BLACKBOX.AI, on the pages public for this profile, is an enterprise inference layer. You can reserve an open-weight model on capacity that the site says is not shared, or send many models through one router endpoint. A separate Agents surface creates repository runs over HTTP, and a terminal CLI can plan a change before applying it. The pricing page meters tokens against an annual commit and includes list rates per million tokens. It does not publish the monthly individual plans that review articles still quote.
BLACKBOX.AI fits a team that already knows it wants either a single-tenant open-weight deployment or one bill for many models. The pricing page says to commit annual enterprise spend for a discount on every token, with the balance decreasing as teams consume it. There is no credit-purchase fee and no per-seat charge. Unused commitment expires, and the page says it sends spend alerts at 75% and 90%. The router and the dedicated deployment are not the same promise: the router page qualifies zero retention with provider support, while Enterprise Inference says the workload is not forwarded to another customer.
A good fit for
Buying many models through one endpoint and one committed token balance.
Keeping an open-weight model on capacity the site describes as single-tenant.
Sending a repository task to an agent and reviewing the branch it opens.
Know the trade-offs
Unused committed spend expires at the end of the billing period.
The CLI FAQ says code context is sent. It does not describe a local-only run.
Dedicated inference, one router, and two agent surfaces.
A deployment reserved for one customer
Enterprise Inference is the page for a model that should not share a machine. You choose the open-weight model, and we run it as a deployment dedicated to you and isolated from every other customer, at independently verified top speed. The page says there is one tenant, reserved capacity, and nothing shared, and that Blackbox operates the deployment. You do not install or operate it in your own infrastructure. A forward-deployed engineer is described as configuring it with you at contract start. Weights you already trained are a separate control. The FAQ says Enterprise includes a private model registry, and that you upload those weights and call them through the same endpoint on the same reserved capacity. When traffic jumps, endpoints autoscale across the fleet and route to warm GPU capacity. Bursts return 503 only while capacity scales, and the page tells you to retry with backoff. Compatibility is a client change: change the base URL and key on the official client, and existing requests keep their shape. Streaming uses the same stream flag. Those are the published controls, not a load test.
The Router page puts 300+ open and closed models behind one OpenAI-compatible endpoint, and the gateway enforces zero data retention and no training. The quickstart says to keep the client you already run, collapse provider keys into one, and point the base URL at the Router so changing models is one string. Load balancing can serve the same model id from more than one provider. When a provider degrades or has an outage, the gateway moves the request to the next provider. After the response, the working content is not reused as a storage layer. The gateway enforces zero data retention and no training, wherever the provider API supports it. That last clause is the limit on the promise. The pricing page is not a seat menu. Commit annual enterprise spend for a discount on every token: the balance decreases as your teams consume it, with dedicated deployments, customer-managed keys, and custom contracts when you need them. No platform fees, no markups, and no seats. Every surface uses the same balance: Enterprise Inference, the Blackbox Router, the API, the CLI, and the Agents API. It expires at the end of the billing period. We send spend alerts at 75% and 90% of your commit, so nothing expires without warning. You can connect your own OpenAI, Anthropic, or Google accounts and keep routing, failover, caching, and the dashboard while the provider bills the usage. List rates are priced per 1M tokens. We bill input, output, and cached reads separately. One shown row is moonshotai kimi-k3 Text 1M $3.00 $15.00 $0.30. A separate line reads $0.44 BLENDED / 1M ON NEMOTRON 3 ULTRA · #1 SPEED. The enterprise Closed −5%+ Open −10%+ as discount markers, and it says the contract rate improves as committed spend grows.
The Agents page is an HTTP control surface. Create agent runs over HTTP, stream every terminal event, inspect sandbox files, and continue work that already started. Send one task to multiple agents, let Chairman LLM rank the implementations, then promote the cleanest branch to review. The parallel set named in the body is Blackbox, Claude Code, Codex, and Gemini. A successful run can become a pull request with a scoped branch, clean commits, a generated summary, and links back to the run history. The FAQ says you do not install anything: the call is plain HTTP, or the same agents run in the browser after you sign in with Google or GitHub. GitHub, GitLab, and Bitbucket are the named hosts. Enterprise adds self-managed instances and SSO. Each task runs in an isolated cloud sandbox with its own file system, network, and process boundaries, and the page says that sandbox can install dependencies and run tests. The CLI is a different entrance to a similar comparison. Blackbox CLI is an open-source AI coding agent for your terminal. Plan changes before you apply them, send four agents in parallel, and use it in any script or CI pipeline. Plan mode divides complex tasks into ordered steps that you can approve, comment on, or rewrite before any change applies. After approval, each change shows as a clean diff. Run /multi-agent to send the same task to Blackbox, Claude Code, Codex, and Gemini in parallel, each on its own worktree. A later answer adds Goose, OpenCode, and Qwen, and says Chairman LLM selects a result. The code-handling answer is direct: code context is sent, with end-to-end encryption and AES-256 encryption at rest. Enterprise can place that on a single-tenant deployment. The same FAQ says the CLI is fully open source, with the code on GitHub.
The privacy page treats exposure as a routing choice. The privacy filter replaces names, emails, account numbers, secrets, and other identifiers with semantic placeholders before a prompt reaches a closed model. Closed models receive placeholders, and the response is restored for the user. The filter is described as a local step in the Blackbox layer, before provider routing. Enterprise teams can tune the categories that must be masked. The homepage states the same enterprise rule in shorter form: on Enterprise, PII is removed before prompts reach a closed model. Zero Data Retention keeps prompts, code context, and completions in memory only. The response is returned, then the working data is discarded. The page says that means no model training, no prompt review queue, and no product analytics copy of the content. Training opt-out is on by default and is enforced with upstream providers wherever the provider API supports it. Audit logs record operational events such as access, model invocation metadata, repository access, and policy changes. The FAQ says they do not store prompt bodies. A workspace policy can reserve single-tenant Enterprise Inference for source code or records the team will not send to a third party. The homepage encryption line sits earlier in the path. Prompts are encrypted before they leave your machine and decrypted only where the model runs. Nothing readable crosses the wire, and zero data retention means nothing is kept after the answer. A separate line says the proxy, the host, and the operator never hold the key.
A suggested router pass starts from a client that already speaks the same request shape. Change the base URL, use one key, and pick a model string from the catalog. Read the list rate for that model before the commit is sized, because the shown row is a reference point and the page says the contract rate moves with committed spend. If the workload cannot accept a third-party host, the inference FAQ says that job belongs on the dedicated deployment instead.
Reserve capacity for one model
A suggested deployment pass names the open-weight model, asks for single-tenant capacity, and treats the private registry as the path for weights you upload. The page says another customer's job should not change your throughput, and that prompts are processed in memory and discarded. The speed line on the site cites Artificial Analysis for Nemotron 3 Ultra in July 2026. Do not treat that citation as a measurement of your own traffic.
Review a branch before anyone merges it
A suggested agent pass connects a GitHub, GitLab, or Bitbucket repository, sends one task, and reads the pull request the page says a successful run can open. If you use the CLI instead, approve or rewrite the plan before execution, then read the diff. Chairman LLM is the page's ranker, not a substitute for that reading. The sandbox is allowed to install dependencies and run tests inside its own boundary. The pages do not say those tests replace your review.
How to get started with BLACKBOX.AI
1
Read the pricing page, not a review table
Confirm the annual token commit, the expiry rule, and the list rates. Ignore monthly seat prices that are not.
2
Choose router or dedicated inference
Use the router for many models through one endpoint. Use Enterprise Inference when the page's isolation line is the requirement.
3
Send a task only after the plan is readable
On the CLI, approve the ordered steps. On Agents, expect a branch and a summary tied back to the run.
4
Check retention against the surface you picked
Router retention is qualified by provider support. The privacy page says prompts are discarded after the response and are not the audit log.
The public pricing page has one named offer, Enterprise, priced as custom annual committed spend. List rates are reference rows, not a second plan. One row reads moonshotai kimi-k3 Text 1M $3.00 $15.00 $0.30. Another line reads $0.44 BLENDED / 1M ON NEMOTRON 3 ULTRA · #1 SPEED.
Enterprise
A team committing spend across the router, dedicated inference, CLI, and Agents.
Custom.
Annual purchase order, balance decreases as tokens are used.
No platform fee and no per-seat charge on the public page
Same token balance for Enterprise Inference, the router, the CLI, and Agents
Unused commitment expires; alerts are described at 75% and 90%
There is no credit-purchase fee and no per-seat charge. You pay for tokens at or below the list rates. Closed −5%+ Open −10%+ beside the custom price. Connecting your own OpenAI, Anthropic, or Google account bills that usage through the provider. Figures from review articles were not copied.
You are picking an editor, not an inference contract.
The BLACKBOX.AI pages describe a router, a dedicated deployment, and agents that open pull requests. Choose Cursor when the editor is the job.
BLACKBOX.AI: common questions
Is there a free plan on the public pages?
They do not publish a free inference tier. Enterprise is listed as custom annual committed spend. Implementation is shown as included on that card, which is not the same as a free token allowance.
They are priced per 1M tokens, and input, output, and cached reads are billed separately. The page calls them list rates and says contracts are quoted against them. The row moonshotai kimi-k3 Text 1M $3.00 $15.00 $0.30 is one example. $0.44 BLENDED / 1M ON NEMOTRON 3 ULTRA · #1 SPEED is a separate shown line.
It expires at the end of the billing period. We send spend alerts at 75% and 90% of your commit, so nothing expires without warning. The page says to size the commit to minimum usage. It does not describe a rollover.
Can the router use provider accounts the team already has?
Yes. Connect your OpenAI, Anthropic, or Google accounts and route through your existing provider relationships and negotiated rates. Routing, failover, caching, and the dashboard remain available, while usage bills through your provider.
The CLI FAQ says the opposite of a local-only claim. It says code context is sent, with end-to-end encryption and AES-256 encryption at rest. Enterprise adds a single-tenant deployment on reserved capacity.
The Agents FAQ names GitHub, GitLab, and Bitbucket, with automatic pull request creation and branch management. Enterprise adds self-managed instances and SSO. The page says each task stays inside its own sandbox.
They say Blackbox CLI is an open-source AI coding agent for your terminal, and a FAQ says the code is on GitHub. They do not name a license. They also do not state a named privacy regulation. Retention and the identifier filter are described in product language instead.
NavHunt did not run BLACKBOX.AI. No account was created, no agent run was started, no repository was connected, and no token request was sent. Speeds, list rates, and retention rules come from BLACKBOX.AI pages reviewed on 6 October 2026. The Artificial Analysis figure is quoted from those pages and was not repeated. This description is based on BLACKBOX.AI's official materials reviewed on 2026-10-06, including BLACKBOX.AI homepage, Token pricing, Enterprise Inference, Blackbox Router.