ChatGPT is very good at the things OpenAI's models are good at. The friction starts when one piece of work needs three different specialists at once: a research pass over the live web, a code model that can build the page, and an image model that produces a hero shot you would actually ship. Inside a single-lab product you solve that by opening three tabs, or by paying three companies.
Terminal X is built the other way around. One prompt goes into a parser, the parser splits it into tasks, and each task goes to whichever model is strongest at that specific thing. Independent tasks run in parallel, so wall-clock time is the slowest task rather than the sum of all of them.
What follows is the actual data: our real per-model credit costs, the published token rates we bill against, routing rules lifted from our own parser, and a section on what Terminal X does not do. You should know the last part before you switch anything.
Most real work is not one task. 'Research the three competitors, write the launch page, and generate a hero image' is three jobs wearing a trench coat. A single-model product handles that by making one model attempt all three, or by making you handle the handoffs manually: copy the research into a new chat, copy the copy into an image prompt, copy the image back into the page.
The cost of that is not only the subscription stack. It is that the wrong specialist is doing the work. A general chat model writing a full HTML build is a different quality tier from a code-tuned model doing it. An image model asked for a packaging mock-up produces a different result from one tuned for photoreal product still-life. Terminal X's whole premise is that the routing decision is worth making explicitly, on every task, every time.
The parser caps a normal prompt at four tasks so it splits a mixed request without exploding into busywork. Website builds are the deliberate exception, because a page plus its images plus its copy is genuinely more than four pieces.
These are not marketing summaries. They are lines from the routing table the parser runs on, reproduced as written:
WRITE NEW CODE → claude-code. READ or EXPLAIN existing code → claude-sonnet. Refactor, improve readability, or restructure existing code → claude-sonnet. Quick lookup or single fact → perplexity. Long-document summary with many pages of input → gemini-pro. Strict-JSON output or a function-calling demo → gpt-5. Multilingual writing where a non-Latin script is primary → qwen-next.
On the image side: movie poster or key art with a title, tagline or credit block baked in → gpt-image-2. Product photography, food, still-life, packaging mock-ups → nano-banana-pro. Editable vector output explicitly requested → recraft, which is the only one producing true SVG. Logos and brand marks default to recraft-v4.1-vector. Editing an attached photo, swapping a background, or keeping a character consistent → nano-banana. Talking head, spokesperson, or lip-sync → heygen. Premium cinematic or long establishing shots → veo.
There is a hard limit written into the same table, and it applies to every image model on the market: image models paint letterforms, they do not typeset. One to eight words of display text is fine. Paragraph-length copy comes out as garbled pseudo-words on every single one. So text-heavy pieces like a brochure, a menu, or a resume get built as a pipeline with real typography, not as one image generation.
Take the prompt from earlier: research three competitors, write the launch page, generate the hero image. The parser splits it and routes it to the research agent (20 credits), claude-code (60 credits), and nano-banana-pro (5 credits). Total 85 credits, which is about $0.51 of underlying API spend at our calibration of 1 credit = $0.006. On the Starter plan at 2,500 credits, that is roughly 29 runs of that exact shape per month.
A lighter one: blog post plus cover image plus a voiceover. That routes to claude-sonnet (4), nano-banana (2), and elevenlabs (17). Twenty-three credits, about $0.14. The $4.99 Basic plan's 400 credits covers about seventeen of those.
Video is where the numbers change character. A five-second Pika loop is 36 credits. Seedance is 70, Luma 88, Veo 90, Runway 105, Kling 245. Those are not rounding errors on a 400-credit plan, which is exactly why Terminal X quotes the estimate before it runs anything rather than after.
One credit equals $0.006 of real API spend. Every per-model figure in the table below starts from the provider's published price converted at that rate, then carries a deliberate margin buffer on top: Nano Banana's real cost is about $0.005 an image and it quotes at 2 credits, GPT Image 2's is about $0.08 and it quotes at 16. The quote is a ceiling, not the charge — you are never billed more than you were quoted, and metering after the run reconciles the deduction down to what the call actually cost.
That is also why the spread is so wide. A single image edit and a full website build differ by roughly seventy times in real API spend — about $0.005 against about $0.36 — which is what puts Nano Banana at 2 credits and Claude Code at 60.
The token rates we bill against are published too. Claude Sonnet is $2.00 in / $10.00 out per 1M tokens. GPT-5 is $2.50 / $15.00. Gemini Pro is $2.00 / $12.00. Grok Fast and DeepSeek are both $0.20 / $0.50. GPT-5 Nano is $0.05 / $0.40. Claude Fable, the premium flagship, is $10.00 / $50.00, which is why the routing rules reserve it for prompts that explicitly ask for the single most capable model and never use it as a default.
Before a run, a token-aware estimator counts the real input tokens of each task, including outputs propagated from upstream tasks, and returns a point estimate plus a min and max range. After each task, metering reconciles the deduction against actual usage rather than leaving the flat table figure in place. Tasks that error are excluded from the deduction pass entirely, so a connector that fails does not get billed.
This is not an anti-OpenAI page. Several routing rules send work to OpenAI specifically because nothing else does the job as well.
GPT-5 (5 credits, $2.50 / $15.00) is the default for strict-JSON output and function-calling work, where its schema adherence is the reason to pick it. GPT-5 Codex is 4 credits at $1.25 / $10.00 for agentic coding. GPT-5 Pro is 12 credits at $5.00 / $30.00 when a problem genuinely needs the reasoning budget. GPT Image 2 at 16 credits sits first in our own routing table for posters and key art with baked-in title text, ahead of the Google models, because poster hero art lives or dies on legible typography and GPT Image 2 wins on layout and text fidelity while still rendering photoreal faces. Whisper handles transcription at 1 credit. GPT-OSS 120B and GPT-5 Nano are both 1 credit for high-volume work where the cheapest competent model wins.
The difference is not that OpenAI is worse. It is that you get GPT-5 when GPT-5 is the right answer, and Claude, Gemini, Grok, Qwen, DeepSeek, Mistral, Kimi, Recraft, Flux, ElevenLabs or Suno when one of those is, without you having to hold that decision tree in your head or pay four bills.
There is no free plan. It existed and was removed in July 2026 after two days of signup-farm abuse dropped the allotment to zero. A card on file is the only gate that worked. The cheapest way in is $4.99 a month.
Gmail is send-only. Terminal X requests the gmail.send scope, not full mailbox read, so a prompt like 'summarise my inbox' does not work on Gmail today. Outlook has broader email access if that is your provider — full list, search, read, send, reply, archive and delete. Google Drive is scoped to files Terminal X created, not your entire Drive.
Kling video and Runway Aleph are Business-only, and Business is scoped and quoted with you rather than listed with a price. If your work depends on Kling specifically, that is a conversation, not a checkout button.
MCP servers work over streamable HTTP only. Local stdio servers cannot be plugged in. And on the product side: the brand kit is one per account — a single palette, type pairing and logo the parser treats as binding — not a per-client set you switch between. There are no human seats, no managed-service staff, and no white-glove team behind the software. It is a routing product operated by one person, priced accordingly.
Basic is $4.99 for 400 credits with 100 rolling over. That is roughly 100 Sonnet drafts, or 66 live-web lookups, or about six full website builds. It unlocks every model and every connector except the two Business-gated video families.
Starter is $19.99 for 2,500 credits with 500 rollover, and adds AI Teams. Pro is $49.99 for 6,500 credits with 1,500 rollover, plus memory and personalisation. Business is sized to your volume and adds Kling, Runway Aleph, data residency controls and custom integrations.
Your monthly bill does not move based on which models the parser picks. The credit cost of each call is published, the estimate is shown before the run, and the deduction is reconciled against real usage afterwards. That is the whole pricing model.
| Job | Model the parser routes to | Credits | ≈ API spend |
|---|---|---|---|
| Edit an attached photo, swap a background | nano-banana | 2 | $0.012 |
| Blog post, marketing copy, general prose | claude-sonnet | 4 | $0.024 |
| Product, food or packaging shot | nano-banana-pro | 5 | $0.030 |
| Strict-JSON output, function calling | gpt-5 | 5 | $0.030 |
| Quick factual lookup with live sources | perplexity (runs Sonar Pro) | 6 | $0.036 |
| Deep reasoning, long analysis piece | claude-opus | 6 | $0.036 |
| Vector logo with true SVG output | recraft-v4.1-vector | 8 | $0.048 |
| Slide deck (.pptx) or spreadsheet (.xlsx) | powerpoint / excel | 8 | $0.048 |
| Poster or key art with baked-in title text | gpt-image-2 | 16 | $0.096 |
| Voiceover, roughly 30 seconds | elevenlabs | 17 | $0.102 |
| Multi-step research with a cited report | research (agent) | 20 | $0.120 |
| Five-second social video loop | pika | 36 | $0.216 |
| Full website or new codebase | claude-code | 60 | $0.360 |
| Premium cinematic video, long establishing shots | veo | 90 | $0.540 |
| Top-end video, named explicitly or as fallback | kling (Business only) | 245 | $1.470 |
Yes. GPT-5 is 5 credits per call and the parser routes strict-JSON and function-calling work to it by default. GPT-5 Codex (4 credits), GPT-5 Pro (12 credits), GPT Image 2 (16 credits), Whisper (1 credit) and GPT-5 Nano (1 credit) are all in the routing table too. You also get Claude, Gemini, Grok, Qwen, DeepSeek, Mistral and Kimi on the same subscription.
Research three competitors, write the launch page, generate the hero image routes to the research agent (20 credits), claude-code (60) and nano-banana-pro (5). That is 85 credits, about $0.51 of underlying API spend. On the $19.99 Starter plan's 2,500 credits, roughly 29 runs of that shape per month.
No. The free tier was removed in July 2026 after signup-farm abuse, and a card on file turned out to be the only gate that worked. Basic is $4.99 a month for 400 credits with 100 rolling over, and it unlocks every model and connector except Kling video and Runway Aleph, which are Business-tier.
Not yet. Terminal X requests the gmail.send scope, so it can send mail but cannot read your inbox. If you are on Outlook, email access is broader there. Google Drive is similarly scoped to files Terminal X created rather than your whole Drive. Be aware of that before you plan a workflow around inbox triage.
No. Terminal X runs on pooled provider keys, so there are no accounts to open at OpenAI, Anthropic, Google, xAI and the rest, and nothing to rotate. The tradeoff is that you pay in credits at our published rates rather than at each provider's wholesale rate.
The tasks that errored are excluded from the credit deduction, so you are not billed for a connector that did not connect or an image that got refused. Some runners also fall back automatically. Flux, for example, falls back to Grok Imagine on any failure, since Grok has the most permissive policy on real-person likenesses.
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.