Engineering insights on AI routing, model orchestration, and the future of LLM infrastructure.
59 articles — Page 1 of 5
Novita retired 32 models with no replacements. Three other vendors sent notices the same week. The interesting part was not the retirements — it was that acting naively on one of them would have deleted most of our catalogue.
A 763B mixture-of-experts model that activates 8B parameters during prefill and 16B during decode, with a KV cache four times smaller than its predecessor. The headline is not the parameter count — it is that re-reading a large context stops being the expensive part of an agent loop.
DeepSeek V4 Flash is served on Antbase by nine independent providers — DeepSeek direct, Fireworks, SiliconFlow, Vercel, Scaleway and RedPill among them. You write one model string; Antbase picks the cheapest healthy route and fails over when one degrades. Fireworks moves to $0.44 per 1M input on 21 August, with cached input at $0.014 — 31x cheaper.
Google's Gemini 3.7 Flash is live on Antbase. A 1M-token context window, up to 65,536 output tokens, and text, image, video, audio and file input — through the same OpenAI-compatible API you already use. Antbase routes it two ways, OpenRouter and Vercel AI Gateway, and OpenRouter is running it at 50% off through 27 August.
Hermes is one of the fastest-growing open-source coding agents — and like most agents it ships hardwired to a single provider. One command points it at Antbase and it inherits automatic cross-provider failover, prompt caching, cost routing, and a routing trace for every request across 20,000+ models. Here's the 30-second setup.
Coding agents now dominate token spend, and each one is hardwired to a single provider — no failover, no cost routing, real lock-in. Point your agent's base URL at Antbase and it inherits automatic cross-provider failover, prompt caching, cost routing, and per-request routing traces across 20,000+ models. Here's how to wire up Claude Code, Cline, and Aider in a few lines.
Antbase is now a team platform. Invite your team into a shared workspace with roles, set hard spending caps per team and per project so a runaway job can't produce a surprise bill, and whitelist exactly which models each team or project may use. All governed from one dashboard, enforced on every request.
Turn on semantic caching and Antbase remembers answers to prompts it has already seen. The next time the same — or a close-enough — prompt comes in, it returns the stored answer instantly and for free: no model is called, and your balance doesn't move. Per-workspace, private, and off by default.
Claude Sonnet 5, Anthropic's most capable Sonnet-class model, is live on Antbase. Frontier coding, agents, and computer use on a 1M-token context with 128K output — reachable through the same OpenAI-compatible API. Antbase routes it two ways: Anthropic direct, or through NetMind at $2/$10 per 1M instead of $3/$15. You call one model string; the router picks the cheaper healthy path.
Two new Qwen3.7 models are live on Antbase through NetMind: Qwen3.7-Max, the reasoning and long-horizon agent flagship, and Qwen3.7-Plus, a lower-cost multimodal model that reads images and video. Both run on 1M-token context through the same OpenAI-compatible API — switch tiers by changing one model string.
Image and video generation on Antbase aren't a separate product or a new SDK. They're two model strings — ant:image-gen and ant:video-gen — that you call exactly like a chat model, through the OpenAI-compatible endpoint you already use. So they work anywhere that speaks OpenAI: your app, your scripts, Cursor, or the playground.
Cerebras has retired the public endpoints for Llama 3.1 8B and Qwen 3 235B Instruct. If you called them directly, you have to migrate. On Antbase, neither model was ever pinned to Cerebras — they route across every provider that still serves them, so nothing changes for you.