TerraFlow

Governed AI control plane

Govern every AI request. Pay only for the model it needs.

TerraFlow runs inside your network or cloud account. It checks every AI request against your signed rules, then sends it to the cheapest model that is good enough for that kind of question, with budgets, a tamper-proof audit record and a person to approve risky answers.

Log in See the proof

Apps connect through OpenAI- and Anthropic-compatible APIs: change a base URL and a key. Read the API guide.

What happens to one request

  1. IdentifyWhich app, which use case, which signed rules apply
  2. ClassifyCard, health and personal data found in the prompt and fetched documents
  3. Rules decide firstOnly models cleared for this data stay on the list
  4. RouteThe cheapest allowed model good enough for this type of question
  5. BudgetCost reserved before the call; stops where you set the limit
  6. Check and recordAnswer checked, held for a person if required; audit record written, with no prompt text

The problem

AI spend defaults to the premium model, and nobody can show it was safe

Everything goes to premium

Teams wire one top-end model into every app, whether the question is trivial or critical. Routine questions pay premium prices.

Sensitive data drifts outside

Card numbers, personal and health data reach public AI services, in prompts and in the documents fetched to answer them, with nothing to stop them.

Savings are claimed, not proven

Headline percentages are quoted, but nobody checks that answer quality held, that the rules were obeyed, or that the audit trail adds up.

How it works

Rules first, then cost. Every time.

Each AI use case has a signed contract: its rules. The rules pick the allowed models before any cost logic runs, so saving money can never break them.

  1. Check the rules

    The request is matched to its use case. Signed rules say which models may see which data. Card, health and personal data are detected in prompts and in fetched documents; models not cleared for them are dropped.

  2. Route to the right model

    Among the allowed models, TerraFlow picks the cheapest one that is good enough for this type of question, and reuses checked answers for repeats.

  3. Hold to a budget

    Cost is reserved before each call, against budgets per department and use case. A spent budget stops the request with a clear reason code.

  4. Check, approve, record

    The answer is checked on the way back. Where the rules require it, a person approves it first. Every decision is written to a tamper-proof audit record that never stores prompt text.

Your data stays inside

Installed in your network or cloud account, not ours

TerraFlow runs on your servers or in your cloud account, and can run fully offline. Data leaves only when a signed rule allows that specific model to see it.

Never leaves

  • Prompts, answers and documents, processed inside your boundary
  • Answers held for review, stored encrypted
  • The audit and cost records, in your own PostgreSQL
  • Your provider keys: TerraFlow stores only references to them

Leaves only if the rules allow

  • A request to a named cloud or public model
  • Only data that model is cleared to see
  • Each call recorded with the rule that allowed it

Never in logs or monitoring

  • Prompt, answer or document text
  • Credentials
  • The audit record holds codes, ids, costs and timings only

The proof

GPT-5-level quality at up to half the cost

Every figure below says where it comes from. Your own number comes from the same test on your own tasks and models.

Public benchmark
24%

lower cost than GPT-5 at the same quality

Balanced setting: 65.6% correct vs GPT-5's 65.9%, for $96 vs $126. LLMRouterBench, Nov 2026.

Public benchmark
50%

lower cost, keeping 96% of GPT-5's quality

Max saving setting: 63.3% correct for $63. LLMRouterBench, Nov 2026.

Live run
28%

lower bill than sending everything to Claude Opus

159 enterprise tasks on Claude and Azure OpenAI models, Oct 2026.

Against single models

Accuracy versus cost for 13 AI models and TerraFlow Grey dots are 13 models used on their own. GPT-5 is the best single model at 65.9% for $126. TerraFlow's line runs from Max saving (63.3%, $63) through Balanced (65.6%, $96) to Max quality (66.8%, $155). 42% 49% 56% 63% 70% $0 $50 $100 $150 $200 $250 Cost to answer the test questions (USD), cheaper to the left Quality: % answered correctly Gemini 2.5 Pro DeepSeek R1 Qwen3 235B Claude Sonnet 4 ● each grey dot is one AI model used on its own GPT-5 · best single model Max quality Balanced Saver Max saving TerraFlow · one dial
Public benchmark: LLMRouterBench (2025), 12,446 questions, 13 models. TerraFlow learns on 70% of the questions and is scored on the 30% it has never seen, over five splits. Run November 2026.

One TerraFlow control trades quality for cost. Each setting is a point on the green line.

Cost = total to answer the test questions.
SettingCorrectCostvs GPT-5
GPT-5 alone65.9%$126—
TerraFlow Balanced65.6%$96−24%
TerraFlow Max saving63.3%$63−50%
TerraFlow Max quality66.8%$155+23%

Max saving keeps 96% of GPT-5's quality. Max quality beats GPT-5.

Against other optimisers

TerraFlow compared with published research routers on the same benchmark Research routers: RouteLLM 67.7% for $111, Avengers 68.6% for $133, GraphRouter 67.8% for $150. GPT-5 alone 65.9% for $126. TerraFlow's line runs from Max saving (63.3%, $63) through Balanced (65.6%, $96) to Max quality (66.8%, $155), in the same range, while also enforcing data rules, auditing every decision and supporting human approval. 62% 64% 66% 68% 70% $50 $75 $100 $125 $150 $175 Cost to answer the test questions (USD), cheaper to the left Quality: % answered correctly RouteLLM Avengers GraphRouter GPT-5 alone Max quality Balanced Saver Max saving ● TerraFlow governed: rules, audit, approvals ■ Research routers cost only
Public benchmark: LLMRouterBench (2025), same questions and models. Router figures as published for that benchmark.

Published research routers land in the same range. They optimise cost only; TerraFlow is the only one that also enforces data rules, audits every decision and supports human approval.

OptimiserCorrectCost
RouteLLM67.7%$111
Avengers68.6%$133
GraphRouter67.8%$150
TerraFlow, Balanced to Max quality65.6–66.8%$96–$155

No saving counts until six checks pass, on your own tasks

TerraFlow reports a saving only after all six pass on the customer's own tasks and models. A feature that fails is rejected, with the check it failed.

  • Quality holds
  • Critical work stays safe
  • Rules were obeyed
  • The audit trail verifies
  • The saving beats its own cost
  • Speed holds

For each team

One control plane, four owners

Security CISO · DPO

  • Signed rules per AI use case decide which models see which data
  • Card, health and personal data caught in prompts and fetched documents
  • Rules are checked before any model is chosen

Finance FinOps

  • Cost reserved before each call
  • Budgets per department and use case
  • Savings reported only after the six checks pass

Platform AI platform

  • Apps connect by changing a base URL and a key
  • OpenAI- and Anthropic-compatible APIs
  • Your servers, your cloud, public models only where allowed

Audit Risk · audit

  • Tamper-proof record of every decision, with no prompt text
  • Human approval where the rules require it
  • Verify the record yourself with one command

How to start

From install to your own number

  1. 01 · Install

    Inside your boundary

    Containers on a Linux host in your network or cloud account, online or fully offline from a signed bundle. Install guide

  2. 02 · Connect

    Point your apps

    Keep the OpenAI or Anthropic SDK you use. Change the base URL and the key. API guide

  3. 03 · Set rules

    Each team signs off

    Write a contract per use case: allowed data, models, budgets and review. A second person approves it before it runs.

  4. 04 · Prove savings

    On your own tasks

    Run the six checks on your own tasks and models. You see your number before anything goes live.

Already have a workspace?

Sign in to the console. If you were sent an invite link, open it to choose your username and password.

Log in Connect your first app