TITAN / THEUS

Many intelligences. One decision.

You set the goal. TITAN picks the right AI for each task, within your plan.

Start with TITAN See how it works

Included in your Theus plan

Simulator

Watch it decide.

Pick a request, your plan and a priority, and see which model TITAN chooses, which stay as fallback and which are out. These are real decisions from TITAN's code with its factory values; in production it also learns from every answer and measures how long each model takes.

1 · The request
2 · Your plan
3 · Priority

In the API you can also ask for «Speed»: it prefers the models that answer sooner.

Chosen: Claude Sonnet 5. Fallback 1: Gemini 3.1 Pro. Fallback 2: Claude Opus 4.8

1 · Understands

«Refactor this function to make it clearer and write its tests:»

How hard it is
LightMediumFrontier

Level frontier: frontier get priority; models without a known level also compete. The rest stays as fallback.

2 · Needs
  • Nothing special: any model in the plan can
3 · Compares and picksPlan Theus Max

164 eligible candidates · 0 excluded by requirements or identity.

  1. 01Claude Sonnet 5Anthropic · Frontier · usage ×3Chosen
  2. 02Gemini 3.1 ProGoogle · Frontier · usage ×3.6Fallback 1
  3. 03Claude Opus 4.8Anthropic · Frontier · usage ×5Fallback 2
  4. 04GPT-5.5OpenAI · Frontier · usage ×9Fallback 3
  5. 05GPT-5.6 SolOpenAI · Frontier · usage ×9Fallback 4
  6. 06DeepSeek-V4 ProDeepSeek · Medium · usage ×1Fallback · other level
  7. 07GLM 5.3Z.ai · Medium · usage ×1Fallback · other level
  8. 08MiniMax M3MiniMax · Medium · usage ×1Fallback · other level

The winner is whoever best balances quality and usage for this request among those competing at its level; if it fails, the next one answers.

4 · Receipt
selected_model  claude-sonnet-5eligible_count  164excluded_count  0

What the answer carries in theus_titan_evidence (excerpt).

Why it matters

Choosing well in every message is work.TITAN does it for you.

A greeting does not need the most expensive model and a very long contract does not fit in just any of them. Getting it right in every message saves quota and gives you better answers, without you having to know every model.

The decision line

  1. 1

    Reads

    Your message and what it carries: text, image, document or tools.

  2. 2

    Measures

    How hard your request is and what it needs to be solved well.

  3. 3

    Filters

    Out goes whoever cannot: your plan, capabilities, context and provider health.

  4. 4

    Scores

    Quality against cost, with the weight your plan and your priority set.

  5. 5

    Answers

    The first in line. If it fails, the next, under the same conditions.

  6. 6

    Accounts for it

    The receipt: which model answered, how many attempts it took and how much it used.

Three requests, three AIs · Theus Max plan
  1. 01 · A greeting«Hi, how are you?»
    → TITAN picksDeepSeek-V4 ProLight task: frontier models stay as fallback
  2. 02 · Refactor code«Refactor this function to make it clearer and write its tests:»
    → TITAN picksClaude Sonnet 5Hard task: frontier models compete
  3. 03 · A very long contract«Analyze this contract and compare its termination clauses.»
    → TITAN picksGemini 3.1 ProOnly the ones that read 260,000 tokens compete

Computed with TITAN's code, without usage history

  • 0 TTChoosing does not spend your quota: it is a formula that runs in milliseconds, not another AI thinking
  • 3 levelsA greeting does not spend a frontier model; a hard task does not drop below frontier
  • 3 attemptsAt most per stage if a route fails, always with a model that meets the same bar
  • 1 receiptIn every answer: which model answered and how much it consumed

What weighs in the decision

Quality, usageand what it learns.

TITAN does not ask another AI: it decides on the server, in milliseconds and without spending your quota, with criteria Theus defines and tunes. This is what it looks at in every message.

  1. The quality the task needs

    Every model has a Theus score for what it does well. The harder the request, the more quality weighs and the less price does.

  2. What it uses from your plan

    Between two that perform alike, the one that spends less of your quota wins. On entry plans usage weighs more, so your week goes further.

  3. How long it really takes

    TITAN measures how long each model takes to answer and does not keep you waiting when another, just as good, is faster.

  4. What it learns

    Whether an answer finished well, arrived fast, or you regenerated it, edited it or it was cut off: all of that counts next time. It never judges the content.

  5. Your priority

    By default it looks for balance. In the API you can ask for more quality, lower usage or more speed, and set your own criteria.

The exact calculation belongs to Theus. What you always see is the result: which model answered, how many competed and how much it used.

Before scoring

What is not negotiable.

A preference never makes up for an exclusion. Before adding a single point, TITAN leaves out whoever cannot do the job.

  1. Your plan

    TITAN considers the full catalogue enabled for your plan and filters it for this request. Counts vary with capabilities, context, routes and your constraints; manual-only routes do not enter automatic routing and aliases are not additional models.

  2. What the message asks for

    Does it carry an image? Only models that see images. Is it an agent with tools? Only models that use them. Do you ask for JSON? Only models that declare support for it.

  3. That it fits

    The request and the answer must fit in the model's context. If it fits in none, TITAN tells you instead of cutting it silently.

  4. That it is healthy

    A provider that is failing does not compete until it recovers.

  5. The task's level

    Difficulty helps order eligible models. Levels guide priority without closing the catalogue to a fixed list; a model with no known level can still compete.

  6. Your usage window

    On paid plans, a model that would eat too much of your usage window with this request goes to the back of the line.

And if the chosen route fails?

TITAN tries the next one and your turn carries on.

  1. Up to 3 attempts per stage

    First another route of the same model; then the next in line. The order is not random: it is the same ranking as the decision.

  2. The fallback meets the same bar

    It comes from the same line of eligible models: if your message carries an image, the fallback sees it too; if the request is long, it fits too.

  3. An empty attempt is not charged

    If a model returns an empty or invalid answer, TITAN discards it and that attempt costs you 0 TT.

  4. What you already saw is not replaced

    If the text already started showing on your screen, that model finishes: TITAN does not swap it halfway.

  5. You pay for what answered

    Each attempt is charged at the weight of the model that served it, not the one requested.

When it is needed

Two AIs: one writes, the other reviews.

In Theus Max and Theus Max 10x, TITAN can ask a second model —from another engine— to review the answer, and the author rewrites it with that review. It does not happen on every message: only when you ask for it or when there is evidence that something went wrong.

  • You ask for it

    With «Contrast between AIs» in TITAN's criteria, the review happens whenever two engines are eligible.

  • Something was not met

    An automatic check detects a delivery failure —a page or a presentation that was cut off, or that came back unchanged when you asked for changes— or one of your criteria that was not met.

  • And if it does not fit

    If the answer is very long or there is no time left to review, the draft is delivered and the receipt says so.

How do I know which model answered?

Every answer carries its receipt.

  • In the chat

    Above every answer, «Evidence for this answer»: each attempt with the model that served it, its route and its status; the checks, the recorded usage and the uncertainty of the choice, which is not a probability of being right.

  • In the API

    The answer carries theus_titan_evidence: the chosen model, how many competed and how many were left out, each attempt and its usage.

  • In your usage, not yet

    On the usage page the turn shows as TITAN, the model you requested. The breakdown by model lives in each answer's receipt.

theus_titan_evidence · decision
selected_model   claude-sonnet-5eligible_count   164excluded_count   0

Example computed with TITAN's code · Theus Max plan · code request

See the selector's measurements, their limits and the files of every evaluation in our public TITAN benchmarks.

And if you want to be in charge

TITAN decides. You set the rules.

  1. TITAN's criteria, in the chat

    Strategy (Adaptive, One AI or Contrast between AIs), priority (Balance, Quality, Lower usage or Speed), a TT limit per call and up to 12 delivery criteria.

  2. A specific model

    Pick it in the selector and the turn goes straight to that model, without going through the choice.

  3. In the API

    Ask for the titan model and, if you want, add theus_routing with the priority, the strategy, allowed, preferred or blocked models, a weight or TT limit and your criteria.

Questions

What people usually ask.

  1. What is TITAN?

    Theus's automatic mode: you give it the task and it picks, among the models your plan includes, the one that solves it with the best balance of quality and cost. It is not bought separately or installed: in the chat it answers when you do not pick a model, and in the API it is the titan model.

  2. Among how many models does it choose?

    Across the full catalogue enabled for your plan that can handle the request. Capability, context, availability and budget constraints apply, and aliases are grouped by engine. Routes reserved for manual selection do not enter automatic routing. It compares the eligible catalogue before prioritising. Consult the catalogue for models and conditions.

  3. Does deciding cost anything?

    No. The choice is a formula that runs on the server in milliseconds; it does not ask another model. What uses your quota is the answer and, when it kicks in, the review.

  4. How do I know which model answered?

    In the chat, above every answer, «Evidence for this answer» shows each attempt with the model that served it. In the API, the answer carries theus_titan_evidence with the chosen model and each attempt. On the usage page the turn still shows as TITAN.

  5. What happens if the chosen model fails?

    TITAN tries another route of the same model and then the next in line, up to 3 attempts per stage, always with a model that meets the same constraints. An empty or invalid attempt is not charged, and each attempt is charged at the weight of the model that served it.

  6. Does it use several models at once?

    In most messages a single model answers. In Theus Max and Theus Max 10x, if you ask for it with «Contrast between AIs» or if a check detects a delivery failure, a second model from another engine reviews and the author rewrites.

  7. Does it learn?

    Yes, from service signals: whether the answer finished, whether it arrived fast, whether you regenerated or edited it, whether the model had to be replaced or whether it was cut off. It never judges the content.

  8. Can I pick the model myself?

    Yes. Pick a specific model in the selector and the turn goes straight to it. Or let TITAN decide and adjust its criteria: strategy, priority, TT limit and delivery criteria.

  9. Does TITAN use my local model or my own key?

    TITAN chooses among the managed routes included in your plan. The public CLI requires a Theus account. Desktop separately offers Crea tu LLM on supported plans and hardware; those local sessions are outside TITAN’s managed catalogue.

  10. How much does TITAN cost?

    Nothing extra. What you pay is the plan, and the plan decides which models TITAN chooses from.

Don't pick another AI.Let TITAN decide.

Included in every plan · Decides in milliseconds

Start with Theus →