Cortex
CortexFoundry · Smart routing

Stop paying frontier prices for routine work.

Cortex routes every AI request to the cheapest model that passes your evals, enriched with your org's context before it runs.

Summarize this board deckQueued
People+1ApplicationsMax budget$0.04
Spans Finance + Product
Draft follow-ups from today's callsQueued
PeopleApplicationsMax budget$0.01
High repeated-work → cache candidate
Review the key-rotation PR for exposureQueued
PeopleApplicationsMax budget$0.60
Touches security scope → restricted route
Plan multi-region failover architectureQueued
PeopleApplicationsMax budget$1.20
Net-new knowledge → Knowledge Object
Reconcile these invoice mismatchesQueued
PeopleApplicationsMax budget$0.02
High repeated-work → batch candidate
Claude Sonnet 5AnthropicM01
Task fit
Reasoning Medium640 ms$0.03 / call
Gemini 3 FlashGoogleM02
Task fit
Reasoning Flash240 ms$0.03 / call
Fable 5AnthropicM03
Task fit
Reasoning High1.4 s$0.19 / call
GPT–5.6 SolOpenAIM04
Task fit
Reasoning xHigh2.8s$0.42 / call
Grok 4 FastXAIM05
Task fit
Reasoning Low310 ms$0.01 / call
Cortex-triage-v3Cortex FoundryM06
Task fit
Reasoning Eval-tuned380 ms$0.006 / call
Routed
RiskLowUrgencyNot urgentPriorityP2AccessInternal
Best-value route

Every prompt is a buying decision.

Cortex keeps work on the Pareto Frontier.

Smart routing
PromptTriage this support escalation, with prior resolutions
Route candidates
GPT-4o$0.041/1,180msOverkill
cortex-triage-v3$0.006/380msBest value
Claude Sonnet$0.018/690msBackup

Best value, not biggest model

The cheapest path that passes your evals, never the frontier model by default.

Enriched before it runs

The prompt arrives carrying what your org already knows.

Inside your policy

Every route respects your cost, latency, and risk limits, and every permission.

Prompt enrichment

Sparse prompt in. Your whole org behind it.

At run time, Cortex searches your company context graph and memory stores that Atlas builds. The task, the artifacts, and the history attach before the model runs. Ten words in, a grounded answer out, at a fraction of the cost.

Enriching meta…
Who is asking and why it matters?
Prompt

“Triage this support escalation”

Meta
Priya NairL2 support
TaskSupport triage
P0/Urgent
Max budget$0.60/task
Permissionssupport-restricted
Applications
SEC-1204Exposed key rotation
PR #4821rotate-keys +312 -88
#support-esc14 messages · 22m ago
Runbook / Key rotation v3Last update Jul 2
Context
3 similar resolutions96% accepted
Acme CorpEnterpriseEU region
PolicyPII scope · Cost cap
Prior fix / cortex-triage-v3Accepted Jun 12
0 tokensContext 0%
Beyond the prompt

Other routers read the prompt. Cortex knows the work.

Cortex picks the model using signals no prompt-based router has: the task being done, the person's role and permissions, and which model produced accepted work on that task before.

Without CortexRaw prompt

GPT-5.6 Sol41 tokens$0.42 / call

|

Answer quality0.42

With CortexEnriched prompt

GPT-5.6 Sol1,286 tokens$0.42 / call
Meta (5)Priya NairP0$0.60cap
Applications (3)SEC-1204PR #4821#support-esc
Context (3)3resolutionsAcme Corp.PolicyPII

|

Answer quality0.80
Grounded / 6 sourcesPasses evalWithin $0.60 budget
How we measure

Cortex Router benchmarks

Router measurements and directional pilots.

Tau2 · GPT-5.6 family

Directional pilot: Terra was 9 points better and 14% cheaper.

Learn more →

Tau2 · Open models

83% success at $0.309 per task on the paired cohort.

Learn more →

Cortex Router vs Sol

Sol quality with 27% savings.

Learn more →

Cortex Router vs Not Diamond

91% of Opus quality with 40% savings.

Learn more →

See Smart Routing & Enrichment on your own data.

We'll scope it as part of Cortex Foundry on your stack in days.