Gateway or direct Anthropic API

Direct Anthropic and codexkey both end up calling the same model weights. The choice is operational: who you pay, how you pay, and what the account can do.

Account & onboarding

Payment methods

Direct Anthropic: corporate card, ACH, wire (US and EU). Prepaid balance, $5 minimum; monthly limits depend on the tier.

codexkey: SBP (Russia), cards (Russia, CIS, EU), crypto (USDT TRC20/ERC20, BTC, ETH, TON). Minimum top-up $1, no tiers.

Latency

codexkey adds roughly 30-60 ms to direct Anthropic for one extra proxy hop through an EU edge. TTFT in streaming is kept, since nothing is buffered. For p99-sensitive workloads, direct access can be better.

Feature parity

Model availability parity

Both codexkey and direct Anthropic serve the same Claude model lineup through Anthropic-compatible request formats. When Anthropic releases new versions (Sonnet 5, Opus 4.8, Haiku 4.5, the computer-use beta), codexkey adds them within a few hours, so there is no gap in model availability. Only the IDs differ: codexkey keeps stable gateway IDs for the native Anthropic entry point (`claude-sonnet-5`) and for the OpenAI-compatible one (`claude-sonnet-5` through /v1/chat/completions), so one model can be called in either dialect.

Code sample: switch in 2 lines

If the Anthropic SDK already works for you, the difference comes down to two values:

import Anthropic from '@anthropic-ai/sdk'

const client = new Anthropic({
  // BEFORE: defaults to https://api.anthropic.com
  // AFTER: point it at codexkey
  baseURL: 'https://api.claudexia.tech',
  apiKey: process.env.CLAUDEXIA_API_KEY,  // sk_cdx_...
})

const msg = await client.messages.create({
  model: 'claude-sonnet-5',
  max_tokens: 1024,
  messages: [{ role: 'user', content: 'Hello' }],
})

The same swap works in the Python SDK, the Go SDK and any community SDK where you can set a base URL and API key.

SDK & client compatibility

At the protocol level codexkey is compatible with the Anthropic Messages API. Anything that speaks Anthropic (the official SDKs, LangChain, Vercel AI SDK, Continue.dev, Cline, Cursor's BYOK mode, custom HTTPS clients) works unchanged. For clients that only know OpenAI (older OpenAI SDK code, some open-source agents), codexkey serves /v1/chat/completions on the same base URL.

Rate limits and quotas

Direct Anthropic applies account and model quota rules. codexkey can also be subject to account or provider-channel limits. The dashboard key API does not set RPM, TPM, concurrency or a budget for a key, so no fixed quota or extra throughput is promised. Handle HTTP 429 in your client.

Compliance and data handling

The direct API is covered by the terms of the chosen supplier. codexkey works through a chain of API suppliers for which retention periods and training terms are not yet fully confirmed. Do not carry the direct API's guarantees over to the gateway. A DPA needs a separately agreed appendix naming actual recipients, processing locations and safeguards.

Pricing transparency

We do not hide the markup. The live /pricing page shows the rate per 1M tokens, and comparing it with anthropic.com/pricing gives you the exact margin. For most workloads from Russia and the CIS the markup is lower than currency conversion costs and foreign card fees, so the net cost in codexkey can be lower than with direct Anthropic.

When to pick which

Direct Anthropic

Choose direct Anthropic if you have a payment method Anthropic supports, need zero proxy latency, need first-party support contracts, or your compliance team requires a direct vendor relationship.

codexkey

Choose codexkey if you cannot pay Anthropic directly, want pay-per-token with no monthly minimums, need payment options for Russia and the CIS, need an OpenAI-compatible format without your own translator, or need sub-org billing for multi-tenant apps.

FAQ

Will my code break when I move from direct Anthropic to codexkey?

No. codexkey mirrors Anthropic's Messages API one to one. Changing ANTHROPIC_BASE_URL and the API key is enough.

Are model versions identical?

Yes. claude-sonnet-5 in codexkey points to the same Anthropic model snapshot.

Does codexkey charge a markup over direct Anthropic prices?

Yes, codexkey adds a small per-token markup that covers payment processing and infrastructure. The exact rate is on the public pricing page; for Russia and the CIS it is often lower than international acquiring fees.

Does codexkey store my prompts or responses?

No. For billing and analytics codexkey logs only the token count, model id, latency and account id. Message content is proxied and not stored.

Does prompt caching work in codexkey?

Yes. Anthropic's cache_control markers pass through unchanged, so you get the same savings of up to 90% on cached input as with direct Anthropic.

Are tool_use, vision and JSON mode supported?

Yes, every Anthropic Messages API feature passes through one to one: tool_use, vision (images as base64 and by url), system prompts, stop sequences, temperature, top_p, top_k, metadata, streaming SSE.

What about computer-use and the new agent skills API?

The computer-use beta is supported through the same anthropic-beta header. New Anthropic features usually show up in codexkey within 24-48 hours of GA.

Will I lose Anthropic's native error semantics?

No. codexkey keeps Anthropic's error structure (type, message, request id). If Anthropic returns 529 overloaded, you see the same status with the same retry-after. SDKs that already parse Anthropic errors keep working.

Can I get a refund on unused balance?

Yes, an unused balance is refunded under the public ToS. Direct Anthropic credits are non-refundable once credited.