CodeGateway vs Voyage AI

Voyage AI is an accuracy play on the retrieval half of RAG. The lineup splits three ways: general-purpose embedders, domain-tuned models for finance, legal and code, and company-specific models fine-tuned on proprietary data. The Harvey case is the concrete one — a voyage-law-2 model fine-tuned on Harvey's datasets cut the irrelevant-document rate by 25% and cut embedding dimensionality and vectorDB costs threefold. voyage-multimodal-3.5 puts images and text through one retrieval pipeline; voyage-context-3 pairs chunk-level detail with global document context; rerank-2.5 and rerank-2.5-lite add instruction following so ranking behavior can be steered, with the lite tier as a cheaper price-performance point. Shorter vectors (3x–8x), a 4x smaller model with faster inference, 2x cheaper inference, 32K-token context and a Batch API round out the engineering story.

CodeGateway solves a different problem entirely: consolidating access and billing. One key covers 89 models across Claude, GPT, Gemini, DeepSeek, xAI, Qwen, Mistral, Moonshot and Zhipu, exposed through three endpoint styles (Anthropic Messages, OpenAI Chat Completions, OpenAI Responses) plus an OpenAI-compatible images route billed per image to the same wallet. The practical hook is a one-line base_url swap for Claude Code, Cursor, Codex CLI, Aider, Crush, opencode and Copilot CLI, with native SSE streaming pass-through and per-model cache write/read rates on a live table. Team workspaces with owner/admin/member roles, shared balance pools and usage export cover the admin side. Nothing here embeds or reranks; nothing in Voyage routes or bills.

The two price in ways you can't put side by side. CodeGateway is transparent to the point of tedium: a flat 1.07× consumption multiplier over Anthropic, OpenAI and Vertex official list prices, a prepaid wallet with no monthly fee, $5 minimum top-up and balances that never expire. Token billing is metered to $10^-10 precision with no round-up loss, and per-model cache write and cache read rates are published on a live pricing table. Watch two things: the 1.07× is on top of whatever the underlying vendor charges (its 2026-10-22 post compares Claude and OpenAI list prices as of 2026-10-15), and fixed Stripe fees mean frequent small top-ups erode the headline rate. Its 2026-10-24 billing-precision whitepaper documents integer micro-dollar accounting, streaming metering and refund top-up rules — useful if you need the audit trail.

Voyage AI is paid with no tier structure disclosed in the data here, so the honest comparison is value-per-outcome rather than list price. Its savings claim is structural: 3x–8x shorter vectors cut vector-search and storage spend, a 4x smaller model runs faster, and 2x cheaper inference lands with superior accuracy. The Harvey result — threefold lower embedding dimensionality and vectorDB costs alongside a 25% drop in irrelevant documents — is the number to model against your own corpus. If you're a hobby project still testing whether RAG works, neither the price nor the tuning effort is justified yet.