Frequently Asked Questions

Vangu provides API access for chat, coding, agent and image workflows. The text catalog includes gpt-6.1-terra, gpt-6.1-sol, gpt-6-astra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna. Image models appear in the live catalog below. The plans page shows available offerings; authenticated GET /v1/models shows models accessible to your key.

gpt-6.1-terra supports visual reasoning, natural writing and general-purpose tasks. It accepts text and image input and returns text; it is not an image generation model. Tool calls are available through the Responses API. Supported-tool restrictions and your package quota rules still apply.

No. gpt-6.1-sol and gpt-6-sol are distinct model IDs. 6.1 Sol is the newer Sol catalog option for advanced coding, debugging and architecture analysis. Use the exact ID in the request’s model field rather than a generic “sol” name. Model access, pricing and request multipliers are separate settings.

Consider 6.1 Sol or the Sol family for coding and technical analysis, Astra for deep complex analysis, Luna for short chat, summaries and classification, and Terra for general text and translation. Compare accuracy, latency and token usage on a small representative task. Not every task needs the highest-tier model.

No. On request-quota plans, the multiplier determines how many request units a call consumes. General catalog examples are Luna 1x, Terra 1.5x, Sol/6.1 Sol 2x and Astra 4x; your purchased plan’s coefficients apply. A 2x call consumes two request units. Pay-as-you-go instead prices input, cached input and output tokens using the USD tariff.

On supported models, reasoning effort changes the amount of reasoning allocated to a task. Higher settings can increase latency and output/reasoning token usage. For 6.1 Sol, 6 Sol and Astra, choose low for lower latency or an appropriate higher level for harder tasks. Check the documentation for accepted model-specific values.

The public catalog does not mean every plan includes every model. Check your active plan’s models, key model allowlist and endpoint scopes, plan validity, remaining quota or balance. Authenticated GET /v1/models and your dashboard’s plan details are the relevant sources for your access.

Daily allowances reset at 00:00 Türkiye time (UTC+3). Unused daily allowance does not roll over. Weekly or monthly subscription validity is separate from the daily reset period: a midnight reset does not reactivate an expired plan.

A weekly subscription applies its daily plan rules during its validity period. A one-time term-based pack uses the fixed shared pool and expiry specified by the product. An add-on with a daily allowance may reset daily during its own validity. Check quota type, reset period and validity together before purchase.

“Unlimited” means the text request allowance is not capped at a fixed request count. Concurrency, included models, image allowance, validity and any package token budget still apply. Endpoint input/output and operating limits also remain in effect. Current terms appear on the plan card and in your dashboard.

A configured package token budget covers total input and output tokens over the package’s validity period. Keys and chats using that plan share the budget; cached input is included in total input. Reasoning tokens are part of the provider’s output total and counted once. This budget is separate from the daily request allowance.

Eligible daily plans can use the early-reset option shown in the dashboard. The default weekly allowance is two; your assigned remaining allowance appears in the panel and resets on Monday. Same-day reuse is restricted. An early reset removes one day of validity from each daily plan reset together; the operation is rejected if sufficient time is unavailable.

An add-on works alongside the existing base plan and keeps its own expiry. Model access and available quota are evaluated against active entitlements. If base and add-on daily allowances are reset early together, each loses one day of validity. Replacing a base plan is a different operation from assigning an add-on.

An administrator-assigned hourly plan starts when it is assigned: one hour is exactly 60 minutes. Access ends at expiry; an hourly plan does not automatically renew into a weekly or monthly subscription. Upcoming-expiry and expired-plan emails are not sent for hourly plans.

Concurrency is the number of API requests that can remain open for your account at the same time. All keys and applications using that account share the capacity. A streaming request may remain active while content is being delivered, so generation time alone is not the full request lifetime. Check your plan’s limit in the dashboard.

No. When concurrency capacity is full, a new request is rejected with 429 concurrency_exceeded rather than queued on the server. Capacity is released when an active request completes, fails or its connection closes. Keep client parallelism within the account’s shared limit.

No. Concurrency measures open requests; RPS/RPM measures calls over time. Long tasks hold capacity longer. Throughput depends on the model, context, output length, network and protective controls; there is no fixed per-second throughput guarantee for every model.

Set stream: true on a supported request and consume SSE events through the terminal event. Verified text or tool fragments are forwarded as they arrive. keepalive/ping keeps the connection alive and is not answer text or billed tokens. Client, proxy or UI buffering can make fragments appear together at the end.

Deep reasoning tasks can reason before emitting visible answer text. Raw reasoning tokens are not ordinary answer text and are not fully exposed in Chat format. Time to first text and total completion time differ. Configure suitable client timeouts for long tasks and distinguish connection keepalives from actual completion.

Eligible transient connection failures before output begins may be retried safely. Once text or tool output has started delivery, a regenerated answer is not spliced into the existing response; an interruption ends explicitly as an error. A client must not treat a response without a terminal event as complete or execute a partial tool call.

The catalog includes gpt-image-2, gpt-image-2.5, gpt-image-2.5-flare and gpt-image-2.5-sunburst. Flare and Sunburst IDs select their modes directly; Image 2.5 requests can also use supported mode/quality options. Your plan must include image access and have remaining image allowance.

In package mode, image calls use a separate image allowance rather than the text request pool. Its amount and daily/weekly/monthly reset period depend on the plan. In balance mode, the published image-model tariff applies. Check current entitlements on the plans page and in the dashboard’s quota section.

Use /v1/images/generations for new images, /v1/images/edits for changes to an existing image and /v1/images/variations for variations. Editing requires image files and, when applicable, a mask; clients may need multipart input. Match model-specific parameters to the API reference examples.

Supported requests can select response_format: "url" or "b64_json". URLs may expire, so store images in your own storage for permanent use. Resolution, quality, file format and transparency depend on the selected model and endpoint; invalid combinations are rejected.

Use https://api.vangu.dev/v1 as the base URL. Chat uses /v1/chat/completions; native Responses uses /v1/responses. Configure Python with OpenAI(base_url="https://api.vangu.dev/v1", api_key="VANGU_API_KEY") or Node.js with new OpenAI({baseURL:"https://api.vangu.dev/v1",apiKey:"VANGU_API_KEY"}), supplying your own key and exact model ID.

Use the OS-specific quick setup or desktop setup tool in the Codex documentation. Configure provider base URL https://api.vangu.dev/v1 and wire API responses. Supply your dashboard key and an entitled model, such as gpt-6.1-sol. Existing terminal or application sessions may need to restart to load new settings.

The Anthropic-compatible base URL is https://api.vangu.dev/anthropic; its messages endpoint is /anthropic/v1/messages. Follow the Claude Code documentation, supply your Vangu key and map default/large/small-task settings to entitled Vangu model IDs. Protocol compatibility does not mean Claude models are included in your plan.

For client-defined function/tool calls, the model generates a tool name and arguments; your application executes it and sends the result back. Preserve tool definitions, call IDs and result pairing. The server does not automatically execute a client tool’s shell command on your computer.

When the model requests a tool, return the result with the same call ID in the next request. Tasks may require multiple rounds; each API round follows its own quota/billing rules. Your application should manage round limits, timeouts, cancellation and tool permissions. Do not execute incomplete JSON arguments as a successful call.

Provider prompt caching may report reused identical context prefixes as cached input; it does not mean unique answers become identical. API-side compression can reduce unnecessary volume in supported command results without extra client installation. Not every result is compressed, and no fixed quota-saving percentage is guaranteed.

Choose Package Quota or Balance: Pay-as-you-go when creating a key. Package keys use active entitlements; balance keys use the wallet. Exhausting a package does not automatically move all requests to wallet billing. Manage the relevant key’s billing mode explicitly in the dashboard.

Text models have separate USD prices per million input, cached input and output tokens. Reasoning is included in the provider’s output total. TRY billing converts the stored USD tariff using the automatically refreshed USD/TRY rate. Image models use their image tariff. Check current API prices; request multipliers do not replace token pricing.

Card payments are processed through Stripe. External crypto, cash or bank-transfer collections can be recorded by an administrator and appear with their actual method in the customer payments page. Plan access is assigned when the payment entry includes that option; a payment-record-only entry does not itself grant a new plan.

No. A wallet top-up and a wallet-funded plan purchase are different transactions. Even if the balance was originally topped up by card, the later plan purchase is shown as Wallet. The purchase deducts wallet funds instead of requiring a second card collection for that amount.

Check the order’s status and details on the payments page. If card authentication or verification is ongoing, complete the same order rather than creating a second payment. Manual collections also require the plan-assignment option to be processed. If the issue persists, open a support ticket with the order ID and payment time.

Use separate production and test keys. Do not expose keys in public browser code, repositories or support messages. Configure scopes, models, IP restrictions and expiry in the panel; rotate or revoke keys as needed. Enable TOTP/2FA and keep recovery codes secure.

Vangu does not use your requests or responses for model training. Requests are sent to the selected provider for service, and its data-processing terms also apply. When enabled, conversation content is retained encrypted for a limited period. Usage, cache and replay records follow separate retention rules; see the privacy page for current details.

chat.vangu.dev is the chat interface linked to your account. Model access, quotas and configured token budgets follow account rules. Messaging, files/images and supported tool workflows depend on the chosen model and active entitlements. You can return to the customer dashboard from chat.

Share your personal link from the dashboard’s Referral Program section. Customers attributed to your link appear in the panel. Eligible successful external payments credit commission at your assigned rate, defaulting to 5% and adjustable per account. Wallet spending is not a new external collection; refunds and payment reversals may adjust commissions.

For 401, check that you send the correct Vangu key without extra whitespace, that it is not revoked and that the app is not using an old key. For 403, check endpoint scopes, model/IP permissions and account status. Share error codes and request IDs with support, without sending the secret key.

No. concurrency_exceeded means open-request capacity is full; insufficient_quota indicates a plan, entitlement or balance limit. Check validity, remaining request/token/image allowances and key sublimits. Continuous retries cannot resolve exhausted quota; the relevant allowance must reset or new valid access must be added.

These errors can involve provider connections, temporary load, incomplete responses or deadlines. Correlate the error code, time and request ID with logs. Safe transient failures before output may be retried; regenerating an already-started answer should be treated as a separate request. Also check client/proxy timeouts and streaming support.

Input tokens represent context; output includes the answer and any reasoning included by the provider. Cached tokens are a subset of input and are not added again. Cache rate is cached input divided by total input. Separate estimated or failed measurements from actual provider usage; a high cache rate does not guarantee a fixed subscription-quota saving.

Include the model ID, endpoint, approximate Türkiye time, request ID or available cf-ray, error code and reproduction steps in a dashboard support ticket. Remove API keys, passwords and personal data from code samples. Include the order ID for payment issues.