It's been three weeks since #008, which ended on a model nobody could get yet and a weekly cap that burned out in a day. Since then Anthropic and OpenAI shipped frontier models on the same afternoon, a Sonnet 5.5 release nearly matched them, OpenAI reopened Pro $200 with about half the allowance, and a run of Steam pages went up. The group passed 500 members somewhere in there too. A lot of what follows is you all sorting out which model to use for what, and what it costs now.

Early September: Spawn, ownership, and the walls

The window opened on the old argument about where an AI-built game should live. A few of you are happy in three.js and WebGPU and think the browser is good enough; at least one of you left the web stack after a game crawled on a handheld and wants standalone binaries now. Spawn kept picking up converts in between, on a pragmatic pitch: build it there, watch the numbers, prove it's fun, then build it properly. If you've never run live ops, the built-in multiplayer is most of the appeal. @OP__Nico's Pokémon-meets-Mario-Party mash runs "multiplayer games that are always filled with people: either AI or real people in the main lobby."

Ownership came up next, and Spawn's FAQ doesn't dodge it. An outside coding agent can help you port to another engine, but "Spawn does not provide a ready-to-run Steam export," and "keeping a copy of your files is different from having a standalone game." The same page says Savi, Spawn's builder agent, now runs on Claude Opus 5.5 by default.

On Sep 23 somebody asked whether a Spawn game could go to mobile stores or Steam. "It's not 'your game' on spawn," one of you replied. Another said the terms give you ownership and "it's basically a repo you can fork." Asked whether people would just build on the free compute and leave, the Spawn side said "our bet is that the platform is good enough you won't want to export" and that it had seen "about 10x more imports than exports."

Then the money. Somebody asked what justified Spawn keeping half, and the answer from the Spawn side was that their half pays for the free AI tokens, hosting, live ops, moderation, distribution, and scaling if a game blows up. One of you said that made the split almost generous. Another said 50/50 is "outrageous to put any serious development effort into" while Steam, a standard engine and heavily subsidized AI are all still on the table: "we're all aspiring to make the next hit game and if we do we want to be the bigger beneficiaries." By the end of the month the subsidized-AI part of that math had started going away.

Meanwhile the craft problems kept coming, and a better model doesn't fix most of them. Vocals were "in a pretty bad state," and one of you had better luck with ElevenLabs than Suno. Another skipped generation entirely and had a cheap Opus setting write C and Python that drives VSTs, which got them a six-track soundtrack that shares instruments across keys and tempos.

Humanoid animation was the wall people named most. On Sep 15 somebody asked "so what are we doing about humanoid animations?" and called it "the biggest bottleneck." Kimodo handled wheelies and dances but struggled "with complex moves like a spinning backfist," which they put down to training for robotics, where "a robot wouldnt need to break out in kung fu psychopath killer mode." Another said "Text to animation is comically bad even for fairly basic idles," with video reference the only thing producing decent results. That afternoon somebody asked how to get a looping walk into a game. Mixamo came up first, Meshy's animations turned out to be "incompatible with mixamo and standard skeletons," and Kimodo, Cascadeur, Nira and Unimate all got pitched, with no clear winner.

3D that has to fit together got a more unusual fix. One of you handed a model the NASA systems engineering handbook as the spec for a Blender scene, and it came back with measurable requirements, explicit interfaces where parts meet, door gaps and clearances. Not prettier, but the parts fit.

Meta's Muse arrived as the cheap "grunt worker." The cleanest early loop paired it with Astra: Astra wrote the tech spec for about 2% of one person's budget, then Muse built for two-plus hours, passing its questions back to Astra.

Sep 22: Opus 5.5 and GPT-6 Sol, two hours apart

On Tuesday Sep 22 Anthropic shipped Claude Opus 5.5, which it says performs "at the level of Claude Fable 5.1 for most tasks" and "costs 40% less to run than Opus 5." Less than two hours later OpenAI shipped GPT-6 Sol and Luna, and @thsottiaux added that OpenAI was "permanently reducing the API price by 50%."

Not many of you picked a team. On launch afternoon one of you said "at this point cheaper is the moat," and two hours later had a number for it, "Sol xhigh = Astra Medium performance," at a tenth of the price. The next day they put it more bluntly, "The models are good enough." Somebody answered "I still have ceilings models cannot hit," starting with an equipment system and procedural islands.

Opus became the one you call for visual work. @NicolaManzini ran it through threejseval on day one and called it "one of the best so far." @aollivier82 asked for an Amano-style aquarium in three.js, declared "3d graphics are now a solved issue," and got a "Lost for words" from @mrdoob. Same-prompt comparisons kept everyone honest, or at least entertained: @majidmanzarpour's procedural cheetah puts GPT-6.1 Sol Max against Opus 5.5 Max, and @aollivier82's sequel test ran Sol 6.1 on the same prompt and left him "actually crying of laughter."

The workflows that came out of the week were more useful than the leaderboards. The common one goes spec, build, polish, with Sol or Muse doing the bulk implementation between an Astra plan and an Opus pass on 3D, rigging and art. One of you wrote out the long version as "Design bible → Astra breadth pass → Opus art/technical polish → engine validation → game systems → content integration."

Effort level matters too. One of you warned that at the top settings Astra "overthinks the loop and closes a lot of lean endpoints for expansion, it builds it as 'software' not a game." They now plan with Astra and implement with Sol, and rate Astra medium as the better "thinker" for games.

Opus also got used as a QA pass. @AIandDesign reopened Rummy 500 for free after an Opus overhaul that "ironed out several obscure bugs," and later had Astra and Opus build a Godot version for Windows, Mac and Linux. Elsewhere, one of you had a model build a test framework just to prove that procedurally generated levels contain no show-stoppers.

The surprise was models cutting video. One of you pointed a model at a folder of gameplay footage, gave it a piece of music, touched nothing else, and got five usable trailers in a few hours; they've since burned a week of usage on around a hundred trailer and content videos. @AIandDesign had Astra go "into the code and captured all the footage by itself, then assembled this trailer" for It Wants You To Stay. @boona11 went furthest with Low Tide: "Claude Opus 5.5 built the game, then played it, filmed it, wrote the soundtrack and cut its own trailer." One of you even had the agent set up and take its own store screenshots.

Store pages and AI disclosures

Once the Steam pages started going up, the AI disclosure was often the most interesting thing on them.

@GeneStoneGames put AirGap on Steam: a first-person stealth roguelite about sabotaging an AI datacenter, due Q4 2026. The disclosure is front and center and reads "This game was Designed by AI for AI." It's part of the pitch.

The Final Like from @willandwordx, a UK fruit-machine sim crossed with social-media horror, goes the other way and says only "A small number of game assets are AI generated." Its free Steam demo has been live since Sep 11, and the full game is listed for Q1 2027. Graven, Basalt Games' Frankenstein-inspired Gothic RPG due December 2026, scopes its disclosure narrowly: a procedural 2D rig, a reinforcement-learning animation layer, base sprite designs, and no AI for writing, voice or music.

Driver Survivor from @Obsolete_Robot ("The dead own the streets. You have a car.") is on Steam with release date TBA and says plainly that AI was used for "code, art, and marketing materials." And one of you wrote the long version: every tool named, from coding agents to image, music and video models, plus a note that if you'd rather not support games made with AI, this one may not be for you. Of the lot, the specific disclosures read better than the vague ones.

Not everything has a Steam page yet. It Wants You To Stay from @AIandDesign has a site up for its liminal survival horror across eleven endless worlds, marked "Coming Soon to Steam Early Access" but with no store link. His Radial Drift is on the App Store, and Canyon Overdrive is a vaporwave canyon shooter in the browser, "not bad for one day," made with Astra, Sol 6.1, ElevenLabs and Suno and now mobile-optimized. Shatterwake: Ascension from @nazimadakli has a new site with a "Wishlist Now" button, though so far it's a site, not a Steam page.

On Sep 29 one of you posted a PSA, "Opus 5.5 has an issue of literally copying entire github repos silently," to the point of being "very-plagiarism-like." Remakes rebuilt from old game binaries were circulating that week too. On Oct 1 somebody pointed out "you could always clone any game, just took longer." Another put the response plainly: "the old 'build small' meta is gone. it used to take a year to 'build small' now it takes a day. so you take a year and use AI to build larger."

Animation, two weeks later

Two weeks after "comically bad," @nazimadakli posted: "Opus 5.5 is also the first LLM that does animation decently. Today I animated cervid (deer) and canid (dog) families with it." A few days later he compared Opus 5.5 Ultra plus Blender against Unimate ("Opus is the clear winner right now"), adding that if Unimate scales, "it'll be far cheaper than agentic animations."

Rigging moved too. @majidmanzarpour had Opus model, texture and rig a character in Blender from a concept sheet, "in pure code," and turned the workflow into blender-game-skills. On Oct 1 he followed it with threejs-procedural-animals (announcement): 24 species "sculpted from SDFs, meshed, rigged, skinned and furred at runtime, then animated with IK," with no models, textures or hand-made rigs. @startracker is on cloth now, because "simple Tripo meshes/textures won't cut it" once characters fly and their outfits flutter. @Lunarsong put up a character-models viewer of rigged work-in-progress characters for a fantasy RTS.

On Sep 29 one of you wrote that "algorithmically autorigging anything right now sucks vs agentically rigging in headless blender," and the next morning called Blender MCP "likely deadweight" next to having the model drive Blender directly with Python. Another agreed, "MCP spends more tokens and is slower."

The same builder wants players to "generate and wear any equipment using 3d generators, across 7 different playable races and 4 body types, without a human in the loop." On Oct 1 somebody who had worked on the same problem coached it. The first pass is "rotating it till the reference image and the 3d model's render from the front match exactly," then a critic model reviews a student model's suggestions until "the critic accepts." Humanoids are still open, though. The public wins this month were creature families and cloth, not bespoke human performance.

More of the recipes are public now. Besides Majid's two repos, @AIandDesign published token-saver, a Codex setup for delegating to Muse and choosing context with Jev, and godot-game-skills, agent skills for building, debugging and verifying Godot games, distilled from months on It Wants You To Stay. @cranberryclock has Cranberry Forge, a set of three.js dev tools and templates.

Sonnet 5.5, and Pro $200 buys half as much

Anthropic shipped Claude Sonnet 5.5 on Sep 28: "30%+ faster" and "up to 30% less for most work." @NicolaManzini put early runs "in the same band as the best Opus and Astra we've seen" on his bench.

A few hours later @thsottiaux said OpenAI was reopening Pro $200 to new subscribers with a new usage calculation that will "net out at half the dollar in API spend compared to the old Pro $200 plan." Same price, roughly half the allowance. Business Insider reports it as 10× Plus, down from 20× (The Verge has the same story). The Dev Day recap the next day added a $500 Pro tier and GPT-6.1 Sol, billed as near-Astra at one-fifth of Astra's price.

Overnight on Sep 29 one of you wrote that "they are nuking subsidies." Somebody called that "a huge leap to conclusions," and the first came back with numbers, "$200 used to give us thousands of dollars in API credits (when maxed out). Now it gives $400." A third read it as "50% less usage not 90% less usage," and within ten minutes the first agreed it was about half. Their longer take was that "we will get better models but use it way less or pay more," and that "opus 5.5 level intelligence will be likely the plateau wrt intelligence/$."

People had started adjusting before the announcement. One of you found a browser-automation tool eating 21–31% of their usage, banned it in their global agent instructions, and now just tests in Chrome themselves. One of you made "an insanely cool shooter" with Astra on about a fifth of a week's budget; another is maxing out accounts this week to bank assets while the current models are at their best.

TypeSafe's Jev showed up as a decision engine inside games. One of you priced pack-movement AI at about a penny per half-million decisions and is sketching a harness where players bring their own AI to operate the game.

@AIandDesign asked publicly for "real, finished games. Blood, sweat and tears stuff. Not token-maxxed demos for social media eyeballs." @mercutio_93 posted the other half: "keeping an idea small but polished works wonders!"

Projects Worth Watching

• Hybrid MMORPG / sim — @startracker. "A true living city, with all the NPCs you see being simulated," made mostly with Astra and Opus 5.5, with Spawn as the backend for a rough first combat and dungeon pass. No play link yet.
• DEUS — @majidmanzarpour. A dark-fantasy boss rush built in conversation with Savi on Spawn. The play link is in the thread.
• Pakt at Strella's Bar — @SeloSlav. A card game on Spawn, solo or with friends, borscht included. He also made a dragon racer in a day.
• Cookie Crypt — @cranberryclock. Slide rows of rune cookies to swing swords, on Spawn. He also has a creator mod, Source Audit Workbench, for browsing and diffing your world's real source.
• No Man's Land, 1885 — @josephepstein. "An endless country you walk in third person," in the browser and "drawn entirely in code with painted surfaces," hosted on his own studio site.
• Impact Soccer — @DeepThinkLabs. "A physics-based soccer game w/ no ball": you knock opponents down or kick them in to score. Headed for Roblox, and he's asking for playtesters.
• Norngine — @SpikeRiser. A creature life sim where "tiny artificial minds learn from what happens to them… None of it is scripted." He asked for Windows beta testers in late September.
• Drift Club — @DriftClubGame. Multiplayer touge drifting in Chrome, built on three.js. Waitlist only for now, with Steam "coming soon."
• Fishies — @CRTkafa. A living browser aquarium, in alpha, that "keeps moving when you step away."

Short takes

• Pixel Pack is an arcade anthology of microgames you play in the browser.
• Mythraval, a browser wave-and-build ARPG, has "actually achieved fun gameplay state," according to one of you.
• Spaceplate, the SpacetimeDB + Threlte (Svelte three.js) boilerplate that got shared in the group, lives at the repo root now. The WebGPU branch link that went around 404s.
• @theothersuzu had Claude build an 8-bit music tracker with MIDI import. Not hosted yet.
• One of you had Astra pre-analyze a game's music tracks into a data file so enemy-destroy sounds harmonize with the soundtrack, and said handing the agent an ElevenLabs key to sort out sound and music "feels like a superpower."

Procedural Press — AI + Game Dev • Subscribe