TypingMind is a good front-end. The problem it hands you is the one every bring-your-own-key tool hands you: before you can talk to a model, you have to go become that provider's customer. Sign up, verify, add a card, pre-load a balance, mint a key, paste it in. Then do it again for the next lab, and the next, and the video vendor, and the voice vendor.
Terminal X is the version where that step does not exist. One subscription, one balance, one place. The provider accounts sit on our side. You type a prompt, a parser decides which model or models the job needs, and several can run in parallel on the same request.
This page is written for the person who already knows what an API key is, already has four of them, and has started to resent the admin. It includes the actual credit costs and the actual routing rules, because that is the part you would otherwise have to reverse-engineer.
Start with the count. The model picker inside Terminal X lists 87 model variants across 23 vendor brands: Anthropic, OpenAI, Google, xAI, Perplexity, Meta, Qwen, Kimi, DeepSeek, Mistral, Black Forest Labs, Stability, Ideogram, Recraft, ElevenLabs, Suno, Runway, Pika, Luma, Seedance, Kling, HeyGen and Meshy. To reproduce that lineup with your own keys you are opening 23 accounts, agreeing to 23 sets of terms, tracking 23 balances and 23 rate limits, and keeping 23 secrets in one browser's local storage.
Most of those vendors are not a card on file with a monthly invoice. They want money up front before the first token, and the image, video and audio houses generally sell credit packs rather than metered billing. So you pre-commit to each one, separately, before you know which you will actually use. Money parked in a Stability balance is money that cannot pay for the Runway render you needed instead.
The time cost is worse than the money, and it is quieter. A new model ships on a Tuesday. With keys, trying it means finding the docs, checking whether your existing account covers it, possibly opening a new one, possibly waiting on verification. So most people do not try it. They keep using the model already wired up, which is the exact failure a multi-model front-end was supposed to prevent. Add the ongoing maintenance: rotating a key you pasted into a browser tab, catching a deprecation notice in an inbox you do not read, noticing that a provider quietly reorganised its model IDs.
The honest counterweight: at high volume on a single model, BYO keys win on unit price. You pay the provider's list rate with nobody in the middle. If your usage is one model, all day, every day, and you have already done the setup, keep your keys. The trade below only pays off if you use many models, or if you want to stop thinking about this.
Pooled provider access. You never create an account at OpenAI, Anthropic, Google, xAI or anyone else, there is nothing to rotate, and there is one card on file. Text, image, video, audio, 3D and document generation all sit behind the same balance, so trying a model you have never used costs you a prompt rather than a signup.
Every paid plan unlocks every model except two families, and the reason is arithmetic rather than upsell. Kling video is quoted at 245 credits per render and Runway Aleph at 491: one Aleph render costs more than a Basic plan's entire 400-credit month, and one Kling render eats most of it. Both are gated to Business by canAccessModel, which matches on model family, so every Kling variant is covered rather than just the base id. Everything else, including Veo, Runway Gen-4, HeyGen and Seedance, runs on any paid plan.
Alongside the models there are 32 live app connectors with real read and write access (Gmail send, Slack, GitHub, Stripe, Notion, Linear, Airtable, Google Calendar, Drive and more), plus 8 marked coming soon rather than pretended live. Under the hood Terminal X uses a mix of direct provider accounts and aggregation. What you are buying is that the balances, the rate limits and the 3 a.m. deprecation email are our problem.
The parser is a system prompt of roughly 25,000 tokens running on Claude Sonnet. It reads your request, splits it into tasks, picks a model per task and writes each one a tight brief. The rules are specific. A representative sample, straight from the prompt:
Write new code or build a page goes to claude-code, which runs Opus 5. Read or explain existing code goes to claude-sonnet, and so does a refactor, because explaining and restructuring are not the same job as authoring. A quick lookup or single current fact goes to Perplexity. A summary of many pages of input goes to gemini-pro for the context window. Strict-JSON output or a function-calling demo goes to gpt-5. Writing in a non-Latin script goes to qwen-next.
On the image side: a logo, wordmark or app icon defaults to recraft-v4.1-vector, because Recraft is the only route that returns genuinely editable vector output. A poster or key art with a baked-in title or credit block goes to gpt-image-2, which handles text in-image best. A cinematic still with no text goes to nano-banana-pro, as does product photography, packaging and interiors. Editing a photo you attached goes to nano-banana. Ask for the single most capable model regardless of cost and you get claude-fable at $10 per 1M in and $50 out, which the prompt explicitly forbids as a routine default.
That is the layer a BYO-key power user is running in their head every day, and it is the layer that goes stale fastest. It is also why a single prompt can produce a page, its hero image and a voiceover at once: independent tasks are dispatched in parallel, so wall time is the longest task, not the sum.
It is a judgment layer, and judgment layers are sometimes wrong. That is why the manual controls exist and are not buried.
The model picker sits in the composer. The default is Terminal X, which auto-routes. Pick any of the 87 variants instead and the executor skips the parser entirely and sends your prompt straight to that model as a single task. The choice persists in local storage, so if you want to talk to Opus 5 all week, you can.
Next to it is a routing mode: Default (best model for the task), Fastest (lowest latency), Cheapest (lowest cost), Best (highest quality, slower and premium), and Ensemble, which fans a text task out to Claude, GPT and Gemini in parallel and fuses the answers at roughly triple the cost. AI Teams, on Starter and up, goes one level further and scopes the parser to a lineup of models you curated, which is the closest thing to pinning your own routing table.
If you would rather not use a UI at all, there is a first-party API: keys prefixed txk_live_, credits attached to the key rather than the account, a planning call that returns the task list plus a cost estimate, and a streaming execute call. The usage log records models, credits and duration per call.
One credit is calibrated to about $0.006 of real API spend. Each model's quoted cost is derived from that: for text models, the real cost of a reference call of roughly 2,000 tokens in and 1,000 out, rounded up and carrying a deliberate margin buffer over the raw arithmetic so a provider price rise does not put a model underwater. Before you press enter you see a pre-flight estimate. That estimate is a deliberate upper bound: actual usage is metered from real tokens, characters or seconds afterwards and is usually lower, and you are never charged more than the quoted figure.
On the web app the routing step itself is free. You are billed for the tasks that run, not for the parse that decided them, and connector tasks that soft-error are excluded from the deduction. Through the API a plan call costs 5 credits, because there it is a billable request rather than part of a subscription.
Plans: Basic is $4.99 for 400 credits with 100 rolling over, Starter is $19.99 for 2,500 with 500 rolling over, Pro is $49.99 for 6,500 with 1,500 rolling over, and Business is quoted rather than listed. There is no free tier. At Pro that works out to about $0.0077 per credit; Basic is deliberately the most expensive per credit at $0.0125, because it exists as an entry point rather than a value plan.
No bring-your-own-key mode. You cannot paste an Anthropic or OpenAI key and pay that provider directly. If that is a hard requirement, because you hold committed spend, an enterprise agreement or a zero-retention arrangement with a specific lab, then TypingMind or another BYOK front-end is the right tool and this page will not argue otherwise.
No human seats. Accounts are individual. AI Teams means a saved lineup of models, not a shared workspace you invite colleagues into.
No per-client brand kits. There is no place to store a client's fonts, palette and tone of voice and have every generation inherit them. Memory helps across sessions, but if you run an agency with eight brands to keep straight, that constraint is real and you should test it before switching.
Google access is deliberately narrow. Gmail is send-only, and Drive is scoped to files Terminal X created, so it cannot read your inbox or your existing Drive. MCP servers are supported on the web app over streamable HTTP only, so local stdio servers cannot be plugged in.
Or skip the comparison shopping: Terminal X routes one prompt to the right model automatically — and runs several in parallel when a job needs more than one.
| Job | Where the parser sends it | Published rate (per 1M in / out) | Credits quoted |
|---|---|---|---|
| Quick lookup, single fact | Perplexity (executes on Sonar Pro) | $3.00 / $15.00 | 6 |
| Read or explain existing code | Claude Sonnet | $2.00 / $10.00 | 4 |
| Write new code, build a page | claude-code (runs Opus 5) | $5.00 / $25.00 | 60 |
| Summarise a long document | Gemini Pro | $2.00 / $12.00 | 3 |
| Deep analysis, long reasoning | Claude Opus 5 | $5.00 / $25.00 | 6 |
| Strict JSON, function calling | GPT-5 | $2.50 / $15.00 | 5 |
| Writing in a non-Latin script | Qwen3 Next 80B | $0.09 / $1.10 | 1 |
| Ultra-cheap volume work | GPT-5 Nano | $0.05 / $0.40 | 1 |
| Best available, cost no object | Claude Fable 5 | $10.00 / $50.00 | 15 |
| Vector logo or wordmark | Recraft V4.1 Vector | per render | 8 |
| Product or packaging shot | Nano Banana Pro | per render | 5 |
| Poster with baked-in title text | GPT Image 2 | per render | 16 |
| Voiceover | Eleven v2 (ElevenLabs) | per render | 17 |
| 5-second hero video | Pika 2.2 | per render | 36 |
| Kling video (Business plan only) | Kling v3.0 Standard | per render | 245 |
No. Terminal X runs on pooled provider access, so there is nothing to sign up for, nothing to pre-fund per vendor and nothing to rotate. The flip side is that there is no BYO-key mode either: you cannot supply your own key and pay a provider directly. If you already hold committed spend or a negotiated agreement with a lab, a BYOK front-end like TypingMind will serve you better.
You can pick. The composer has a model picker with 87 variants across 23 vendor brands, and selecting one skips the parser entirely and sends your prompt straight to that model as a single task. The choice persists across reloads. There is also a routing mode (Default, Fastest, Cheapest, Best, Ensemble) and AI Teams, which scopes the parser to a lineup you curate.
A pre-flight credit estimate appears before execution. It is a deliberate upper bound: after the run, cost is metered from actual tokens, characters or seconds, and you are never charged more than the quoted number. One credit is calibrated to roughly $0.006 of underlying API spend, and each model's quote is set from the real cost of a 2,000-in / 1,000-out reference call for text models, rounded up with a margin buffer on top.
Not at the margin, and it would be dishonest to claim otherwise. Direct API access at list price is cheaper per token than credits, which sell at about $0.0077 each on Pro and $0.0125 on Basic against a $0.006 cost basis. What the difference buys is no per-vendor minimums, no key management, no idle balances scattered across 23 accounts, and routing you do not maintain. Heavy single-model usage favours your own keys; broad multi-model and multi-modality usage favours this.
Four things worth knowing before you switch. There is no bring-your-own-key mode. There are no human seats or shared workspaces, so accounts are individual. There are no per-client brand kits, meaning fonts, palettes and tone of voice go in the prompt rather than a stored profile. And Google access is narrow by design: Gmail is send-only and Drive is scoped to files Terminal X created, so it cannot read your existing inbox or files.
Yes. There is a first-party API with keys prefixed txk_live_, where credits live on the key rather than the account balance. You call a plan endpoint that returns the task list and a credit estimate, then a streaming execute endpoint that emits per-task events and a final credits-used total. A plan call costs 5 credits. Every call is written to a usage log with models, credits and duration.
On the web app, yes, over the streamable-HTTP transport. The auth header you give a server is encrypted at rest. Local stdio servers, the kind you would launch on your own machine, are not supported on the web app.
Or skip the comparison shopping: Terminal X routes one prompt to the right model automatically — and runs several in parallel when a job needs more than one.