Blackbox

The organization Blackbox targets is the enterprise engineering or security group that cannot send prompts into a shared multi-tenant pool. Blackbox sells trust controls as the primary argument: end-to-end encryption on every connection, zero data retention enforced at the gateway, training opt-out on by default with the upstream providers Blackbox routes to, and PII stripped before prompts reach closed models on the Enterprise agreement. Blackbox also advertises a contract option in which abuse monitoring and human review are disabled, negotiated into the contract, and then enforced technically.

Adopting Blackbox in daily work means changing a base URL. The Blackbox homepage shows a short curl against https://enterprise.blackbox.ai/chat/completions with a standard messages array and streaming enabled, so existing OpenAI-compatible SDK code keeps running. Beyond raw completions, Blackbox ships an Agents API for cloud coding agents, Remote Agent cloud sandboxes with dedicated runners on Enterprise, a VS Code extension, a CLI, and SSO sign-in across every surface, and Blackbox says each of those surfaces draws down the same token balance.

Routing behavior is part of what Blackbox is selling rather than a side feature. Blackbox includes smart routing, failover, prompt caching, and cost-aware model selection, and the Blackbox pricing FAQ claims routing plus caching extend the same commit by 10 to 20 percent, even on closed models. Customers holding negotiated rates with OpenAI, Anthropic, or Google can connect those provider accounts to Blackbox, keep routing, failover, caching, and dashboard visibility, and have that usage bill through the original provider instead.

Pricing at Blackbox is a single Enterprise motion with no self-serve card. Blackbox says there are no platform fees, no markups, no credit-purchase fee, and no seats; buyers commit annual spend through a purchase order and the balance burns down at published per-million-token rates, with the effective rate improving as committed spend grows. Blackbox lists indicative discounts of 5 percent or more on closed models and 10 percent or more on open-weight models, and publishes reference rates such as nvidia/nemotron-3-ultra-550b-a55b at $0.32 input, $0.80 output, and $0.08 cached reads per 1M tokens, moonshotai/kimi-k2.7-code at $0.74 input and $3.50 output, and zai/glm-5.2 at $1.40 input and $4.40 output.

The constraints around Blackbox are structural rather than hidden. Unused commitment expires at the end of the billing period, and Blackbox sends spend alerts only at 75 percent and 90 percent, so Blackbox itself advises sizing the commit to minimum usage. Blackbox states that an engagement starts "with conversation, not signup," which rules out a card-swipe trial, and PII removal before closed models, data residency, dedicated Remote Agent runners, custom SLAs, and a dedicated forward-deployed engineer are all tied to the Enterprise contract. Blackbox also qualifies its retention and training guarantees on routed traffic with the phrase "wherever the provider API supports it."

Against other multi-model gateways, Blackbox differentiates by operating the metal for open-weight models rather than only brokering third-party capacity. Blackbox cites Artificial Analysis measurements of 454 tokens per second on NVIDIA Nemotron Ultra as the #1 provider at 2.7x lower price than the #2 fastest provider, and points to a Microsoft AI partnership for Azure enterprise deployments. Buyers who want a metered pay-as-you-go account with a credit card will find Blackbox harder to start than a pure router, because Blackbox packages routing, dedicated hardware, and implementation engineering into one negotiated contract.

Evaluating Blackbox therefore comes down to whether the trust stack justifies the sales cycle. Teams that already need a DPA, SAML SSO, SCIM, RBAC, audit logs, data residency, and contractual zero retention get those items bundled into the same Blackbox commit that pays for tokens, with implementation quoted at $0. Teams that only need cheap access to many models can get the routing layer elsewhere without agreeing to an annual Blackbox purchase order.