If you are shopping for a Poe alternative, the reason usually is not the model list. Poe's list is good. The reason is that you cannot look at a task before you run it and know what it will take out of your allowance. Compute points move with the bot you picked and with how long the exchange runs, and you find out the real number after the fact. That is a fine way to meter a service and a terrible way to plan a month of work.
Terminal X meters the way providers actually bill: tokens and calls, converted into one unit. One credit is $0.006 of real API spend. That is not a rounded marketing figure — it is the constant the cost table in our executor is built on, calibrated in May 2026 against each provider's published rates. Language-model work is charged at exactly ceil(real_dollar_cost / 0.006) on the tokens you actually used. Flat-rate work carries a fixed credit price set so that credits × $0.006 covers the provider's real cost with a deliberate margin buffer on top, because a provider raising its rates mid-month should not be your problem.
Below are our own numbers, the formula we bill with, three jobs priced end to end, and a section on where Poe is genuinely the better product. Nothing here is a benchmark claim. It is arithmetic you can redo.
The problem with a points system is not that it is expensive. It is that the conversion is invisible and it moves. Two messages to the same bot can cost different amounts. Switching bots changes the rate. Nothing on the screen tells you what a point is worth in dollars, so you cannot answer the only question that matters when you are budgeting: how many of these can I do this month?
A Terminal X credit has a fixed anchor. $0.006 of underlying API spend, one number, applied to every model in the system. The consequence is easy to check. Starter is 2,500 credits for $19.99, so the most API spend that plan can consume is 2,500 × $0.006 = $15.00. We are not hiding where the margin is.
What a credit costs you depends on which plan you bought it under. Basic is $4.99 for 400 credits, which works out to $0.0125 per credit. Starter is $19.99 for 2,500, or $0.0080. Pro is $49.99 for 6,500, or $0.0077. Business is scoped and quoted per deal, so it has no list rate. A credit buys identical work on all four. The plan only changes the price you paid for it, and unused credits roll over up to a per-plan cap (100 on Basic, 500 on Starter, 1,500 on Pro). Prepaid credit packs for API-key use are priced separately, from $0.0085 to about $0.010 a credit, because those credits never expire.
Two mechanisms, depending on what kind of task it is.
Flat-rate work has a fixed quote because its real cost barely varies, and each quote sits a little above that real cost so a provider price move does not put us underwater. An image edit on Nano Banana costs us about $0.005 and is quoted at 2 credits. A 5-second Pika clip costs $0.20 and is quoted at 36. A 5-second Kling clip costs $1.40 and is quoted at 245. A .pptx deck costs $0.048 and is quoted at 8. Those numbers are the same every time you run them.
Language-model work is metered on real tokens after the call returns, and there the arithmetic is exact. The formula is in the source: per-call cost = (input_tokens × input_rate + output_tokens × output_rate) / 1,000,000, then credits = ceil(cost / $0.006). The rates we bill against are the providers' published per-million-token prices. Claude Sonnet is $2.00 in / $10.00 out. Claude Code runs on Opus and bills at $5.00 / $25.00. GPT-5.6 Terra is $2.50 / $15.00. Gemini Pro is $2.00 / $12.00, Gemini Flash $1.50 / $7.50, Gemini Flash Lite $0.30 / $2.50. Grok Fast and DeepSeek are both $0.20 / $0.50. Mistral is $0.10 / $0.30. Claude Fable, the flagship, is $10.00 / $50.00, which is exactly why the router will not reach for it unless you explicitly ask for the best possible output regardless of cost.
Run that formula on a normal writing task. Two thousand input tokens and twelve hundred output tokens on Sonnet is (2,000 × $2 + 1,200 × $10) / 1e6 = $0.016, which is ceil(0.016 / 0.006) = 3 credits. The pre-flight quote for that task is 4. You get billed 3. Audio is metered on real clip duration and video on real seconds, for the same reason.
Job one: "write me a 900-word post on single-origin espresso." The parser rule for prose is explicit — PROSE / writing / script routes to claude-sonnet. One task, quoted at 4 credits, which covers up to $0.024 of API spend. Metering usually lands it at 3.
Job two: "build a landing page for my roastery, put a hero shot of the beans on it, and record a 30-second voiceover." That is three tasks and three different vendors. HTML and website builds route to claude-code (60 credits). Product and packaging photography routes to nano-banana-pro (5 credits). Standard voice routes to elevenlabs (17 credits). Total quote: 82 credits, covering $0.492 of API spend. On Pro that is about $0.63 out of your pocket. The image and the voiceover run in parallel; the HTML task is deliberately scheduled after the image so it can embed the result rather than describe it.
Job three: "make me a five-second video." With no premium signal in the prompt the router defaults to Pika at 36 credits, because it is the best dollars-per-second on a hero loop. Ask for cinematic and you get Runway at 105 for the same five seconds. Kling is 245. Runway Aleph — a clip-editing model, quoted on the ten-second length it clamps to — is 491. Those last two are gated to Business, and the reason is arithmetic rather than upsell: 491 credits is more than an entire Basic month's allotment of 400, and 245 is more than half of it.
None of those numbers required a support ticket to discover. They are in the table below.
Before anything runs, the estimator walks the task graph, counts the real input tokens each step will receive (including the output that upstream tasks will feed into downstream ones), and returns a point estimate plus a minimum and a maximum. You are told "this could cost up to N credits" before you commit, not after. The balance check that admits the run reserves a realistic figure — twice the point estimate, and never more than that modelled maximum — deliberately, so a Basic account is not refused a job it could comfortably afford.
The deduction is capped at the modelled maximum. If a run meters higher than that, the difference is absorbed by Terminal X and logged on our side. That is not theoretical. One website build metered 327 credits against a 76-credit midpoint quote and cost about $1.51 out of our own pocket, which is the reason claude-code is now quoted at 60 credits instead of 4. The deduction is also clamped to your remaining balance, so an account can never go negative. A live credit counter ticks during the run so you can watch it land.
Two honest caveats. First, cheap prompts are cheap but not free: the router itself is a model call, though greetings and other trivial prompts take a fast path that skips it entirely. Second, runs are durable. If you close the tab or your phone locks, the job keeps going and it bills. Only pressing Stop cancels a run and refunds it. Tasks that error out are excluded from billing entirely.
On Poe you pick a bot and talk to it. You can pull a second bot into a thread by name, but you are the one sequencing them and stitching the outputs together. Terminal X inverts that: you describe the outcome and a parser assigns each piece of the job to whichever model is best at that piece, then runs the independent pieces at the same time.
The routing table is opinionated and specific, not a vibe. Quick single-fact lookups go to Perplexity, which always executes on sonar-pro. A long-document summary goes to Gemini Pro because of the context window. Writing new code goes to claude-code and is explicitly not downgraded to Sonnet; reading and explaining existing code goes to Sonnet. A vector logo goes to Recraft because it is the only one of our image models that emits true SVG. A movie poster with a baked-in title block goes to GPT Image 2, while a cinematic still with no text goes to Nano Banana Pro. Ask for an image of a named real person and the parser schedules two tasks: sonar-pro to find real reference photos, then Nano Banana to generate against them, so the face is the right face.
Terminal X also connects to 30+ apps with real read and write access (Slack, GitHub, Notion, Linear, Stripe, Airtable, HubSpot, Supabase and more), plus user-added MCP servers. That is a different category of thing from a chatbot directory, and it is the main reason a prompt here can end with something changed in your account rather than text on a screen.
If you want Poe's behaviour back, there is a model picker in the composer. Pin a model and every prompt goes to it until you unpin it.
Terminal X has no bot marketplace. Poe's library of user-created bots and shared personas is a real product with real network effects, and we have nothing equivalent. If your workflow is built on curated public prompt-bots, switching will cost you something. AI Teams is the nearest feature and it is not the same thing: a Team is a named bundle of models you curate for a workflow, not a shareable persona.
There is no free plan. We removed it on 2026-07-30 after two days of signup-farm attacks, and a card on file was the only gate that held. The cheapest way in is Basic at $4.99.
Several connectors are narrower than the logo implies, and we would rather you read that here than discover it. Gmail is send-only; Terminal X cannot read your inbox. Google Drive is scoped to files Terminal X itself created, not your whole Drive. Google-side connectors still show Google's "unverified app" screen during consent while our verification review is open. MCP support is streamable-HTTP only, so local stdio servers cannot be plugged in.
There are no human seats. One account is one person, with no shared workspace and no per-client brand kits. And two video models, Kling and Runway Aleph, are Business-only for the credit reasons above.
If none of that is a dealbreaker, the tradeoff is straightforward: you give up a persona marketplace and get a published exchange rate, automatic routing across 35+ models and every modality, and a quote you are never billed above.
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 | Model the parser picks | Credits quoted | API spend the quote covers |
|---|---|---|---|
| Blog post, copy, general prose | claude-sonnet | 4 | $0.024 |
| Quick factual lookup | perplexity (runs sonar-pro) | 6 | $0.036 |
| Multi-step research report | research | 20 | $0.120 |
| Website or app build | claude-code (runs Opus) | 60 | $0.360 |
| Image edit / background swap | nano-banana | 2 | $0.012 |
| Product or packaging shot | nano-banana-pro | 5 | $0.030 |
| Vector logo (true SVG) | recraft (vector) | 8 | $0.048 |
| Slide deck (.pptx) | powerpoint | 8 | $0.048 |
| Spreadsheet (.xlsx) | excel | 8 | $0.048 |
| ~30s voiceover | elevenlabs | 17 | $0.102 |
| Song or jingle | suno | 20 | $0.120 |
| 5s video, default | pika | 36 | $0.216 |
| 5s video, premium | kling (Business only) | 245 | $1.470 |
| Transcription | whisper | 1 | $0.006 |
A credit is calibrated to $0.006 of real API spend. For language-model work that is exact — the charge is ceil(real dollar cost / 0.006) on the tokens actually used. Flat-rate work (images, video, audio, documents) carries a fixed credit price set so credits × $0.006 covers the provider's real cost with a margin buffer on top: a Pika clip really costs us $0.20 and is quoted at 36 credits, or $0.216. What you paid for that credit depends on your plan: $0.0125 on Basic ($4.99 for 400), $0.0080 on Starter ($19.99 for 2,500), $0.0077 on Pro ($49.99 for 6,500). Business is quoted per deal.
Not above the ceiling you were shown. You get a worst-case number before the run starts and the final deduction is capped at it; when a run meters higher, Terminal X absorbs the difference and logs it. The balance check that admits the run reserves a realistic figure — twice the point estimate, capped at that worst case — rather than the worst case itself, so a small plan is not refused a job it can afford. Because language-model tasks are metered on real token counts afterwards, you usually pay less than quoted rather than more, and the deduction is also clamped to your balance so an account can never go negative.
No, and that is the clearest thing Poe has that we do not. Terminal X has no bot marketplace and no shared persona directory. AI Teams lets you save a named bundle of models for a workflow, but it is a private routing constraint, not a publishable bot.
No. The free plan was removed on 2026-07-30 after signup-farm abuse, and there is no free trial. An account without a subscription sits at 0 credits and cannot run commands. The entry plan is Basic at $4.99 per month for 400 credits.
Yes. There is a model picker in the composer; pin a model and every prompt goes to it until you unpin it. The default is automatic routing, which is where the value is, but the manual path exists because sometimes you already know what you want.
Two: Kling and Runway Aleph, both Business-only. The reason is credit arithmetic rather than tiering for its own sake. Aleph is 491 credits per call against Basic's entire 400-credit monthly allotment, and Kling is 245. Every other model and every connector is available on every paid plan, including Basic — video generation included.
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.