CodeGateway Real Cost & Reliability Data: Reproducible Latency, Success Rate, and Billing Reconciliation

CodeGateway Real Cost & Reliability Data: Reproducible Latency, Success Rate, and Billing Reconciliation

Every number on this page comes from reproducible scripts, not marketing estimates. Latency/success-rate probe: `scripts/probe-gateway-metrics.py`. Billing reconciliation: `docs/pricing/reconcile-vs-direct.mjs`. Dimensions we did not measure are explicitly labeled Not measured. This page does not constitute any SLA commitment. Formal SLA / status page is a separate project.

TL;DR

1. Billing reconciliation: what you actually pay

1.1 Three fee layers (code-level source of truth)

If we only said “1.07×”, we would understate your all-in cost by ~10 points. We chose to show the top-up side too.

1.2 $10 top-up example

credit        $10.00   (value you receive)
platformFee   $ 0.20   (2% of face)
stripeFee     $ 0.73   (ceil((1020+30)/0.961) − 1020)
─────────────────────
charged to card  $10.93   ⇒ $1.0930 paid per $1 of credit

Recompute: node docs/pricing/reconcile-vs-direct.mjs (the script self-asserts; mismatch exits 1).

1.3 All-in multiplier for three representative workloads

Worked example (W1):

• Official = 0.6M×$2 + 4.0M×$0.10 (cacheRead) + 1.2M×$2.50 (cacheWrite) + 0.03M×$10 = $4.90

• CG base prices match official (verified after the 2026-10-15 price fix); ×1.07 = $5.24

• All-in = 1.07 × 1.093 ($10 top-up amortization) = 1.1695

Batch caveat: CG has no batch tier today. If you would have used the official Batch API (50% off) anyway, W3 vs official Batch direct is about 2.34× all-in — that comparison does not apply to interactive work (W1/W2). Micro-amount footnote: if you top up exactly one workload, the fixed $0.30 Stripe fee dominates tiny amounts. In practice users use the $10 tier or free credit; the main table uses amortized economics.

1.4 Historical skew (fixed; kept for the record)

claude-sonnet-5-5 cacheRead was billed at $0.20/MTok while official is $0.10 ⇒ cache-heavy workloads briefly reached +26.5% all-in. Fixed 2026-10-14 and verified in production (GET /v1/pricing: cache_read_per_mtok=0.1). All workloads now uniformly +17.0%.

2. Reliability measurements: success rate first, latency with caution

2.1 Measurement definition (read this first)

• Environment: https://api-test.codegateway.dev (test)

• Window: 2026-10-14T08:17:21Z → 08:22:40Z (5m 19s)

• Definition: client-visible end-to-end milliseconds, cold connections (new TLS per request), including this host’s network path

• Re-run: see `docs/agent-cli-integration/metrics/README.md`

2.2 How to read latency (honest disclosure)

Network floor from the same session:

The floor from this machine to the public internet is on the order of 1.4–5 seconds. Gateway latency in the table below is dominated by the client path and must not be read as gateway server performance.

Therefore this page does not use absolute latency as a performance claim. The probe and floor-correction method are in the repo. We will publish comparable latency numbers only after a prod re-run from a clean network path. Until then, “we are faster” should not appear anywhere — including our own marketing.

2.3 Success rate (usable)

Success = HTTP 2xx and parseable body (non-streaming needs choices[0]; streaming needs ≥1 non-empty SSE delta).

Sample size and window as above. This is not a long-term SLA — it is a reproducible reading we can show in public docs.

2.4 Streaming TTFT (archived readings; not a conclusion)

TTFT = first non-empty SSE data: chunk (content or reasoning_content). Also polluted by the network floor — no performance conclusion.

3. Agent CLI compatibility

Compatibility matrix and per-CLI setup: `docs/agent-cli-integration/` (drafts + tools for Cursor / Codex / Gemini CLI, etc.).

4. Reproducible integration examples

4.1 One-shot probe (verify your path)

# API key required (see 4.2)
export ANTHROPIC_BASE_URL=https://api.codegateway.dev
export ANTHROPIC_AUTH_TOKEN=<your CodeGateway key>

# Protocol self-check
curl -sS "$ANTHROPIC_BASE_URL/v1/messages" \
  -H "x-api-key: $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-sonnet-4-6","max_tokens":16,"messages":[{"role":"user","content":"ok"}]}'

4.2 No official key / don’t want a foreign card?

Register on CodeGateway and get a key. Alipay and other local methods are supported (no international credit card required). free credit (current amount on the signup/pricing page) is enough to run the probe above.

4.3 Troubleshooting cross-links

• Connection timeouts: Claude Code connection timeout solutions

• Billing / quota questions: use GET /v1/pricing + §1 on this page; do not trust third-party cached prices

• CLI setup: `docs/agent-cli-integration/`

5. What we did not measure (explicit list)

6. Takeaways

1. Cost: token-side flat 1.07×; with top-up and payment fees fully counted, you pay about 1.17× official list. A $10 top-up yields $10.00 credit and charges $10.93 to the card. We publish the arithmetic and the scripts.

2. Reliability: short-window success rate 15/15. Latency numbers are not concluded because of the client network floor — methodology is public; we will update after re-measurement.

3. Compatibility: Claude Code verified; Codex CLI untested and unclaimed; Cursor unsupported.

4. Next: prod re-measurement from a clean network path, Codex CLI /v1/responses shape verification, production token-mix measurement.

When numbers change, this page changes with them. Scripts are traceable in git history.