What we ship, when we ship it.
Tracking updates from March 2026 to present. @ctimlucas on Telegram for the latest.
Full support for Anthropic Claude 3.5 Sonnet and OpenAI GPT-4o via the unified endpoint. Both models are available with the same API Key — just change the model parameter. Pricing aligned with upstream providers.
Reduced proxy overhead through connection pooling and request batching. Average gateway latency dropped from 85ms to under 30ms at p50, and from 200ms to 65ms at p99. Upstream model response times unchanged.
Added Google Gemini 2.0 Pro and Alibaba Qwen 3 to the supported model list. Gemini accessible via the same OpenAI-compatible endpoint. Qwen 3 offers strong Chinese language capabilities at competitive pricing.
Fixed an issue where API Key rate limits were enforced globally rather than per-key. Each key now has its own independent token quota and request rate counter. Prevents one noisy user from exhausting limits for others.
Launched the self-service usage dashboard. View real-time token consumption, cost breakdown by model, daily/weekly/monthly trends, and per-usage breakdown. Export data as CSV for accounting purposes.
When an upstream model provider experiences an outage, the gateway now automatically retries with the configured fallback model. Users can set fallback preferences per API Key to control which model to use as backup.
Added support for streaming (Server-Sent Events) through the proxy. Clients can now receive model responses token-by-token instead of waiting for the full response. Significantly improves perceived latency for long outputs.
Corrected token count discrepancies between reported and actual usage for multi-byte character sets (CJK, emoji, special Unicode). Now uses the same tokenizer as the upstream provider for each model. Billing accuracy improved to within 0.1%.
Users can now create, manage, and rotate multiple API Keys from the dashboard. Each key has independent rate limits, model access restrictions, and usage quotas. Ideal for team environments and service isolation.
Expanded model catalog with DeepSeek V3 (cost-effective alternative to GPT-4) and Zhipu GLM-4 (strong multilingual support). Both models accessible through the standard OpenAI-compatible endpoint.
Upgraded gateway to support HTTP/2 multiplexing and persistent connections to upstream providers. Reduces connection establishment overhead by 60% for burst request patterns.
All error responses now follow the OpenAI error format consistently, regardless of the upstream provider. Clients using the official OpenAI SDK no longer need special error handling for non-OpenAI models.
CTIM AI gateway launched in early access. Initial support for OpenAI (GPT-4, GPT-3.5) and Anthropic (Claude 3 Opus/Sonnet/Haiku) models through a single unified endpoint. OpenAI-compatible API, zero code changes required.
Core proxy layer implemented with Go. Features include request routing, response streaming, token counting, and basic rate limiting. Internal testing with seed users.
Want to try the latest features?
Get Early Access →