The best LLM for Godot

Asked for a Node2D that blinks, Claude Haiku 4.5 called AnimationPlayer.add_animation, and Godot 4.7.2 stopped on the line: Nonexistent function 'add_animation' in base 'AnimationPlayer'. That call is from Godot 3. In Godot 4 an animation goes into an AnimationLibrary first. It was one of four tasks out of 15 that Haiku 4.5 got wrong on the first try. Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol got none wrong.

Flockbay ran the test on 6 October 2026. Each task asks for one GDScript file with a stated API, in a fresh Godot project: move a CharacterBody2D with an input action, save a dictionary as JSON, paint cells on a TileMapLayer, cast a physics ray, find a path on a grid around walls. A check scene loads the file and drives it headless on the official Godot 4.7.2 at a fixed 60 frames per second. It presses input actions, steps physics, waits on timers and signals, and compares what happened with what the task asked for. A task that fails gets the check's output once, and one chance to fix the file.

Each model ran as a coding agent allowed to read and write files and nothing else: Claude Code for the three Claude models, Codex for GPT-6.1 Sol. Every task touches an API that changed between Godot 3 and Godot 4, or arrived since: move_and_slide, FileAccess, signals, tweens, Timer nodes, PackedScene, TileMapLayer, ray queries, AStarGrid2D, AnimationLibrary and lambdas.

Haiku 4.5's other three misses are the usual Godot 4 traps. Its health script emitted the died signal twice. Its grid path returned an untyped Array from a function declared Array[Vector2i], and Godot stopped it with a script error. Its coroutine that waits for a signal or a timeout never finished inside 10 seconds of game time. Shown the error, it fixed three of the four. The coroutine still failed: it returned false for a signal 0.2 seconds away with a 1 second timeout.

Speed was close at the top. The median task took the agent 8 seconds on Opus 5.5 and Sonnet 5.5, 10 on GPT-6.1 Sol and 14 on Haiku 4.5, timed from the agent's start to its reply, so Claude Code's and Codex's own start-up are inside it. Every task, the code each model wrote and what Godot printed are on the per-model pages: Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol and Claude Haiku 4.5. The 15 tasks, their checks and the runner are open source under MIT at [github.com/flockbay/godot-llm-bench](https://github.com/flockbay/godot-llm-bench), so any model can be run against the same tests.

More on this: How to make games in Codex: the Codex CLI, the ChatGPT app, or your ChatGPT plan inside Flockbay

Fifteen small scripts say whether a model knows Godot 4. They do not say whether it can build a game system and test it. So the same models took on bigger jobs in Flockbay over MCP: a physics truck, a six-legged walker, doors, a rigged character twice and a save system, with the same prompt for every model in a build. They pressed Play themselves and decided when they were done. Then both sides were played or filmed the same way, and each video ended on a verdict.

Opus 5.5 took the physics truck, with a trail more than twice as long, and the doors, where one rule about what counts as in the way jammed Sol's doors and stranded its guard. It also came first of four on a rigged character built with Flockbay's modeler. GPT-6.1 Sol took the same character built in Blender, and beat Sonnet 5.5 on a six-legged walker: Sonnet's could walk into a spot it never walked out of. Save and load was a draw, with both games matching the game left running on 25 of 26 loads.

The bill went the other way. Across the five builds Opus and Sol both played, Opus's runs came to $2.93 to $10.48 at public API prices and Sol's to $0.63 to $1.55, so Opus cost 3 to 11.5 times as much for the same job. Sonnet 5.5's walker cost about seven times Sol's at the same price per token. Every one of these runs went through a subscription, so the dollar figures are estimates at API prices, not what the runs were billed.

Claude Opus 5.5. Verdict: the strongest Godot model measured here, 15 of 15 and ahead in three of five game builds, at 3 to 11.5 times the cost of GPT-6.1 Sol.

GPT-6.1 Sol. Verdict: the value pick, 15 of 15 with two build wins and a draw, each of its builds against Opus costing $0.63 to $1.55 at API prices.

Claude Sonnet 5.5. Verdict: level with Opus on short GDScript, 15 of 15 at the same 8-second median, but it lost the six-legged walker to Sol at about seven times the cost.

Claude Haiku 4.5. Verdict: fine for plain scripts when it sees Godot's errors, at 11 of 15 on the first try and 14 after one fix.

A GDScript function is a short coding task with a strict runtime. The model has to emit Godot 4 names, typed where the project is typed, and a signal connection that actually exists. On that job the three top models tie at 15 of 15, so pick on speed and price. Opus 5.5 and Sonnet 5.5 had the lowest median time, 8 seconds, and on 3 October Anthropic's pages listed Sonnet 5.5 at half the API price of Opus. Haiku 4.5 handles the plain ones, a mover, a label, a save file. Check its signals, typed arrays and coroutines, because three of its four misses were there, and the fourth was a Godot 3 call.

Scene editing through MCP is tool use. The model reads a scene, calls a tool, reads the tool result and calls the next one. Every build above was that job, done through Flockbay's MCP tools, and on the doors the difference was how hard each model tested its own work. Opus made 67 model calls and 5 Plays and wrote a three-minute stress test with the player in the doorways. Sol made 23 calls and 2 Plays, and its 46 checks moved the player out of the house first. The Claude docs read on 3 October 2026 say Opus 5.5 and Sonnet 5.5 return an error when a request forces a specific tool, so an agent loop on them should let the model choose.

A long agent loop is a context-window problem before it is a quality problem. Every tool result stays in the conversation until something summarises it. On their own pages, read on 3 October 2026, Claude Opus 5.5, Claude Sonnet 5.5 and Gemini 3.8 Flash each publish a window of about a million tokens, and OpenAI publishes 1,050,000 for GPT-5.5 on the API but 400K for GPT-5.5 in Codex. The builds show the cost side of the same loop. On the truck, the walker and the doors, Opus and Sonnet re-read 8 to 17 million cached tokens a run, and Sol 1.3 to 2.9 million. On a subscription that is usage. On an API key it is the bill.

Haiku 4.5 went from 11 of 15 to 14 of 15 with one look at Godot's output. That is the useful finding for a cheap model: the error loop matters more than the first draft. A model that runs the game, reads the GDScript parse errors and the debugger, and tries again recovers most of what it got wrong.

Claude Haiku 5.5 came out after the 15-task run and is not in it yet. In its build log it built a Celeste-style jump in Godot over Flockbay's MCP server, with the measurements sent back after each pass and never the fix. Its first pass met 4 of 6 bars, coyote time and jump buffering among them. Its third met all 6. The three passes came to about $0.86 at API prices, and the same tokens on Opus 5.5 would have come to about $7.60.

A local model is the cheapest of all, and none ran these tasks. If you run one, give it the same loop: a headless run and the full error text after every change. Making games with local LLMs covers the setup.

More on this: Trying to call an async function without "await" in Godot 4

A table that names one best model for Godot is wrong the day the next model ships. Claude Opus 5.5 came out on 22 September 2026 and Claude Sonnet 5.5 on 28 September. Six days apart, both replaced the recommendation a spring article would still be printing.

So every score here carries the day it was measured. The 15 tasks ran on 6 October 2026, and the game builds went out between 5 and 10 October. When a model ships, the same tasks run again on the engine Flockbay ships, the model gets its own dated row, and this URL stays the same.

What lasts longer than a score is the job. A GDScript function, a scene edit through an MCP tool call and an agent loop that runs for an hour do not want the same thing from a model. The vendor sections below are dated to the page they were read from.

More on this: Best AI game maker

The 15 tasks are small, single-file and checked by running them. They say whether a model's Godot 4 code works. They do not say whether it designs a good level, picks a good art direction or keeps a 50-file project coherent.

Each game build is one run per model and one reviewer's verdict, so treat a single win as a data point, not a ranking. The times include the agent, so Claude Code and Codex are part of what they measure. And the dollar figures are estimates at public API prices for runs that went through subscriptions.

Dated 3 October 2026, from Anthropic's model page. Released 22 September 2026. Model id claude-opus-5-5. Anthropic describes it as built for long-running agentic coding and knowledge work. Context window 1M tokens, max output 128K, adaptive thinking always on, default effort medium, knowledge cutoff June 2026. API price on that page: $4 per million input tokens and $20 per million output tokens.

For a long Godot session that is the relevant line: long-running agentic coding, and a 1M window. The limitation on the same page is that forced tool use returns an error, and thinking cannot be turned off. The page does not mention Godot, GDScript or MCP. Flockbay's runs are the Godot evidence: 15 of 15, and ahead of GPT-6.1 Sol in three of five game builds.

Dated 3 October 2026, from Anthropic's model page. Released 28 September 2026. Model id claude-sonnet-5-5. Anthropic calls it the best combination of speed and intelligence. Context window 1M tokens, max output 128K, adaptive thinking, default effort high, comparative latency fast, knowledge cutoff June 2026. API price on that page: $2 per million input tokens and $10 per million output tokens.

Forced tool use returns an error on this model too. Setting temperature, top_p or top_k off the default returns a 400. The page does not mention Godot. In Flockbay's runs it passed 15 of 15 at the lowest median time, then lost the six-legged walker to GPT-6.1 Sol at about seven times Sol's cost.

Dated 3 October 2026, from OpenAI's API model page and the 23 April 2026 announcement. The API page calls GPT-5.5 a flagship model for complex professional work, with a 1,050,000-token context window, 128,000 max output tokens, and a knowledge cutoff of 1 December 2025. API price on that page: $5 per million input tokens and $30 per million output tokens. Prompts over 272K input tokens are priced higher for the whole session.

The announcement says GPT-5.5 in Codex has a 400K context window, on Plus, Pro, Business, Enterprise, Edu and Go. A ChatGPT sign-in runs in Flockbay on any plan, Free included, and Codex is how it builds. Neither page publishes a Godot score. The GPT model in Flockbay's runs is GPT-6.1 Sol in Codex, and OpenAI's page for it was not read for this edition.

Dated 3 October 2026, from Google's Gemini API model page, which lists Gemini 3.8 Flash as stable. Google describes it as its most intelligent Flash model, for long-horizon software engineering and autonomous agents. Input token limit 1,048,576. Output token limit 65,536. Function calling, code execution and structured outputs are supported. Computer use is marked preview. Thinking supports low, medium and high, and the page says minimal returns an error. The page's latest-update line says September 2026. It does not mention Godot.

Gemini is not one of the sign-ins in Flockbay's chat, and it was not in Flockbay's runs. Gemini CLI connects over Flockbay's MCP server like any other agent. It belongs in this comparison because the search is for a model.

Dated from the Qwen team's post of 22 July 2025, which is the open-weight card this edition uses. Qwen3-Coder-480B-A35B-Instruct is a mixture-of-experts model, 480B parameters with 35B active. The post says it supports 256K tokens of context natively and 1M with extrapolation, and that it was trained for agentic coding and tool use. The same post compares it with Claude Sonnet 4, which is two generations behind the Claude models above. Treat that comparison as historical.

256K native context is the smallest window in this set. A long Godot agent loop that keeps tool results will fill it before it fills a 1M window. Running the weights yourself is the reason to pick it: the machine is yours, and the vendor page is not metering the turn. It was not in Flockbay's runs, and no Godot score is on the post.

Flockbay builds and play-tests a whole Godot game for free: free AI every day or your own Claude or ChatGPT, on the official Godot engine. On a Claude plan, Pro or Max, the model menu lists the models Claude Code offers your account. Signed in with ChatGPT, on any plan, Free included, it lists your account's models. Your own agent, such as Claude Code, Codex or Cursor, connects over Flockbay's MCP server and brings its own model. With none of these, Demo, Flockbay's hosted model, is free each day. Claude Free does not run here, and plans for more hosted AI are optional. The setup is Claude or ChatGPT with Godot.

More on this: Godot MCP: connect Claude Code, Codex or Cursor to your Godot game