Capture & fix · Model routing

The right modelfor every job.

Quick answers go to fast models, hard bugs to deep reasoning, fixes to coding models, and a different vendor checks every fix. If a provider goes down, the next one takes over. Leave it on Auto, or set any task yourself.

Join the waitlist See how routing works

Tore opens soon. Join the waitlist and we'll email you the day you can start.

app.tore.ai/beacon-data/routing

Example routing

Example routing: six jobs arrive and each is sent to a model that fits it. A how-to question goes to Gemini Flash at low. A pre-sales question goes to Claude Sonnet at medium, a choice Alex set by hand. An error spike is grouped by Grok Fast at low. The root cause for a bug is sent to Claude Opus at high; when that provider times out, the fallback, Gemini Pro at high, takes over. The fix and its test are written by Codex Astra at high, and a different vendor, Gemini Pro at medium, reviews it.

Model routing · Beacon DataAuto (Recommended)Manual

JobWhy this modelModel · level

"How do I export to CSV?"Maya · chat widget

fast answer from your help center

Gemini Flash · low

"Do you integrate with Linear?"Visitor · pricing page

conversational, pre-sales

Claude Sonnet · med

set by Alex

1,284 errors in 10 minutesSentry · beacon-web

high volume, keep it fast

Grok Fast · low

Root cause for TCK-482Export stops at 1,000 rows

deep reasoning over your code

Claude Opus · high

Gemini Pro · high

timeout, fallback took over

Write the fix and its testsrc/export/paginate.ts

coding model

Codex Astra · high

Review the fixbefore your engineer sees it

second opinion, different vendor

Gemini Pro · med

Auto picks each model and shows the reasonSet any job yourself, any time

A how-to and a hard bugare different jobs.

So Tore doesn't use one model for both. Each task gets the model that does it well, at the level it needs.

Fast where it can be

Quick answers use fast models. Deep reasoning is saved for the hard bugs.

paying top-model prices for a how-to

Support keeps running

If a provider has an outage, the next model in the chain takes the job.

an inbox that stops when one vendor does

A second opinion on every fix

A model from a different vendor reviews each fix before your engineer is asked.

taking one model's word for it

See what each model costs

Every AI action is metered, so spend shows up per model and per product.

one AI line on the invoice

Auto when you want it.Yours to set when you don't.

Per-task routing

Each job, its own model and level.

Turn on Auto and Tore picks a recommended set-up, with the reason for each choice. Prefer to decide? Set the model family and level for any task, and your choice sticks.

One switch: let Tore route everything
The reason for each choice, shown next to it
Override any task, and see who set it

settings · model routing

Example routing

Gemini FlashlowmedhighAnswer a how-to · Auto: short answers from your help center

Claude SonnetlowmedhighAnswer a pre-sales question · set by Alex, Oct 21

Claude OpuslowmedhighFind the root cause · Auto: deep reasoning across files

Codex AstralowmedhighWrite the fix and its test · Auto: coding model

Fallback chain

A provider goes down. Support doesn't.

Every task has a backup. If a model times out or a provider has an outage, the job moves to the next model in the chain and carries on. You see it in the log, not in your inbox.

A fallback set for every task
Switches on timeouts and outages, on its own
Every switch recorded, with the reason

routing · fallback log

Example data

Answer a how-to

Gemini Flash

Claude Haiku

14:02Maya asks how to export to CSV

14:02Gemini Flash: provider timeout

14:02Switched to Claude Haiku, the next in the chain

14:02Answered with the help center article cited

14:31Gemini Flash back. Routing returned to normal.

Second review

One writes. A different vendor checks.

The model that writes a fix never reviews its own work. A model from a different company checks it first, so blind spots don't get approved twice. Your engineer sees the fix after both agree.

The reviewer always comes from a different vendor
Notes go back to the writer and get fixed before you see them
By default it then waits for your engineer's approval

fix · TCK-482 · review

Example data

Writer

Codex Astra · highvendor: OpenAI

Reviewer

Gemini Pro · medvendor: Google

Change matches the root cause at paginate.ts:57

Note sent back: keep the header row on an empty export. Fixed.

Ready for your engineer

Reviewer must be a different vendor from the writer

Learning and spend

Better per dollar, and you see the bill.

Routing learns from what works across Tore and suggests changes that keep results and lower cost. Every AI action is metered, so you can see spend per model and per product, and decide.

Routing that learns from what works across Tore
Suggestions you accept or keep as they are
Spend per model, attributed to each product
Budgets, caps and a kill switch in AI controls

usage · spend by model · October

Example data

Claude Opus$96.40

Codex Astra$70.10

Claude Sonnet$33.80

Gemini Pro$24.90

Grok Fast$8.20

beacon-web 71% · beacon-mobile 29%$233.40

Pre-sales answers on Claude Haiku have matched Sonnet's results across Tore. Switch to save on this task?

SwitchKeep Sonnet

Early access

Bring your own AI keys during early access. Early teams can plug in their own Claude, OpenAI, Grok or Gemini keys while Tore still routes every task.

Tune it

Routing AutoFallback on for every taskReviewer different vendor

All AI controls →

1switch to let Tore route everything, with reasons shown

2models on every fix, from two different vendors

4model families to route between: Claude, OpenAI, Gemini and Grok

3levels per task: low, medium and high

Keep exploring.

Stop picking one model.Use the right one each time.

Start a trial with routing on Auto. Change any task whenever you like.