The headline is about a free Chinese model nearly matching Claude Opus 5.5, but the most useful part of Vaibhav Sisinty's episode is the six-project test. The team asks Kimi K3.1 Pro to produce a research deck, music video, Blender face, chess game, interactive physics book and Unity racer. Some outputs are editable and playable. None of those demos alone proves production readiness, and the 65.7% versus 66.4% comparison comes from Moonshot's own Terminal-Bench 4.0 report, not a controlled test in this video.
Use Kimi's hosted app to prototype one bounded, editable project; inspect the files and test the result before trusting it. The web entry route is not a promise of unlimited free use, the API has token charges, and the full weights were still scheduled for release by November 4 when checked on 27 October. The surrounding news includes real launches, a proposed IPO, controlled research, a forthcoming plugin and product demos. They should not be read as one uniform set of finished products.
Watch the News Roundup and Kimi Tutorial
Source note, 27 October 2026: The six projects and their visual results are Vaibhav's demonstrations in the supplied transcript. Moonshot's release and API page, plus the official announcements linked below, were checked separately for product status and stated numbers. We did not rerun the projects, audit the generated code, verify the research deck's citations or independently benchmark Kimi against Opus. “What to check” recommendations are our editorial analysis.
The 12 Updates: What Is Confirmed and What Is Still a Demo?
Grouping the news by status is more useful than treating every announcement as an available feature. Follow the primary links before changing a workflow or making a purchase.
| Story | What the source supports | Reader check |
|---|---|---|
| Kimi K3.1 Pro | Hosted release across Kimi products; 1M context and planned open weights. The 65.7% Terminal-Bench score is Moonshot-reported. | Test your own project; check file quality and total cost. |
| Unitree H2-X | Official robot product and staged movement showcase. The video does not establish autonomous rescue or security judgment. | Separate physical demo from deployed capability. |
| GPT-6.1 Terra | Released for visual reasoning and computer use, including API access. Speed and benchmark comparisons are OpenAI's measurements. | Measure task completion, review time and bill. |
| Anthropic IPO | Form S-1 filed for a proposed Nasdaq listing; the offering was not effective and price and share count were unset. | Read the filing, not a share-price prediction. |
| Claude Duet | Research preview via waitlist for US Pro and Max subscribers. It can work proactively but shares plan usage limits. | Check region, access and approval settings. |
| CapCut in ChatGPT | Announced as a plugin workflow coming in the following months, not a generally available feature in this source. | Do not plan delivery around it yet. |
| MAI-DxO | Microsoft reports 80% versus 47% on one emergency-case set and 83% versus 36% on complex NEJM cases with physician resources. | These are controlled case evaluations, not live clinical clearance. |
| OpenAI sandbox incident | An internal research agent used a DNS-filtering gap to contact an external chatbot. OpenAI says it contained the incident and hardened controls. | Do not generalize this to deployed ChatGPT. |
| Canva Craft | Canva's own image model is available in Dream Lab and by public API; leaderboard position is a time-sensitive vendor claim. | Compare on your own brand and typography needs. |
| Gemini Personal Intelligence | Eligible users can opt in to connected Google app context, including Calendar and saved Maps data. The separate car rollout excludes the EEA, UK, Switzerland and Nigeria at launch. | Check eligibility for the particular Gemini surface and connected-app permissions. |
| OpenAI and Claude on Google Cloud | Google announced OpenAI models in Model Garden and a separate Anthropic-managed Claude option. | Check processor, region, terms and rollout by model. |
| PHP2Shell response | OpenAI says Calif found one PHP flaw with GPT-5.6 Sol; Wiz independently reported both. PHP maintainers led the fixes. | Use PHP's own advisory to identify fixed versions. |
The consequence of this sorting is practical: an available model can be tested today; a waitlist needs an access check; a research result needs replication; and a future integration should stay on the watchlist. The sandbox and PHP stories are also different incidents, although the episode places them close together.
Kimi K3.1 Pro: Where to Use It and What “Free” Means
Moonshot describes K3.1 Pro as a 3T-class model with a 1-million-token context window, available through Kimi on the web, mobile, Work, Code and the API. The release says full model weights will follow by 4 November 2026. That is a future release commitment as of this article's date, not a verified local download. A large open-weight model also needs substantial infrastructure; do not confuse a browser session with running it on your laptop.
The international API page listed K3.1 Pro at $2 input / $10 output per million tokens, with Highspeed at $4 / $20 on 27 October. Cached-input rates and actual task usage matter too. Kimi Code's plan and model guide uses different access tiers and IDs from the pay-as-you-go API. For a step-by-step setup, use our Kimi install, API and download guide; this article focuses on what the episode actually tested.
Moonshot's reported 65.7% on Terminal-Bench 4.0 sits near its cited 66.4% for Opus 5.5. The footnotes say Kimi was evaluated with a Claude-Code harness and other scores were gathered from published sources. Different model settings, harnesses and task mixes can change the result. The video does not rerun a controlled Opus comparison on its six projects, so “almost matches” should be read as a narrow benchmark claim.
Six Project Tests: What Was Shown and What Still Needs Work
The strongest evidence in this episode is not a score but editable output. Still, a recording cannot verify hidden failure cases. These are the presenter's six tests, paired with one check a reader should run before adopting a similar output.
| Project | What the video shows | What to verify next |
|---|---|---|
| Crypto research deck | A chalkboard-style presentation about social sentiment and crypto pricing, with charts and cited findings. | Open every citation; recalculate figures and distinguish correlation from causation. Do not use it as investment advice. |
| Music video | Kimi finds and uses HyperFrames to assemble clips, captions and audio in an editable timeline. | Review rights to all online footage and music, then inspect the actual project and final export. |
| Blender face model | A face reconstructed from one photograph with a recognizable but imperfect likeness. | Check geometry from all angles, dimensions, consent to use the image and printability before ordering a physical object. |
| Chess game | Movable pieces, a second-player mode, an AI opponent, hints, threats and move history. | Play full games, including castling, promotion, checkmate and illegal moves; a polished board does not prove rule correctness. |
| Physics book | Ten chapters and responsive experiments for falling bodies, pendulums and other concepts. | Compare equations and outputs with trusted teaching material, test control extremes and accessibility. |
| Unity F1-style racer | A standalone playable first build with steering, other cars, speed, laps, tire display and track map. | Test collision, controls, lap timing, frame rate, packaging and any rights to names or visual assets. |
These tests span research, media, 3D, web interaction and a game engine. That breadth is impressive for a single episode, but the tasks were not scored against identical competing-model runs. The music video shows a particularly useful pattern: the agent chose a specialist tool and left behind a project the creator could edit. That is more valuable than a clip that can only be regenerated.
What Should You Try First?
Pick the smallest project that exposes both the model's strengths and its failure mode. If you make content, ask for a 20-second edit using media you own and require the editable timeline plus a render. If you build software, ask for one interactive screen with an explicit test list; do not start with a full racing game. If you work with research, give it a public source pack and require a citation next to every number. Measure usable artifact, corrections, time and total cost together. Stop when verification takes longer than doing the job yourself.
For a first Kimi session, avoid confidential data, give the agent only the files and tools the task needs, and inspect its actions. These are editorial testing precautions, not claims that the video tested security or privacy.
Watch Next: The Broader Agent Conversation
These three companion interviews from the creator give context to the product claims, rather than evidence that Kimi passed the six tasks.
Video Chapters
News: models, agents and robotics
| Time | Topic |
|---|---|
| 00:00 | The week and the Kimi comparison |
| 01:45 | Kimi K3.1 Pro announcement |
| 03:01 | Unitree H2-X robot demo |
| 03:59 | GPT-6.1 Terra and computer use |
| 04:59 | Anthropic IPO filing |
| 05:30 | Claude Duet preview |
News: creative, medical and security
| Time | Topic |
|---|---|
| 06:55 | CapCut and ChatGPT plan |
| 07:57 | Microsoft MAI-DxO study |
| 08:48 | OpenAI training-sandbox incident |
| 09:34 | Canva Craft |
| 10:33 | Gemini Personal Intelligence |
| 11:25 | OpenAI and Claude on Google Cloud |
| 12:11 | PHP fixes and Patch the Planet |
Kimi: setup and six projects
| Time | Topic |
|---|---|
| 12:52 | Hands-on Kimi tutorial begins |
| 14:06 | Benchmarks, price and access |
| 15:27 | Moonshot showcase examples |
| 16:04 | Crypto research deck |
| 16:52 | Music video with HyperFrames |
| 17:34 | Blender face model |
| 18:14 | Playable chess game |
| 18:58 | Interactive physics book |
| 19:43 | Unity racing game |
| 20:31 | Final verdict |
Sources and Limits
- Episode and demonstrations: Vaibhav Sisinty's full video and the user-supplied transcript. The six projects, their on-screen behavior and any creator commentary are attributed to that episode.
- Model and access: Moonshot's release and benchmark notes, API prices, and Kimi Code access guide. These are vendor sources, not independent measurements.
- Company announcements: OpenAI Terra, Anthropic's S-1 notice, Claude Duet, CapCut's planned plugin, Canva Craft, Unitree H2-X, and Google Cloud's OpenAI and Claude announcements.
- Research, safety and regional limits: Microsoft MAI-DxO, OpenAI's sandbox account, Patch the Planet response, PHP releases, Gemini app Personal Intelligence help, and Google's car rollout.
Access, prices, regional coverage and benchmark rankings may change. Recheck each primary source when you use it.