Synthorai
New Sign up free, 10 calls on us. Up to $1, no card needed.

The AI gateway with governance built in.

One endpoint for every model, at list price. Prompts versioned, every run traced, spend per key and per model. On Enterprise, budgets, approvals and guardrails on every request.

Get your API key How billing works → 
# Drop-in. One line changes.
base_url="https://api.openai.com/v1"
base_url="https://synthorai.io/v1"
Live pricing 125 models · 11 providers · list price, USD per 1M tokens
Model Input Output Cache read Platform fee
deepseek-v4.1-flash-1017, DeepSeek · Release date: Input $0.3 Output $1.2 Cache read $0.006 Platform fee 0% gemini-4-argon, Google · Release date: Input $2 Output $10 Cache read $0.1 Platform fee 0% claude-opus-5-5, Anthropic · Release date: Input $4 Output $20 Cache read $0.2 Platform fee 0% gpt-6-astra, OpenAI · Release date: Input $10 Output $50 Cache read $1 Platform fee 0% glm-5.3, Z.ai · Release date: Input $1.4 Output $4.4 Cache read $0.28 Platform fee 0% qwen3.8-max, Alibaba · Release date: Input $2 Output $6 Cache read $0.25 Platform fee 0% qwen-image-3.0, Alibaba, image · Release date: Input $0.03per image Output - Cache read - Platform fee 0% fun-asr-flash, Alibaba, speech-to-text · Release date: Input $0.0021per min Output - Cache read - Platform fee 0%
A cache read is input the provider has already cached. On these rows it costs 2-20% of the input price, and each provider sets its own rate. Browse 125 models →
One endpoint, every major model

Three things you will want to check.

Will it cost more than going direct?

No platform markup, no per-seat fees. Input the provider has cached is billed at 2-20% of list price. Hard spend caps per workspace and per key stop a runaway script.

0% platform feecached input 2-20% of list priceyour own keys (BYOK), free

Where do my prompts go?

Zero retention by default: prompt and completion bodies are forwarded and not stored. Data is encrypted in transit with TLS 1.3 and isolated per workspace. On Enterprise, guardrails scan for PII and secrets before a request is forwarded.

zero retentionTLS 1.3region pin

Can legal sign it?

A DPA is available on request. Region pinning (US / EU / APAC), request logs, an audit log, SSO (SAML / OIDC) and admin and member roles.

DPASSOaudit logs

One URL for your agent, from prototype to production.

Build

A prompt library your code calls by id

Prompts are stored as versions, and a label such as production points at one of them. Your code sends the prompt id and the label, and the gateway uses that version.

The API is also an MCP server at /v1/mcp, so Claude Code, Cursor, VS Code or Codex can call it as tools.

Prompt library docs → MCP docs → Playground docs →
Prompts New prompt
Prompt Labels Versions Updated
support-reply production v6staging v7 7 Oct 20
invoice-extractor production v12 12 Oct 16
ticket-triage production v9staging v10 10 Oct 13
Promote staging v7 to production: 2 checks passed, ready.
Run

One base URL and one key for every modality.

Chat, image, speech, video and realtime voice each have their own endpoint under the same base URL, called with the same key.

Enterprisemodel=auto picks the model from what the request is about and records why.

python 
from openai import OpenAI

client = OpenAI(
  base_url="https://synthorai.io/v1",
  api_key="sk-syn-…",
)

resp = client.chat.completions.create(
  model="claude-opus-5",
  messages=[{"role": "user", "content": "How do I cut my LLM bill?"}],
)
Response
claude-opus-5

Reduce your LLM bill by using smaller/cheaper models for simple tasks, caching repeated responses, trimming prompt length, and setting max token limits.

Drop-in. One line changes. POST /v1/chat/completions
Observe

Trace an agent run, call by call

Add the same X-Trace-Id header to every request of one agent run and the console shows them as one trace. Claude Code and Codex sessions are grouped without it. You see how long each call took, how much time passed in your own app between calls and what the run cost.

Claude Code and Codex sessions group on their own, and the Exports feature forwards every request to S3 or an OTLP collector.

# the headers the first call of this run sent
X-Trace-Id: home-14c500a32dd5
X-Span-Id: plan
Docs: send trace headers →
Tracehome-14c500a32dd5
Wall time
8.7s
In the gateway
8.5s
In your app
0.3s
Model calls
3
Tokens in / out
607 / 375
Cost
$0.0021
plan gpt-6-luna 3.5s
your app 0.3s
answer claude-sonnet-5-5 1.5s
check gemini-3.8-flash 3.5s
Model call In your app
Three providers served this run, and it is still one trace. Cost at list price.
GovernEnterprise

Budgets, approvals and guardrails on every request

Each team has a budget and each member a daily allowance. A request over the allowance goes to the person who can approve it, with the context attached. Guardrail rules block, mask, guide or detect, and every hit is recorded. Available on Enterprise and in private deployment.

The three editions →
Approvals Three requests are waiting on you. Team budget $300.00 a day.
  • Emily Carter Daily allowance $20.00 → $50.00 · quarterly contract extraction Asked twice before, both approved · the team has $248.00 of $300.00 left
    Escalates in 4h Decline Approve
  • Jason Miller Access to claude-opus-5 · 30 days First time asking · spent $4.10 yesterday against a $20.00 allowance
    Decline Approve
  • Sarah Brooks Daily allowance $20.00 → $400.00 Above the team's $300.00 a day, so approving sends this to an admin
    Goes to an admin Decline Approve
Guardrail rules
  • Vault key in a promptalways onblock
  • Customer PIImask
  • Source code to a model that is not approved for itguide
  • Contract text outside Legaldetect

Three ways to run it

SaaS

0% platform fee

Self-serve. Model usage at provider list price, no subscription.

  • 125 models across chat, image, speech, video and realtime
  • Prompt library, traces, playground and MCP
  • Spend caps per key and per workspace
Get your API key

Enterprise

A dedicated instance

Run by us for your company. Model usage at list price; the platform is priced per company.

  • Usage by team and by intent
  • Team budgets, daily allowances and approvals
  • Guardrails, data classification, SCIM and a permanent audit log
Talk to sales

Private deployment

Inside your network

The gateway and the console in your VPC or data center. License and support, priced per company.

  • Everything in Enterprise
  • Prompts leave only for the providers you allow
  • Routing and judge models you host
Talk to sales
Compare every feature →

How it compares with the other gateways

All comparisons →

Synthorai vs OpenRouter

Same per-token price. OpenRouter adds about 8% in fees when you buy credits and 5% on BYOK usage above $25k a month. Here both are 0%.

Read the comparison →

Synthorai vs LiteLLM

LiteLLM is a proxy you run and wire to each provider yourself. Synthorai is a managed service, and the same product runs in your own network with its console.

Read the comparison →

Synthorai vs Portkey

Portkey charges for its platform per logged request, and you must bring your own provider keys. Synthorai has no platform fee, includes traces and lets you pay list price on one bill or bring your own keys.

Read the comparison →

Not sure which model? Compare two.

All 470 comparisons →
Featured comparison

Claude Opus 5 vs GLM-5.2

USD per 1Mclaude-opus-5glm-5.2
Input $5 $1.4
Output $25 $4.4
Cache read $0.5 $0.26

Both share a 1,000,000-token context window and optional thinking, so the real split is modality and price: claude-opus-5 takes image input as well as text, while glm-5.2 is text-only at about 3.6x less on input ($1.4 vs $5) and 5.7x less on output ($4.4 vs $25). Pick claude-opus-5 for prompts containing images or heavier reasoning-plus-tools work; pick glm-5.2 for high-volume text and long-context jobs, where its $0.26 cache reads versus $0.5 and slightly larger 131072-token max output (against 128000) help.

Read the verdict →

DeepSeek V4 Flash (0731) vs Gemini 3.7 Flash

Both are text-in reasoning-and-tools models with roughly a million tokens of context (1000000 for deepseek-v4-flash-0731, 1048576 for gemini-3.7-flash), so the split is really about modalities and output volume. Pick deepseek-v4-flash-0731 for text-only work at lower cost - $0.44 input and $1.32 output against $0.75 and $3.75, about 1.7x and 2.8x cheaper - and for long generations, since its 393216 max output tokens is 6x the 65536 cap. Pick gemini-3.7-flash when you need image, audio or video input, which deepseek-v4-flash-0731 does not accept at all.

Read the verdict →

GPT-5.6 vs Qwen3.8 Max

Both take text and image in and return text, with chat, code, tools and reasoning on each, so the split is mostly price and control: gpt-5.6 costs twice as much per input token ($4 vs $2) and about 3.3x as much per output token ($20 vs $6), with cache reads at $0.4 against $0.25. Pick qwen3.8-max for high-volume or output-heavy work and its long-context flag over a 983616-token window, plus a slightly larger 131072-token max output. Pick gpt-5.6 when you want the marginally wider 1050000-token context or the ability to turn thinking off per request.

Read the verdict →

Three steps to your first call.

We cover the Stripe processing fee on top-ups. SaaS is self-serve. Talk to sales for a managed instance or your own deployment.

  1. Sign up free

    Google, GitHub, or email. You're in the console in seconds, no invite needed.

  2. Pay-as-you-go or BYOK

    Top up in USD and pay each provider's list price, with hard caps per workspace and per API key. Or bring your own provider keys (BYOK) for free. The provider bills you directly and we add nothing.

  3. Change the base URL and the key

    Set your OpenAI SDK's base URL to https://synthorai.io/v1 and use your Synthorai key as the API key. The rest of your code stays the same. An Anthropic-compatible endpoint is also available.

Get your API key