TrueFoundry Pricing: Plans, Usage, and Gateway Cost Worksheet

TrueFoundry pricing has three published plan shapes. Developer is free for up to three users. Pro costs $25 per user each month and includes a per-user request allowance, followed by request overages. Enterprise uses a custom annual commitment. A procurement total also needs model-provider charges and external guardrail charges; deployment and support; retention; plus the operating work required to prove access and policy behavior.

TL;DR

• Developer is free for up to three users and includes 10,000 requests and tool calls per user each month, plus seven-day log retention.
• Pro is $25 per user each month and includes 20,000 requests and tool calls per user. It charges $20 for each additional 100,000 requests.
• Enterprise uses a custom annual commitment and adds custom retention and audit exports; deployment choices and data residency; priority support; plus SLA terms.
• Model-provider and managed external guardrail charges sit outside the displayed estimate. Test request policy and evidence separately from the invoice.

TrueFoundry pricing starts with plan rights

The TrueFoundry pricing page lists Developer at $0 and Pro at $25 per user each month. Enterprise has a custom annual commitment. Developer supports up to three users and two teams, with three service accounts. It includes 10,000 AI Gateway requests and tool calls per user each month; five model accounts and five MCP servers; one virtual model group and one virtual MCP; one budget rule; plus seven-day log retention.

Pro allows unlimited users and up to ten teams, with up to 50 service accounts. Its published allowance is 20,000 requests and tool calls per user each month. Overage costs $20 for each additional 100,000 requests. The plan adds built-in guardrails and up to ten budget rules; up to ten guardrail policies and 30-day retention; SSO and custom roles; plus a support portal.

Enterprise adds unlimited teams and service accounts; custom retention and audit exports; external identity-provider mapping and third-party guardrails; advanced observability; VPC or on-premises deployment; data residency and priority support; SLA; plus onboarding. The public page gives no numeric Enterprise total.

Build the cost worksheet

Put each commercial meter and entitlement on its own row. The worksheet should include the published and contracted values; billing unit and forecast; owner and source; plus the acceptance test. Start with these lines:

• Users: active-user definition and monthly count; joiners and leavers; plus service-account treatment.
• Gateway volume: included requests and tool calls; expected traffic and burst assumptions; plus overage price.
• Teams and service accounts: environment separation and workload identities, plus growth allowance.
• Retention and exports: log duration and export format; deletion behavior; plus evidence access after termination.
• Guardrails: built-in policies and custom or third-party services, plus any separate managed charge.
• Deployment: SaaS and VPC; on-premises and air-gapped; data residency and infrastructure; upgrades; plus support.
• Provider spend: token and embedding; reranking; plus other model charges kept outside the TrueFoundry subtotal.

Make the sheet visually boring. One white row per cost owner is better than a merged "platform" cell that hides user and request charges, plus model and support charges.

For a seven-user Pro example with one million monthly requests and tool calls, the live calculator showed $175 for users and $180 for gateway traffic after 140,000 included requests, producing a $355 monthly estimate. Treat that as a calculator state observed on October 29, 2026, rather than a contractual quote. Recreate the estimate with your own inputs immediately before approval.

Model and guardrail charges stay separate

TrueFoundry states that model-provider costs pass through and that it does not mark up tokens. The pricing calculator also says model-provider charges and managed external guardrail charges are excluded. Keep those expenses in separate ledger rows, even when the gateway presents a unified operational view.

Estimate provider cost by model and token class, with region and workload recorded separately. Add retry and fallback behavior because a failed primary call followed by a secondary provider can change both usage and spend. For guardrails, name the built-in policy or external service, followed by its billing unit and traffic coverage. A plan entitlement to connect a guardrail does not establish the external vendor's price.

The AI gateway guide gives category context. My preference is to reject any gateway quote that collapses subscription and model spend, plus external safety services, into one annual number. Separate owners make invoice reconciliation and incident review much faster.

Gateway capabilities need production tests

The TrueFoundry AI Gateway page describes unified model access and centralized API-key management; observability; per-user or per-service rate limits; cost or token quotas; RBAC; routing and fallbacks; guardrails; plus several deployment modes. It also describes centralized request and response logs with metadata such as user ID and team, plus environment.

Turn those claims into tests. Route a fixed call to a permitted model, then repeat it with a denied identity or exceeded budget. Force a provider failure and confirm the selected fallback. Verify which model received the request and which key was used. Check how cost was attributed and what appeared in the retained log. Export the record and inspect timestamp and identity metadata; destination and action; plus correlation identifiers.

The page also advertises sub-3 ms internal latency and 99.99% uptime, along with 10 billion-plus monthly requests and an average 30% cost optimization. Those are vendor-published claims. A buyer should request the measurement method and contractual SLA, then run a representative workload. This article uses the figures only to define diligence questions.

The LLM observability guide covers telemetry. Observability can show request volume and latency; errors and models; plus cost. Authorization needs evidence that the system evaluated the caller and applicable policy before forwarding the request.

Compare the authenticated request-policy boundary

TrueFoundry's page describes RBAC and quotas, plus metadata policies, guardrails, and audit logging. Test the exact policy object and evidence rather than awarding a generic checkmark. For one deliberately routed HTTP model call, capture the authenticated identity and role or service account; route; prompt classification where applicable; policy version and destination; outcome and timestamp; plus integrity controls.

Change one attribute and repeat the call. A lower-privilege identity should produce the expected decision for a retrievable reason. Exceed a quota and confirm the action. Remove identity metadata and observe the failure posture. Then export the record to the system that owns audit retention.

DeepInspect covers authenticated user or agent-to-LLM HTTP traffic deliberately routed through its stateless proxy. TrueFoundry covers a broader gateway job that includes provider access and routing; fallbacks and cost; observability and quotas; plus deployment. The two can coexist when request order and identity propagation; timeout ownership and retry behavior; plus log correlation are explicit. The AI policy enforcement guide explains the narrower control record.

Direct SaaS AI sessions and local execution, plus calls that bypass the configured route, remain outside DeepInspect. MCP tool authorization also needs its own boundary and record, even when the associated model call crosses an inspected route.

DeepInspect

DeepInspect evaluates authenticated HTTP model calls deliberately routed through its proxy. It applies policy to identity and route, plus destination and request data, before forwarding allowed traffic, then writes a per-decision record that can correlate with TrueFoundry's gateway and observability record.

TrueFoundry subscription and request usage; provider routing and model spend; external guardrails; retention and deployment remain separate worksheet lines. Direct browser sessions and local execution, plus bypass traffic and tool authorization, stay outside DeepInspect's claim unless the relevant model request crosses its configured route. Book a technical deep dive at deepinspect.ai.