One gateway to every frontier model — with Meridian, our own reasoning core, as the default route. Auditable at every step, deployed inside your perimeter.
01 / 03
Meet Meridian.
Our frontier reasoning core — language, vision and control loops in one latent space. This is what the switchboard defaults to.
02 / 03
Every decision, inspectable.
The render becomes the blueprint: each inference pass is traced, logged and replayable.
03 / 03
Deployed inside your perimeter.
Your VPC or dedicated silicon. Nothing — weights, prompts, traces — ever leaves.
Every frontier model. One switchboard.
Opus, Fable, GPT, Gemini, DeepSeek, GLM — all live behind one gateway, one contract, one audit trail. And at the center sits our own frontier core.
We call it Meridian — the line every measurement is taken from.
11 MODELS ON THE BOARD · MERIDIAN-2 DEFAULT ROUTE
One API.
Policy picks the path.
Your teams write against a single contract. The gateway classifies each request and routes it — Meridian by default, a specialist when the policy says so. Every hop is traced.
policy "production" {
route = meridian-2
fast_path = meridian-swift # < 2 ms budget
escalate = claude-opus-4.8 # long-horizon analysis
bulk = glm-5 # batch, multilingual
residency = mistral-large-3 # EU-only data
trace = full # every hop, replayable
}On the board today.
SWITCHBOARD ROSTER · REV 2026-07- Meridian-2OBHYDefault route · reasoning, tools, vision4.2 ms
- Meridian SwiftOBHYLatency-critical paths · streaming UX1.1 ms
- Claude Opus 4.8ANTHROPICEscalation · long-horizon analysis18 ms
- Claude Fable 5ANTHROPICAgentic + creative workloads16 ms
- Claude Mythos 5ANTHROPICResearch perimeter only21 ms
- GPT-5.2OPENAIGeneral fallback · broad coverage19 ms
- o4-proOPENAIVerification cross-check · math34 ms
- Gemini 3 UltraGOOGLELong-context corpora · 10 M tokens27 ms
- DeepSeek-R2DEEPSEEKCost-optimized batch reasoning31 ms
- GLM-5Z.AIMultilingual bulk processing24 ms
- Llama 4 MaverickMETASelf-hosted fallback · air-gapped12 ms
- Mistral Large 3MISTRALEU data residency22 ms
P99 GATEWAY OVERHEAD INCLUDED · THIRD-PARTY MODELS VIA PROVIDER SLA
Measured against the board it runs.
We benchmark Meridian against every model on our own switchboard, continuously, on customer-shaped workloads — then publish the tape.
Reasoning composite
SCORE / 100- Meridian-294.1
- Claude Fable 593.6
- Claude Opus 4.892.9
- GPT-5.291.7
- Gemini 3 Ultra90.8
- DeepSeek-R289.4
HIGHER IS BETTER
Serving latency
P99 · MS- Meridian Swift1.1
- Meridian-24.2
- Llama 4 Maverick12
- Claude Fable 516
- GPT-5.219
- Gemini 3 Ultra27
LOWER IS BETTER
Blended cost
USD / 1 M TOKENS- Meridian-22.10
- DeepSeek-R22.80
- GLM-53.20
- GPT-5.27.50
- Claude Opus 4.89.00
- Gemini 3 Ultra8.20
LOWER IS BETTER
INTERNAL EVAL SUITE · JUL 2026 · CUSTOMER-SHAPED WORKLOADS · METHODOLOGY AND RAW TAPES AVAILABLE UNDER NDA
What ships in the box.
Not a model API with a dashboard. A complete reasoning system — retrieval, execution, audit and serving — delivered as one deployable unit.
SPEC SHEET · REV 2026-07
- 1.2 B docs
Retrieval fabric
A native index over your corpus, permission-aware at query time.
1.2 B docs
- 40 k runs/day
Agent runtime
Plans, executes and verifies with full rollback on failure.
40 k runs/day
- 100 % replay
Audit trace
Every inference step logged, diffable and replayable on demand.
100 % replay
- SOC 2 · II
Perimeter deploy
Runs where your data lives — your VPC, your keys, your silicon.
SOC 2 · II
- 4.2 ms p99
Line-rate serving
Speculative decoding on dedicated hardware, measured at the 99th percentile.
4.2 ms p99
Cleared for regulated work.
- SOC 2 TYPE II
- ISO 27001
- HIPAA READY
- EU AI ACT ALIGNED
- FEDRAMP IN PROCESS
See it reason on your own data.
A 45-minute working session with our deployment engineers — your corpus, your perimeter, live traces on screen. Meridian against any model you bring.
4.1 B TOKENS ROUTED / DAY · 99.98 % GATEWAY UPTIME · 12 REGIONS