Always Use the Best Model — My API Model Roster (Which Third-Party Models I've Wired into Claude Code)

A lot of people have been asking me lately: which models are actually wired into your Claude Code, and how does it manage all these different things?

So this episode I’ll show you the APIs I have connected. I had Claude Code do a pass for me: take every third-party model I currently access via API and turn it into a web page, in the light PPVI style. The page is mainly a list, spelling out each model’s name, its API signup URL, a summary of what it does and what it’s worth, and what it costs. Built that way so it’s easy to walk you through in detail.

I’ll upload this page to the blog and you can download it. How to use it: click the URL to the right of each model, go sign up for the API on the official site, top up, get an API key, hand it to Claude Code, and you can do all of this too.

My rule: always use the best model plus the best tools

There’s been a lot of talk lately about whether Claude Code or Codex is better. My thinking has always been the same — always use the best models and the best tools.

On model choice, I always use the best available today. Subjectively, that’s Claude’s Opus 4.8. And for multimodal scenarios — parsing video or audio, say — I think the best is Gemini, given it’s natively multimodal.

Same for tool choice. For computer use across multiple machines, Codex is the more convenient route; for gpt-image-2 generation, Codex is also more direct and better value. So I keep a low-tier plan alongside as a companion: treat Claude as the brain and Codex as a tool, calling Codex from Claude Code through the terminal in the CLI to get the work done.

Text LLMs: reasoning and conversation

Claude. Claude is still what I use most. My own Claude Code isn’t on an API — it runs on the official subscription plan; but the outward-facing products — script writing for the comic drama platform, drafting WeChat articles and voiceover scripts — all run on Claude. Besides 4.8, Sonnet 4.6 and Opus 4.6 get plenty of use too, because the prose, the long context, the reading of scripts and novels, the attention and reasoning, and the final quality of the writing are all excellent.

OpenRouter. It’s a third-party relay. I’ve talked about relays before — I’m not saying all relays are bad, I’m saying relay APIs are bad for Claude Code, because a lot of tool use breaks through them. OpenRouter is an overseas aggregating relay with a relatively good reputation, but it takes roughly a 5.5% cut. So when a new model appears and I haven’t signed up for its official API yet, I test it on OpenRouter first to gauge what it can do.

Gemini Flash. Fairly cheap, and 2.5 and 3.x Flash are what I’ve used recently. Because it’s natively multimodal, my most common use is having it generate the full transcript for subtitles after I finish recording each day — it converts video or audio into a text draft with almost no errors, and it’s a conversion with reasoning behind it.

Qwen3. I signed up for it through the aggregator SiliconFlow, which came with a certain amount of free credit. I’ve been using it for request routing and intent routing: when my AI support agent or bot receives a user request, it makes a quick first judgment about how I should respond.

Image generation: text-to-image and editing

Nano Banana Pro. The one I use most. Native 4K, and the consistency and quality of the images are very strong. Two shortcomings: one, it’s a bit expensive; two, when there’s a lot of Chinese text in a frame, the Chinese comes out garbled and unreadable. But the output quality and the prompt adherence are both excellent.

Nano Banana 2 (3.1 Flash). Cheaper than Pro, faster to generate, allows more reference images, and renders Chinese better; the fine detail quality is just slightly below Nano Banana Pro — hard to quantify, but the difference is real.

gpt-image-2. Lately I’ve used it a lot for PPTs. Its signature strength is the sheer number of reference images — up to 16 — letting you generate complex multi-person scenes or complex Chinese infographics; accuracy is extremely high, and I use it to fix details on e-commerce images, including logos and text. The downsides are just as clear: the style it generates has a strong machine-made feel, the work is a touch “greasy,” and there’s a rendered stiffness to it. It also charges by token, so small-resolution images are cheap, but 4K images actually cost more than Nano Banana.

Seedream 4.0. For images where the quality bar is lower but the cost bar is tight — connected domestically through Volcano Ark, which also handed out some free credit early on.

Kling. It’s odd: to use Kling’s official API you have to fill out an extremely complex form and go through the whole B2B sales flow of requesting a trial and topping up, which is a huge hassle. So the way I actually use Kling now is through Alibaba Cloud Bailian, rather than Kling’s own interface.

Voice TTS: synthesis and cloning

MiniMax TTS. Covered last episode; I use their speech-2.8-hd. Voice cloning, voice formula blending, cross-lingual capability, emotional delivery, plus the simulation of breath and sibilance are all very strong. It’s sat at 2.8 without an update for quite a while, but it’s more than good enough as it is.

Qwen3-TTS. That product you’ve seen me build myself actually uses Qwen3 TTS to do a polish — repair and refinement — and the value for money is excellent.

Digital avatars / video generation

Shiliu digital avatar. This is an interface built by Vector Equation. We’ve collaborated heavily over a long stretch, and digital avatars have been a consistent line of business for me. My current view is that among Chinese digital avatar models, it’s the one that nails lip sync at 4K precision.

Jimeng Seedance. Seedance is mandatory for comic drama. On its current interface, 720p costs roughly one yuan per second, but it takes few re-rolls and its shot work and multi-shot work are both very good. Apart from some special cases — video editing, say, where I’m talking right now and want the background swapped or an extra object in my hand, which is where Google’s newest Omni model comes in — for all of this AI drama, AI comic drama work, Seedance is absolutely, unquestionably in first place. Reportedly its revenue is already past the 1-billion-yuan mark, another point of pride for Chinese tech.

Separately, for generating 1080p video, besides Seedance, Kling gets the most use, then probably Google’s Veo 3.1 — though I’ve noticed I use it less and less lately.

Wrapping up

My current monthly token spend, on top of the official subscription, sits somewhere in the low thousands of US dollars — that’s the generation volume of a handful of small products.

The models I use are each the best-performing or best-value option in their domain. Wire them into Claude Code via API and you get background removal, PPT generation, video generation, video subtitle extraction, and so on — every one of the capabilities I’ve demonstrated is Claude Code calling out to these models.

Any questions, drop me a comment. See you next episode.

Source: EP0029_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the roughly 11-minute original recording

[00:00] A lot of people have been asking me lately which models I’ve actually wired into my Claude Code — the ones that let me do all these different things. So today I’m going to walk you through the APIs I’ve hooked up here. Pull together a rundown for me: every third-party model currently connected to my Claude Code via API. I need you to build a web page, using the PPVI light style. And the content of the page is basically a list, and the list should spell out

[00:26] every API model I use, the name of the model, the URL where you apply for its API, a short intro to what it does and what it’s worth, and what it charges. Build that page out so we can do a proper walkthrough. All right, let’s wait for it to produce the content. Today is really just a light topic like that. So a lot of people have been arguing lately about

[00:53] whether Claude Code is better or Codex is better. My thinking on this is, is this — it’s always been this: we should always use the best model and the best tool. Which is to say, on model choice, I’m definitely going to use whatever is the best there is today. In my opinion, subjectively — that’s Claude Code, the Opus 4.8 model. But for multimodal work, say parsing video or audio, I think the best model is Gemini’s model.

[01:18] It’s natively multimodal, after all. So on model choice, I’ll always use the best model for the job at hand. On tool choice — take computer use across multiple machines, Codex is probably the more convenient route. And for gpt-image-2 image generation, Codex is a more direct, better-value route. So I’d go open a low-tier plan and use it alongside. Treat Claude — treat Claude as the brain,

[01:43] and treat Codex as a tool. Inside the CLI, from the terminal, I use Claude Code to call Codex and get some of the work I want done. And right now it’s — oh, it’s asking me something. All the models go in. Besides the ones you mentioned, do a full analysis of every model I’ve ever saved an API key for in that config file. So people get a sense of which ones I’m using now,

[02:09] on the models’ official sites. Right, because I’ve got quite a lot of models hooked up here. One business line that’s kept running for a long time is the digital-avatar business, so Shiliu Digital Avatar is a really good partner of ours. So we — oops, forgot to hit next. OK, good. Let’s speed this part up and just show you a page in a minute. There we go, it’s done.

[02:34] Let’s take a look. I’m going to upload this page to the blog so you can download it. How do you use the page? Well, I’ll walk you through it end to end first, and then — see, on the right here there’s a — you click that URL and it takes you straight to the official site to apply for the API, right? Top up, get that API key, hand it to Claude Code, and you can do this stuff. So first, let’s look at how many categories it’s split into. The first category is text —

[02:59] large language models, reasoning and conversation. The one I use most is still Claude. For Claude Code itself I haven’t wired in the API — I’m on the official subscription plan. But for the products we ship, like the comic-drama platform writing scripts, or writing articles, writing the WeChat official-account posts and the voiceover scripts, it’s all Claude. And beyond 4.8, Sonnet 4.6 and Opus 4.6 are actually

[03:24] used fairly often too, because their way with words, plus the long context — for reading a screenplay, or reading a novel, their attention and their reasoning ability, and how the text finally comes out, are all pretty great. Then the second one is OpenRouter. It’s a third-party relay. I’ve talked about relays before. I’m not saying all relays are bad — what’s bad is plugging a relay’s API into Claude Code.

[03:49] There’s a lot of tool use inside, and using those tools tends to break. OpenRouter is an overseas aggregator relay, and its reputation is relatively good. But one problem with it is that it takes a 5.5% — I remember it used to be 8%, or was it 5.5% — a handling, handling fee. So for some models, newly released models, when I haven’t yet applied for their official

[04:14] API, I’ll use OpenRouter to run a test first and see what the model can do. Next down is the Gemini Flash model. That one is relatively cheap. 2.5 and 3, 3.5 Flash are the ones I’ve been using a lot lately. Mainly for what? Because it’s a natively multimodal model, and what I use it for most is: every day after I finish recording a video, I have it generate the subtitle transcript for me. The subtitle transcript — it can turn video or audio into

[04:41] text that’s almost error-free. It’s a conversion with reasoning behind it. Then the Qwen3 series — I applied for that through an aggregator site called SiliconFlow. And at the time they gave, I remember at the time they gave, a certain amount of free credit. So right now it’s been handling what’s called request routing, intent-routing calls: when my AI customer-service agent, or my AI bot, gets a request from a user,

[05:06] it makes a quick call about the way I should respond. All that request-intent classification is done with Qwen3. Next down, this image-generation and text-to-image part. The one I use most, most often, is Nano Banana Pro. It’s natively 4K, and the consistency and quality of the images it makes are very strong. The one shortcoming, the fly in the ointment, is: one, it’s a little pricey,

[05:31] and then the other is its Chinese. If a frame has a lot of Chinese text on it, the Chinese actually comes out garbled — and unreadable. But its output quality, including prompt adherence, is very strong. Next down is 3.1 Flash, the Nano Banana 2 model. That model is cheaper than Pro, and it generates faster, it allows more reference images, and it renders Chinese better. But

[05:57] on the fine-detail quality of the images, it’s a bit behind Nano Banana Pro. Right now that’s hard to quantify or make concrete, but the difference is real. Next down — the one I’ve used a lot lately for slide decks is OpenAI’s gpt-image-2. The most distinctive thing about that model is the huge number of reference images, right? It can take up to 16 reference images. You can generate a complex multi-person scene, or a complex Chinese infographic —

[06:22] you can do all of it with that one. Its defining trait is extremely high accuracy. When I’m making e-commerce images, I can use it to fix details — logos, the text parts. But its downside is pretty obvious too: the style it generates has a strong machine-made feel. How do I put it? It’s a bit like — what it produces is a little slick, or a little like something machine-rendered.

[06:48] You can just about feel a stiffness to it. With gpt-image-2, its low-resolution images are fairly cheap. The high-resolution ones — since it bills by token — its 4K images actually end up costing more to generate than Nano Banana. Then the Seedream 4.0 model. I use it to generate images where the quality bar isn’t that high

[07:13] but the cost bar is. That goes through Volcano Ark for a domestic direct connection. And this one’s also a relay outfit. Early on they also handed out a certain amount of free credit, including offering a vectorization model, so I ran a few simple experiments, and after that I never used the model again. Same with Kling. Kling is really odd: if you want to use Kling’s official API,

[07:38] you go to Kling to apply and they make you fill out an incredibly complicated form, and then it’s that whole B2B sales motion — apply for a trial, then top up. It’s such a hassle. So the way I use Kling now is actually through the Bailian platform, Alibaba Cloud’s Bailian platform, to reach Kling — rather than going through Kling’s official API. Then voice, TTS synthesis and cloning. I covered this in a recent episode: I use MiniMax’s TTS, and their speech-2.8-hd model is just something else.

[08:04] Voice cloning, plus its voice formula, plus its cross-language ability, and its emotional delivery, and the way it simulates breathing and those labiodental sounds — all extremely strong. But it’s been sitting on version 2.8 for quite a while now, no updates. Still, it’s in a completely good-enough state. Then the Qwen3 TTS model — that’s what you’ve seen, the cloud-calligraphy thing I built myself actually uses Qwen3 TTS

[08:29] to do a polish, it does the cleanup and the polishing. Great, great value for money. Then, moving on, digital-avatar video generation. Like I mentioned, that’s the Shiliu digital-avatar API built by Xiangliang Fangcheng. And right now, since we work together a lot, the model we end up using — I’d say I currently think it’s the one, among domestic digital-avatar models, that at 4K precision

[08:56] nails the lip sync — one, one of those models. Then Jimeng video, Seedance, goes without saying. Seedance 2.0 is a must for comic drama. Its current API: at 720p, at that kind of, kind of reso— resolution, the cost is about ¥1 per second. But because it takes fewer rerolls, its camera work, its multi-shot work, is really — right now it goes without saying that,

[09:22] apart from a few special cases — say when you’re doing this kind of video editing, like I’m talking right now and you swap out my background, or put something extra in my hand, for that I’d use Google’s newest Omni model — for every one of these AI dramas or AI comic dramas, right now it really is, if Seedance, this Seedance 2.0, is flat-out, flat-out in the number one spot today. Word is its revenue has already crossed the 1 billion mark, which is another

[09:49] point of pride for domestic tech. So those are basically all the models we use. One thing I haven’t mentioned in here is Gemini — Google’s models, actually — I also use them to generate video. But as things stand, say I want to generate 1080p video, apart from Seedance 2.0 the one I use most is Kling. And after that, after that, after that, it’s probably Veo 3.1, that Google model. So I do still use some of it,

[10:14] but lately I’ve noticed I seem to use it less and less, less and less. My current monthly token spend, on top of the official subscriptions, is somewhere in the few-thousand-dollars range — that’s the scale of generation for a handful of small products. Anyway, the models I’m using right now are these. These are the models — each of them is the best-performing or the best-value option in its own domain.

[10:40] Wire them into Claude Code via API and we can pull off things like cutting backgrounds out of images, then generating slide decks, generating video, then pulling subtitles out of video, and so on — that whole run of features I’ve demoed. All of it is Claude Code calling out to some model’s capability. If you’ve got questions, same as always, drop me a comment. See you next episode, bye.