GPT-6.1 Sol (Fully Tested) + Dots & All DevDay Launches Explained: IT BEATS OPUS 5.5!?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: GPT-6.1 Sol (Fully Tested) + Dots & All DevDay Launches Explained: IT BEATS OPUS 5.5!? · Channel: AICodeKing · 20:14 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, AI models, Local and open-source models

Summary

The video reviews OpenAI’s GPT-6.1 Soul, outlines other DevDay announcements, and tests the model on eight KingBench 3 coding tasks using Codex with medium reasoning effort. The tasks cover simulations, 3D objects, a game, an SVG, a counting problem, and a local fine-tuning workflow; the reviewer evaluates both appearance and functionality. Soul scores 78 out of 80, or 97.5%, with the main observed issues being elevator dispatch and intersecting table geometry. The reviewer concludes that Soul performed well in these tests, while emphasizing that generated work should still be inspected.

Key points

The video gives API pricing and specifications for GPT-6.1 Soul, while clarifying that Codex usage follows separate plan and usage rules.
DevDay announcements covered Dots, faster token generation, reusable Codex Cloud environments, code review and security tools, and computer-using agents.
ChatGPT updates included plugins, shared spaces and pages, team tasks, meeting notes, and using ChatGPT allowances in participating apps.
Eight KingBench 3 tasks were tested in Codex with medium reasoning effort; each task was scored out of 10.
The elevator simulation worked but sometimes left an available elevator idle, causing unnecessary waiting.
The contact lens case, panda SVG, archery game, counting problem, local fine-tuning workflow, and wristwatch each received 10 out of 10.
The folding table had smooth motion but its parts intersected during the folding animation, earning 8.5 out of 10.
GPT-6.1 Soul scored 78/80 (97.5%); the video compares this with scores from earlier reviews and notes the tools and settings differed.

Chapters

Start	Chapter	Summary
00:02	GPT-6.1 Soul: pitch, pricing, and specifications	The reviewer introduces Soul, its intended strengths, API pricing, and the distinction between API specifications and Codex usage.
01:34	DevDay announcements beyond the model	The video surveys announcements for Dots, Codex, agents, ChatGPT, developer tools, collaboration, and enterprise customers.
10:25	KingBench 3 test setup and elevator simulation	The reviewer explains the eight-task scoring approach and evaluates the elevator simulation, noting a dispatch weakness.
12:27	3D lens case, folding table, and panda SVG	Soul earns full marks for the interactive lens case and panda illustration, while the folding table loses points for geometry collisions.
15:02	Archery game, counting problem, and local fine-tuning	The reviewer tests the game and its saved scores, checks the counting solution, and inspects the locally trained panda-fact app.
18:40	Wristwatch and final score	The wristwatch earns full marks, and the reviewer totals the results and summarizes Soul’s strengths and remaining defects.

Quotes

Now, let’s come back to GPT-6.1 Soul and see how the model handles the eight KingBench tasks in Codex. — 10:25

The interaction works in both directions, and the labels move with the caps. — 12:59

The answer, the argument, and the executable calculation agree. — 16:37

GPT 6.1 Soul finishes with 78 points out of 80 on KingBench 3. — 18:40

I would still review generated work before relying on it — 19:43

Concepts and entities

GPT-6.1 Soul, Codex, KingBench 3, Dots, Codex Cloud, OpenAI, Opus 5.5, Gemma 2B, LoRA, MCP events

Full transcript (original language)

00:02 [music] >> Hi. Welcome to another video. So, OpenAI has released GPT-6.1 Soul. And the pitch here is quite interesting. It arrived on September 29th with OpenAI positioning it as a model that gets close to Astra’s performance while costing less. Now, I want to see what that means when you actually give it a coding task and open the result. The focus is complex coding, computer use, and professional work. Think of jobs where the model has to

00:33 follow several requirements, work through a project, and check what it has made. OpenAI’s comparison with Astra gives us something to investigate, but I want the finished examples to do the convincing here. For the API, the standard base price is $2 per million input tokens and $10 per million output tokens. Cached input is 10 cents per million. Those rates apply to requests with up to 272,000 input tokens. Longer requests and other processing options have different rates. So, if you’re building

01:03 something around the API, the amount of context you send still matters. The API also lists a context window of just over 1 million tokens and a maximum output of 128,000 tokens. Those are the model’s API specifications. For this video, I’m testing it through native Codex, and Codex has its own context handling and usage rules. If you sign in with ChatGPT, your plan allowance and credits determine the billing, so the API price isn’t a price tag for each Codex task. Now, Soul was only one part of Dev Day.

01:34 OpenAI also announced changes to agents, Codex, ChatGPT, and the developer platform. So, before we get into KingBench, let’s go through what else was announced and why you might care about it. The biggest one is Dots. These are always-on agents powered by GPT-6 Astra with their own cloud computer and browser. You give a Dot an ongoing responsibility, connect the tools it needs and it can keep working across projects while learning your preferences from feedback. For a developer, that could mean following

02:04 customer feedback, investigating bugs, and preparing changes for review. For a creator, it could mean keeping research and production materials organized as a project changes. You can reach your dot through chat GPT, Slack, or Teams. Including voice conversations in chat GPT. You can inspect its work and set rules for what needs approval. Dots are rolling out to pro and business premium users in eligible markets with an admin enabled beta for enterprise, EDU, and healthcare. OpenAI also previewed specialist dots

02:35 for dedicated company responsibilities and work with Microsoft Agent 365. Those specialist deployments are starting with enterprise pilots. Then there is ultrafast. OpenAI says this premium tier can generate tokens up to eight times faster in Codex and up to six times faster in the API. Astra ultrafast is available now while GPT-6.1 Soul ultrafast is coming soon. That could make repeated coding iterations feel much quicker although token generation speed isn’t the same thing as finishing an entire project that much faster. Access and cost matter here.

03:07 The new pro 500 plan is $500 a month with 25 times the plus usage allowance and access to ultrafast. Eligible enterprise customers also have an ultrafast option. On the API, it has separate premium pricing. Personally, I would judge that upgrade by how much waiting it removes from my actual workload. Codex Cloud is getting reusable development environments. You can have Codex prepare the repositories, dependencies, and tools, review the setup, and publish an environment for future tasks. Each task gets its own

03:38 working state. I like this because repeating the same setup for every job is one of the less interesting parts of working with coding agents. You can also access Codex remotely and follow the work from other devices. The refreshed Codex CLI adds voice control, a slash agents view, and improvements to prompt editing, resuming sessions, work trees, and the terminal interface. Those changes are useful when you’re spending a whole session moving between several coding tasks. Code review has a more complete desktop experience, too.

04:08 You can inspect summaries, diffs, comments, and checks, then ask Code X to investigate a possible issue. GitHub support is generally available, while GitLab merge request support is in preview. Automatic cloud reviews can take a first pass when configured, so you have findings to inspect when you return. Code X security cloud goes further into repository security. It can scan connected GitHub repositories, check new commits, investigate findings, and prepare fixes in the cloud. The product is in research preview, and

04:39 Open AI says it includes access to models from Daybreak Blue without a separate Daybreak application. The useful part is having evidence and a proposed fix to review with the work continuing even when your laptop is closed. On the API side, the existing agents API now supports computer use through an Open AI hosted browser. That gives developers a way to build agents which navigate websites and interact with software alongside tools, context management, and delegation. Your application handles website approvals and sign-in requests. GPT 6.1 Soul also

05:13 gets beta multi-agent support in the responses API, so an application can let it divide suitable work among sub-agents. There is also a preview of Bedrock managed agents developed with Amazon. This brings the core agent capabilities into AWS with integration into AWS resources. If your application and infrastructure already live there, having the agents run entirely inside AWS is the main attraction. Another announcement is the decisions API. This focuses Luna on questions with a fixed set of possible

05:45 answers using text or image context. Think of classifying incoming requests or deciding which action an agent should take next. It is in limited preview with a broader release planned in the coming days. For businesses handling sensitive information, OpenAI introduced private intelligence. For approved zero data retention customers, private safety processing performs automated safety checks without giving OpenAI personnel access to the underlying content. Private inference is a separate preview coming this fall, combining confidential computing with verifiable controls. Those are specific

06:17 deployment options that a company would evaluate against its requirements. Now, ChatGPT itself is becoming more customizable. Plugin extensions let developers build sidebar apps, panels beside a conversation, and their own file viewers and editors. So, a plugin can give you an actual interface for working with your data. Composer mentions are currently desktop only, and web extensions for free and go are coming soon. OpenAI also announced improvements to plugin creator, plugin submission feedback, and discovery through rankings and recommendations.

06:48 Shareable profiles give people a place to showcase sites and their plugin activity. For someone building tools, that makes it easier to show people what they have made and help them find it. Sites can now use supported plugins in eligible workspaces. Imagine a shared project dashboard where you see your own assigned issues and a teammate sees theirs. Each visitor uses their own connected account and permissions. OpenAI also highlighted simpler site automations to help keep shared information current. MCP events adds another useful piece. A connected service can tell ChatGPT when

07:20 something changes. You choose what to watch and how it should respond. For example, a new issue could trigger a task to investigate it and prepare a plan. This uses support for the proposed MCP events specification, and it gives developers a way to build workflows around incoming events. There is also an update for people building browser interfaces. Annotations extensibility lets a website provide richer context and custom controls when you select something and ask for a change. A design tool could let you select a specific canvas object or preview a

07:52 component variation before sending the request. That helps make visual feedback more precise in the desktop apps built-in browser. For collaboration, OpenAI introduced ChatGPT space and pages. Space brings shared files and knowledge together, while pages are documents that people and agents can work on together. You can write, research, add charts or images, leave comments, and mention ChatGPT or your dot to make changes. For a team, I can see the appeal of having the working document, its context, and the AI assistance in the same place.

08:23 Collaborative slides are coming in the next few weeks. The announced experience lets teammates and agents edit a deck together starting from a conversation or template, then present it or export it to PowerPoint or Google Slides. Space is available on Pro, Business, and Enterprise. Creating and editing pages starts on desktop and web with mobile editing coming later. Teams and shared team tasks cover the recurring work. A team can maintain instructions for a weekly update or an event-triggered workflow using its configured connections.

08:54 Those tasks use workspace credits. OpenAI also expanded ChatGPT in Slack and Microsoft Teams, where co-workers can contribute context and refine the same result together. In enabled channels, participation doesn’t require every teammate to have an individual ChatGPT license. The meetings plugin connects to that workflow as well. It creates personalized notes and action items in space, and you can ask ChatGPT to turn them into follow-up work. It is in beta on macOS for Pro and

09:24 Business users with Enterprise and broader platform support coming later. The audio is deleted once the notes are ready, and you should get everyone’s consent before taking note. Sign in with ChatGPT is another announcement I think people will notice. Alongside signing into other apps, eligible plus and pro users can use their ChatGPT allowance in participating tools with controls over how much each one uses. OpenAI named partners, including Devin, Notion, Vercel, and T3. That usage

09:54 counts against your plan limits, so it gives you more places to use the allowance you already have. Finally, OpenAI Marketplace lets eligible enterprise customers apply part of their existing OpenAI commitment toward approved partner software. The initial partners span design, customer support, business applications, security, and other categories. This is mainly a purchasing option for companies already spending at that scale. So, there’s a lot here beyond the model release. The announcements that interest me most for coding are reusable cloud

10:25 environments, browser-capable agents, and the improved review tools. Now, let’s come back to GPT-6.1 Soul and see how the model handles the eight KingBench tasks in Codex. Now, for these runs, I explicitly selected medium reasoning effort. Each task gets a fresh folder with the same KingBench 3 request used for that test. That gives us an output we can open and inspect on its own, and I’ll score what the model actually delivers. There are eight tasks, each worth 10 points, for a total of 80. We have an

10:56 elevator simulation, a contact lens case, a folding table, a panda SVG, an archery game, accounting problem, local panda fine-tuning, and a working 3D wrist watch. That’s a useful mix because each one asks for something different. The elevators need to keep their passengers moving correctly. The 3D objects need convincing geometry and interactions. The game needs hits, timing, and a working leaderboard. And the fine-tuning task needs an actual local model behind the page. I’ll check both how these results look

11:26 and whether the requested behavior works. So, let’s get right into it, starting with the elevators. The first task is an elevator simulation. I want three elevators, each carrying one person, with passengers waiting on different floors and random destinations that appear when you hover over them. So, here’s what Soul made. We have seven floors, a control for how many people to add, and these little passengers waiting beside the shafts. The colors are soft, the layout is clear, and the counters make it easy to follow what’s happening. I like the presentation quite a lot. Let’s add five people on the third floor.

11:57 One passenger boards, the others stay behind, and the cars continue collecting and delivering them. We can add someone on another floor while that is happening. The destination tooltip works, and there are useful pause and speed controls as well. There is one weakness in the dispatch. If I add two people on the sixth floor, an elevator can be sitting right there and still stay idle while the other two travel up to collect them. Everyone eventually gets delivered, but that is an unnecessary wait. So, I’m giving this 9 and 1/2 out of 10. The core simulation works well, with a

12:27 specific scheduling issue to improve. Next is the contact lens case. This needs to be a 3js model with pronounced left and right markings and caps that open when I click on them. This one looks really good. We have a connected case, two ribbed caps, raised letters, and a clean little interface around the object. The details make it feel like a recognizable physical product. Now, I can click the left cap directly, and it lifts away from the case. The right one stays closed until I click that, too. Both sides have proper recessed wells

12:59 underneath, with transparent lens shapes inside. If we rotate it, we can inspect the sides and the underside of the raised cap. I also checked that each cap can close independently. The interaction works in both directions, and the labels move with the caps. This is all in one HTML file, using 3js from a CDN. It meets the request very well, so this gets 10 out of 10. This is a strong example of Soul handling presentation and the actual object together. The third task is the folding table. Moving the slider to the right

13:29 should unfold it and moving it left should fold it with smooth motion in three dimensions. Again, the initial presentation is appealing. There is a wood surface, a metal frame, camera presets, and different finishes. The slider responds smoothly and the table really does have separate moving parts. Both end points work and we can go back and forth through the animation. But watch the legs as we stop halfway. They pass through the opposite piece of wood. Further along, the two frames cross through each other underneath the center. Looking from the front makes

14:01 that easier to see and checking the geometry confirms the collision. That matters for this particular task because the folding mechanism is the main thing we are testing. The movement is smooth but the physical arrangement needs work. So, I’m giving it eight and a half out of 10. It delivers an attractive interactive demo and it loses points for those intersections during the fold. Now, for the panda SVG. The request is simply a panda eating a burger and this is a nice result. We have a clearly recognizable panda holding the

14:31 burger with both paws. There is a bite taken out of the bun. The burger is up against the mouth and the layers of lettuce, cheese, and filling are easy to read. The little crumbs and the expression give it some character, too. It renders as a valid SVG and the proportions and layering work. A static illustration is completely fine for this prompt. I don’t see a missing requirement here. So, this gets 10 out of 10. It’s a simple task and Soul handles it cleanly. Now, let’s come to the archery game and I really like this one. You get a woodland range with four

15:02 targets, a visible bow, and clear instructions for drawing and releasing a shot. The arrows follow a curved path and you can see the impact when they reach a target. I played two complete rounds with four shots and four hits in each. The first took 43.75 seconds and placed me seventh on the a leaderboard. Then I improved to 29.03 seconds, which moved me up to sixth. So, the faster time actually improved my position exactly as requested. The personal best updated, too, and refreshing the page kept that

15:33 score. The board starts with sample records, then adds your own saved runs. The separate logic checks also passed for misses, repeated targets, and score ordering. The presentation is clear, the game is playable, and the complete scoring loop works. I didn’t find a failure in the requested features, so this gets 10 out of 10. Next is the counting problem. We have 60 pairs arranged across 20 columns and three rows. And the question asks, “How many orders satisfy the movement rules?” You start on the top row, end on the bottom row, and visit every pair exactly once.

16:04 Horizontal movement goes forward and wraps around, while vertical movement goes between neighboring rows. The expected answer is 20,460, and GPT-6.1 Sol got it right. It also provided a mathematical argument and a Python script that counts the possibilities. The explanation fixes the starting column, finds 1,023 valid paths, and then multiplies by the 20 possible starting columns. The reasoning handles the wrap-around and checks that the paths remain connected. I reviewed the delivered script and ran it unchanged.

16:37 It returned the same answer, and the smaller cases agreed with independent checks as well. The model had already executed its own solver successfully during the task. So, this is a strong result. The answer, the argument, and the executable calculation agree. 10 out of 10. Now for the local fine-tuning task. I asked it to create a panda data set, fine-tune Gemma 2B, and build a local webpage that generates a fact whenever I refresh it. This is a bigger workflow than the earlier demos. Sol created 51 sourced panda facts and

17:09 turned them into 306 examples with separate training, validation, and test sets. It then ran actual LoRA training locally and saved the adapter weights. So, we have the data set, the trained model adapter, and the app together. The Codex run included its own debugging. Soul adjusted an unstable learning rate, selected the checkpoint with the best validation loss, and fixed a threading issue in the local inference server. It followed those changes with generation and browser tests. Here’s the finished page. I like the soft colors, the panda

17:39 illustration, and the large fact card. The first fact I checked was about a wild panda’s lifespan. The button generated a different fact about conservation, and refreshing the page generated one about accidental snares. The expandable source reference worked it, too. Each request runs the local Gemma model with the trained adapter. The app supplies a reference fact to help guide the generated sentence, and it keeps the displayed response to one sentence. That’s useful context. This is a small, specialized app with reference

18:10 grounding. I verified the saved adapter, the completed training, and the working page. Once setup is complete, generation runs locally. This fulfills the requested workflow, so I’m giving it 10 out of 10. Finally, we have the 3D wristwatch, and there’s quite a bit to look at here. You get a metallic case, a crown, lugs, a stitched strap, a buckle, and actual strap holes. Rotating the camera lets you inspect the thickness of the case and the rest of the watch. The hands move, and there’s a

18:40 separate subdial for the away time. This gets 10 out of 10. So, this is the final chart. GPT 6.1 Soul finishes with 78 points out of 80 on KingBench 3. That works out to 97.5% or 9.75 out of 10 overall. For reference, the earlier GPT-6 Soul result was 82.5%, GLM-5.3 was 91.25%, and Opus-5.5 was 93.75%. These are the retained scores from those earlier reviews. The tools and settings

19:12 differ, so treat them as reference points alongside this fresh Codex run. The gap from the earlier Soul score is 15 percentage points. What stands out to me is how much of the requested behavior survived inspection. The lens caps work independently, the archery scores survive a refresh, and the watch handles a fractional hour time zone correctly. Those details make a demo useful once you start interacting with The table also shows why I still check the output closely. Its presentation is polished, but the folding geometry has a real defect, and

19:43 the elevator has a dispatch decision that adds unnecessary waiting. Those are the two specific improvements I would ask for next. Overall, GPT-6.1 Soul works really well in these Codex tests. I particularly liked the contact lens case, archery game, and wristwatch. This is a small set of tasks, and I would still review generated work before relying on it, but the results here give me a good reason to keep using Soul for interactive prototypes and coding work. Overall, it’s pretty cool. Anyway, let me know your thoughts in the comments. If you like this video,

20:14 consider donating through the Super Thanks option or becoming a member by clicking [music] the join button. Also, give this video a thumbs up and subscribe to my channel. I’ll see you in the next one. Until then, bye.