OpenCode Go + CodeGraph: Can Cheap Models Do the Job? — Transcript
Full transcript
- 0:00Muse Spark used almost twice the tokens with CodeGraph as it did without it in my OpenCode test.
- 0:07DeepSeek and MiniMax used fewer tokens with CodeGraph than without it.
- 0:12So the model made a big difference in these runs.
- 0:16In my earlier graph comparison, I asked if you'd like a test with cheaper models.
- 0:22You suggested Muse and DeepSeek.
- 0:25CodeGraph was my pick, so I tried it in OpenCode with several cheaper models.
- 0:30The new subscription budget was ten dollars.
- 0:33I couldn't finish every comparison before hitting the limits.
- 0:36I'll show you which models used fewer tokens, and whether the code was worth keeping.
- 0:47The task was another feature for The Gray Cut, my video editing app.
- 0:53It already picks screencasts to match what I'm saying.
- 0:56I still add the small window showing my face by hand in Premiere.
- 1:01I wanted the app to place that window automatically, with a rounded white border.
- 1:06The difficult part was framing the face
- 1:09My position sometimes shifts slightly between recordings, so the app needed to center my face.
- 1:16The app needed to find my face and choose how much to zoom in.
- 1:21Then that framing should stay steady throughout each recording.
- 1:25The small window also needed to follow the edited talking head footage.
- 1:30Speeding up a screencast should leave my face synchronized with my voice.
- 1:34I also needed to keep editing the result in Premiere, with as little manual setup as possible.
- 1:41Okay, so OpenCode Go is the ten dollar monthly subscription I used for model access.
- 1:47OpenCode itself runs the coding agents.
- 1:50Go is an optional way to pay for the models they use.
- 1:54The plan was to test each model with CodeGraph and without it.
- 1:58Each attempt started with its own copy of the same code and the same task.
- 2:03Claude Opus 5.5 supervised from Claude Code.
- 2:06It checked the plans and reviewed the completed code, then gave each finished attempt one chance to fix problems.
- 2:14The models in OpenCode wrote the code themselves.
- 2:17They didn't delegate work to other agents.
- 2:21The ten dollars only covered OpenCode Go.
- 2:23Claude wasn't included in that.
- 2:25But Go has usage limits for five hour periods.
- 2:30There are weekly and monthly caps too.
- 2:33Kimi K3 hit a limit during implementation.
- 2:37GLM 5.3 couldn't start coding, and Argon later hit a weekly cap.
- 2:42Leaving Argon overnight mostly meant waiting for resets.
- 2:46I eventually stopped it with the feature unfinished.
- 2:49Some attempts also ran at the same time.
- 2:52So the elapsed times wouldn't give a fair speed comparison.
- 2:56The privacy terms depend on the model option.
- 2:59Muse and Argon Contributor allow training on your prompts and responses.
- 3:04I also tried the free options available during this test.
- 3:08DeepSeek and MiniMax both finished attempts with and without CodeGraph.
- 3:13GLM Flash didn't finish both.
- 3:16Next, I'll compare the models that finished both attempts.
- 3:19Hitting a subscription limit doesn't tell me whether a model can build the feature.
- 3:24Meta's Muse Spark 1.3 Contributor had the biggest increase.
- 3:29Without CodeGraph, it used about thirty-five and a half million tokens.
- 3:34With CodeGraph, it used about sixty-nine million.
- 3:38That's about ninety-five percent more
- 3:41Most of those tokens were text the model had already seen.
- 3:45It could read that text again from a cache at a lower price.
- 3:49I counted that reused text each time, just as in my earlier comparisons.
- 3:54The cost calculations use the cheaper cache price.
- 3:57CodeGraph helps a model find related code.
- 4:00For example, it can show which functions call the function being changed.
- 4:05Muse used CodeGraph fourteen times, all while working on its plan.
- 4:10And that planning actually used slightly fewer tokens than planning without CodeGraph.
- 4:16The extra usage came almost entirely while writing the code
- 4:21The reporting script counted the model responses during that phase.
- 4:25Without CodeGraph, there were a hundred.
- 4:27With CodeGraph, there were
- 4:29twenty-four.
- 4:31Each response also had more text to process on average.
- 4:36So there were more responses, with larger inputs.
- 4:39The logs show Muse repeatedly trying to fix the picture preview.
- 4:44It adjusted the crop and investigated why the speaker wasn't visible.
- 4:49Most of the extra usage came during implementation, including these fixes.
- 4:54The logs don't prove that CodeGraph caused those problems.
- 4:58And the separate preview check still showed an upside down picture.
- 5:03The window
- 5:03was at the top right, although it belonged at the bottom right.
- 5:07Muse's own orientation tests passed anyway.
- 5:11So even after all that work, its tests missed a visible problem.
- 5:16If these comparisons help you choose what to try, subscribe for the next one.
- 5:20Join my Discord if you want to suggest another model or discuss the results.
- 5:25The link is in the description.
- 5:27DeepSeek V4.1 Flash had a different result.
- 5:31With CodeGraph, it used about twenty-four million tokens.
- 5:35Without it, almost twenty-seven million.
- 5:38That's about ten percent less
- 5:41Planning used more tokens with CodeGraph.
- 5:44Most of the overall saving came when DeepSeek fixed problems after the review.
- 5:49That repair stage took thirty-eight recorded model responses with CodeGraph, against seventy-six without it.
- 5:57CodeGraph didn't reduce usage at every stage.
- 6:00Fewer tokens also didn't mean a lower calculated cost.
- 6:03Using the Go rates recorded for this test, the CodeGraph attempt comes to about fifty-one cents.
- 6:10The attempt without it comes to about forty-seven cents.
- 6:13The CodeGraph attempt used more input that wasn't cached.
- 6:17That costs more than reading cached text.
- 6:21DeepSeek was free in these runs.
- 6:23Those amounts show what the recorded usage would cost at those rates; they weren't extra subscription charges.
- 6:30Astra recommended the CodeGraph version as the best starting point based on the code review and separate checks.
- 6:38It had controls to prepare the face framing and retry when preparation failed.
- 6:43It also framed the sample recordings correctly during those checks.
- 6:47The face detection runs locally on the computer.
- 6:51Choosing the crop doesn't require sending a request to an AI provider.
- 6:55A simulated file replacement exposed another bug.
- 6:58The app still accepted the old face crop.
- 7:02MiniMax M3.1 Flash Free had the biggest token saving among the models that finished both attempts.
- 7:09With CodeGraph, MiniMax used thirty-nine percent fewer tokens
- 7:14That's about sixty-four million, compared with a hundred and five million without it.
- 7:20Most of that saving came while writing the code.
- 7:23The CodeGraph attempt needed fewer model responses and processed less text per response.
- 7:29But a separate check found a problem with the rounded window.
- 7:33The app uses a shape called a mask to hide the picture outside the rounded edges.
- 7:39MiniMax placed that shape completely outside the picture, so it couldn't clip the picture correctly.
- 7:46The border also stayed visible when the speaker window should have disappeared.
- 7:51So the better token count still left problems with the feature.
- 7:55Both MiniMax versions ask an AI provider to choose the face framing.
- 8:00The separate checks didn't test that part with a live provider.
- 8:05All six completed attempts passed their automated test suites.
- 8:09Separate checks still found a bug in every completed version.
- 8:13So a passing test run was no guarantee that the feature was ready.
- 8:18Here's the final comparison.
- 8:20Pause here if you'd like to look at the numbers.
- 8:23I tried the DeepSeek version with footage from my Codex Framework Arena video.
- 8:28The results aren't acceptable
- 8:30I'll work on it with a smarter model and make sure everything works correctly before merging it in.
- 8:37For the final version, I decided to drop the rounded corners.
- 8:41They added too many manual steps in Premiere, and I wanted to move fast.
- 8:46For CodeGraph, adding the same tool gave different results with different models.
- 8:51It saved tokens in some attempts and nearly doubled them in Muse's.
- 8:55I only ran each comparison once, so I'd want more runs before trusting the size of those savings.
- 9:02But for this feature, the cheaper models disappointed me.
- 9:06Even the version Astra recommended still needs fixing.
- 9:09After a year using frontier models in Claude Code and Codex, this feels like a step backwards.
- 9:16I tested Kimi K3 back when it came out.
- 9:19The results were decent, but it was very, very slow.
- 9:24This time, the quota stopped it before it could finish.
- 9:28Cheaper models might help with simple tasks if the plan is detailed enough.
- 9:33I'd still check their work.
- 9:35For my earlier Graphify and CodeGraph comparison, watch this video
About this transcript
This page contains the full transcript of OpenCode Go + CodeGraph: Can Cheap Models Do the Job? by The Gray Cat, generated from the public captions YouTube serves with the video. The transcript has 1,375 words across 133 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.