octa-mini Preview
fast drafting
The low-latency tier for high-volume drafting, classification, and routing where speed wins.
latency profile Milliseconds
Models
GTM models, as an endpoint
Preview the octa family tuned for go-to-market over the same API, SDK, and MCP patterns as the
rest of graph8. Trained on a decade of real campaign outcomes, not the public web.
octa model
One GTM brain, called like any endpoint.
octa-mini Milliseconds
octa Default
octa-reasoning Long-horizon
6 task heads one per revenue job
01 / Task-specific GTM models
What the models are. One model per revenue job.
The Octa models are not one general chat model behind a prompt. Each is focused and trained for a specific GTM task: email writing, campaign creation, brand voice and assets, sequence building, reply qualification, landing pages, account research, and pipeline forecasting. The six heads below are those task-specific models.
Every one is built for go-to-market and revenue generation, not generic prose. A generic LLM knows what a cold email looks like. octa is trained on which email got the meeting.
- octa Research Account research
Trained to turn the open web plus your private context into campaign-ready account briefs, persona docs, and competitor maps.
account research, market mapping - octa Copy Email and campaign writing
Trained on which message booked the meeting, not what a cold email looks like. Carries brand voice and brand assets through every draft and sequence.
email writing, campaign creation, sequence building, brand voice - octa Pages Landing pages
Generates the full landing page, layout, headline, form, and proof, scored against what historically converted clicks into pipeline.
landing pages, page intent - octa Qualify Reply qualification
Reads inbound replies against your ICP, scores fit, and routes to the right rep with the context attached.
reply qualification, inbound routing - octa Coach Rep coaching
Reviews live conversations and reply threads, surfaces what worked, and gives reps the next-best move on the deal.
call review, next-best-action - octa Forecast Pipeline forecasting
Predicts pipeline conversion from stage signals and flags slipping deals before the close date, learned from outcomes not guesses.
pipeline forecasting, deal-stage signal
Training data provenance
graph8's own GTM data, not the public webFoundation models are bottlenecked by their training data. The Octa models are trained only on graph8's own go-to-market data: a decade of CIENCE outbound, labeled by humans for why deals converted or died. Trained on outcomes, not the open web.
10+ years
Of real GTM outcomes
Continuous CIENCE outbound from 2015 onward, across 1,000+ B2B brands. Not the public web.
1,000+ brands
In the training corpus
Real B2B accounts across SaaS, finance, healthcare, manufacturing, media, and public sector.
100Ks
Of emails and qualified replies
Thousands of real campaigns, multi-channel across email, voice, SMS, social, chat, and landing pages.
Human-labeled
For why deals converted or died
Reps tagged the why behind every win and loss. octa trains on outcomes, not token likelihood.
02 / Model preview access
Preview access The primitives exposed in preview.
Inference over the supported model catalog, the Skills runtime, knowledge base RAG, and AI enrichment are available in approved preview workspaces. Each links to the model API page.
Model inference Preview access
API / SDK / MCP Call the supported model catalog over the API, SDK, or MCP. List the model ids, then run inference inside an LLM skill.
Model API → Skills runtime Preview access
LLM tasks Author and execute LLM skills as reusable, parameterized tasks. Compose any supported model with your own prompts and context.
Model API → Knowledge base RAG Preview access
embeddings Semantic search over your knowledge base with embeddings, ready to ground any generation in your own content.
Model API → AI enrichment Preview access
per-column Run model-backed enrichment over contacts and companies, one column at a time, against an approved workspace slice.
Model API → Agent runtime Preview access
113 MCP tools Durable, tool-using execution with the graph8 MCP surface exposed for preview workspaces.
Model API → Voice agents Preview access
dialer Real-time voice agents wired to the dialer, with transcription and reply handling built in.
Model API → Reply-intent classification Preview access
inbound Classify inbound replies by intent so your sequences and routing can react automatically.
Model API → 03 / The Octa model family
Preview Three tiers, six heads, one GTM brain.
graph8-hosted Octa inference is in preview. Pick a latency tier, then route to the head that matches the job.
octa Default
balanced agentic
The balanced workhorse for agentic GTM work: tool use, multi-turn reasoning, and grounded generation.
latency profile Default
octa-reasoning Preview
deep multi-step
The deep-reasoning tier for long-horizon, multi-step plans where quality matters more than latency.
latency profile Long-horizon
Proof inside graph8
octa-intent
Internal productiongraph8's own page-intent model. It powers the proprietary intent signal inside graph8 today.
octa-doc
Pilotgraph8's own GTM document model for Studio docs. Validated in a pilot, generating the global document set.
Cost target
Target, not GA pricing10 to 20%
The hosted Octa tiers aim for 10 to 20 percent of frontier model cost on GTM workloads. That is a target for the hosted models, not committed pricing. Preview access does not yet bill at this rate.
04 / Why GTM-specialized
The data moat is the outcome, not the email.
Octa learns from more than ten years of CIENCE campaign outcomes. It does not just know what a cold email looks like. It learns which message got the meeting, which sequence advanced the deal, and which signal preceded the reply.
A frontier model is trained on the open web. Octa is trained on what actually moved revenue. That is the difference between fluent and effective.
Read the full Octa story Frontier LLM
Trained on the open web. Knows what a cold email looks like.
octa
Trained on real campaigns. Learns which cold email got the meeting.
Ten years of campaign outcomes
CIENCE campaign telemetry, joined to who replied, who booked, and who closed. The labels other model builders never see.
Tuned for the meeting
The objective is the booked meeting and the advanced deal, not token likelihood. Octa optimizes for the GTM outcome a frontier model never sees.
Already shipping signal
octa-intent runs the proprietary page-intent signal inside graph8. octa-doc is in pilot, generating Studio documents. The moat is working, not theoretical.
05 / Where to next
Pick your path into the models.
call it today
Model API
List the model ids and run inference over the API, SDK, and MCP.
Open → early accessCustom models
Per-org models tuned on your workspace data and outcomes.
Open → previewOcta-squared
The self-improving octa loop: campaigns feed the next weights.
Open → keys + usageOpen the console
Scoped keys, usage, and logs for your preview workspace.
Open →