Telegram's example shows how quickly a category can disappear.

AI makes the browser the default target

Fly.pieter.com and what it proved

In early 2025 Pieter Levels (levelsio) posted a flight simulator he had built with Cursor, in Three.js, with server code generated by Grok 3. It is still online, described as a free multiplayer flight sim made entirely with AI. 404 Media reported on March 5, 2025 that Levels claimed it made more than $50,000 a month; Xataka put the figure at about $52,000, most of it from in-game ads (22 ads, mainly from AI companies), while an early report said the ads earned $5,000 a month in its first weeks and the game had made $38,000 after ten days. Peak concurrency was given as 17,000 in one early article. These are all the creator's claims or early press, and they conflict, so the right word is "reportedly."

What does it prove? Not that anyone can make $50,000 a month from a vibe-coded game. 404 Media's own line is that the future of AI-assisted game development "will not be that simple," and the reach came largely from Levels' own audience (Xataka notes he had 623,000 followers on X, and the game got a boost when Elon Musk shared it). What it proves is narrower and more interesting. A person whose day job is not game development got a multiplayer 3D game to a public URL, running in everyone's browser, with no store review, no install and no engine license, in a time frame that used to be reserved for jam entries. The reason it was a web game and not an iOS app is exactly the reason it spread: a link is the lowest-friction way to share anything.

Vibe Jam

Levels followed it with a jam. The 2025 Vibe Coding Game Jam required at least 80% AI-written code, recommended Three.js and drew more than 1,000 entries, with The Great Taxi Assignment in first place and mrdoob and Andrej Karpathy among the judges, according to the Utsubo article (who did not link a primary source for the results, so we flag that).

The 2026 edition, officially the Cursor Vibe Jam 2026, ran from April 1 to May 1 with a 90% AI-written code rule, free browser-playable games, a bonus for multiplayer, and a listed prize structure of $25,000 gold, $10,000 silver and $5,000 bronze (a Summer Engine page lists $20,000 for gold, a difference we put down to an older snapshot). A third-party guide (Summer Engine) reports 945 submitted games and more than 242,000 plays. We could not retrieve the organizers' press page, and we could not find an official winners announcement in what we read, so those numbers should be treated as unconfirmed by a primary source.

The more useful analysis came from Colyseus, a multiplayer framework vendor. Its post on the jam counted 48 games out of 945 using its framework, which it says is about 5% of all entries and about 15% of online multiplayer ones (implying roughly 320 multiplayer entries; the post does not state the total, so that is our arithmetic). It describes the bar as "much higher than 2025," with progression, persistence and complete game loops, and names Hollowlands, Mafia City and Spellwright as top entries by social likes. That is a vendor looking at its own users, so the 15% share is as much marketing as measurement. Still, it makes a useful point: roughly a third of the entries had real-time multiplayer at all, in a jam where code is mostly written by a model.

Note what is unusual about the numbers. The 2026 jam had slightly fewer entries than 2025 (945 versus more than 1,000) while the prize pool was reportedly larger. Entries went down while quality went up, if you believe Colyseus. That is a sign the novelty is wearing off and craft is returning, which is what you would expect from a maturing medium.

Why web and not native

We think there are four reasons AI coding agents push games toward the browser. None of the following is proven by data in this article; we list them because they match what the jam and fly.pieter.com show.

First, the loop is short. An agent edits a file, you refresh the page, you see the result. No build farm, no signing, no project settings. Second, the output is a single, inspectable artifact: HTML, JavaScript and assets, which a model can read, diff and rewrite in one context window. Third, web code is abundant in training data. Three.js has been the default way to show 3D in a browser for more than a decade, with a huge amount of public example code, and the library even ships an llms.txt file now. We think this matters a great deal, but we have not seen a controlled comparison of model performance on web code versus Unity C# or Godot GDScript, and we say so. For how the models compare on game code in general, see best AI for coding games. Fourth, sharing is a URL.

LMArena's coding arena gives the closest thing to a measurement of how models do on this kind of work. It evaluates web app generation with anonymous head-to-head votes on single prompts. As of an October 5, 2026 summary in the French outlet Blog du Modérateur, the order was:

The article cautions that the leader has just over 2,000 votes and that the Qwen and Gemini entries are preliminary, and it notes that the full-stack ranking has a different leader (GPT-6 Astra Max, with fewer than 900 votes). The arena also measures a single prompt, not whether code is maintainable in an existing project. We read the table as a shortlist: the top ten is a mix of Anthropic, OpenAI, Google and Alibaba models, with spreads of about 100 points between rank 1 and rank 10. The same arena was called WebDev Arena until it was renamed Code Arena, and the live board is at arena.ai/leaderboard/code. For a fuller model-by-model comparison, see our review of AI models for game development.

So are games moving to the web?

Here is where we land, as opinion informed by the above.

Yes, for these kinds of games.

• Casual and hyper-casual games, which portals and Playables are built to monetize with ads.
• Io-style and light multiplayer games, where instant join matters more than fidelity. The Vibe Jam entries and fly.pieter.com are of this kind.
• Social games inside Discord or chat apps, where the host app supplies identity and a friend graph.
• Prototypes, jam entries and AI-generated first drafts of any genre. This is the category AI changes most.
• Interactive marketing and educational games, where a link beats an install.
• Mid-scope 3D games in the vein of what Unity 6.6's demo scenes show: a dense town, a spaceship interior with GPU particles. That is a claim about what is now possible, not about what is selling.

No, or not yet, for these.

• Large premium games. Asset sizes, memory ceilings (even Unity's new 16 GB WebAssembly 64 builds are not available in every browser) and the lack of a storefront with wishlists and reviews push these to Steam and consoles. We did not pull Steam revenue numbers for this piece, and we are not claiming a figure; the point is that nothing we reviewed suggests premium buyers are shifting to browsers.
• Games that depend on platform integration: controllers on consoles, achievements, cloud saves, modding workshops.
• Anything that needs a direct customer relationship and a price. The web's business models are ads and in-app purchases through someone else's wrapper. The premium, pay-once model that Steam made normal has no equivalent here.

The real shift is smaller and more interesting. For a growing class of developers, the browser is where a game first exists. Prototype it in Three.js or Phaser, share a link to testers, then decide whether it graduates to Steam, to an app store through a wrapper, or stays on the web. AI coding tools reinforce that habit because they are strongest at web-sized projects with a short loop. We think that is the lasting effect, and that it will show up in the shape of indie game development more than in a headline about AAA.

There is also a risk we want to name: discoverability. Poki says it has more than 600 studios per a CB Insights profile, with top performers reportedly earning up to 1 million euros a year, and one third-party guide estimates $500 to $3,000 a month for a typical first web game (a guide's estimate, not data). When AI makes 945 games in a month cheap to produce, supply grows faster than attention. The hit-driven economy of the web is likely to get harsher, not gentler. That is also why curation platforms may matter more than engines over the next few years.

If you are deciding where to build, a rough guide from our reading: pick the web if your game is session-based, social or wants to be shared by a link, if your art budget is modest, or if you plan to iterate with an AI coding agent. Pick a native engine if your game is large, premium-priced or needs deep platform features. Use both if you can; Godot, Unity and Three.js projects all export or run on the web, so the choice does not have to be final. Our look at what the big engines are shipping, in Unity and Unreal's AI tools, is the native-side companion to this piece, and AI game makers compared covers how the no-code builders, including those that output browser games, stack up. CreateGame.ai itself is pre-launch and aimed at this browser-first end of the market.

How we reviewed this

We worked from public sources dated up to 2026-10-09: the npm downloads API (queried directly for the packages and windows shown), the caniuse feature data in its public GitHub repository (read at six commits across time), GitHub star counts via the GitHub API, engine release notes and documentation (Godot, Unity via secondary reporting, Three.js via a studio's summary), press coverage of Poki, CrazyGames, Discord, YouTube Playables and Telegram, coverage and announcements for fly.pieter.com and both Vibe Jams, and a published summary of the LMArena coding board. We did not run any of the games, engines or benchmarks ourselves.

What we could not verify. Poki's 625 million and 100 million figures come from a CB Insights summary, not Poki's own post. CrazyGames' 50 million is a company social post. We found no revenue for either portal. The Discord 25% usage figure appears only in secondary sources, and its date is unclear. YouTube Playables and Telegram numbers are from vendor blogs. Vibe Jam 2026 entry and play counts come from a third-party guide, and we found no official winners list. The Unity 6.6 and Babylon.js 9 details come from secondary reports of the official announcements. The Three.js download chart uses npm API windows that contain at least one missing-data day, and its upward trend is partly an artifact of a registry-wide increase we cannot fully explain. The coding-arena table is from a news summary of the leaderboard, not the leaderboard itself.