OpenRouter

A five-step method for routing between a cheap and a strong model using a structured-output confidence field, with thresholds derived from your own error rates and re-tuned in production.

AssemblyAI's first text-to-speech model, aimed at voice agents, is now routable via OpenRouter. It offers 4,000+ voices with US or EU serving, and Cekura's benchmark reportedly rates it well on speed, accuracy and naturalness at $15 per million characters or less.

OpenRouter labels which models allow their outputs to be reused as training data and adds enforce_distillable_text to hold a request to that set. The post pairs it with NVIDIA NeMo Data Designer for generating structured synthetic datasets and distilling large models into task-specific ones.

OpenRouter benchmarks TypeSafe's Jev decision model against Claude Opus 5 on 3,080 Banking77 classification utterances, finding Jev trails by 3.3 accuracy points but runs 13x faster at 1/22 the cost.

OpenRouter reports Anthropic's Opus 5.5 leads all models in share of platform spend, with Sonnet 5.5 and Haiku 5.5 also launching strongly, pointing to a rebound against OpenAI in real usage.

OpenRouter's January changelog: partition sorting deprioritises providers that miss a throughput or latency threshold, the auto router takes wildcard model filters across 58 models including Opus 4.5, the SDK gains an idempotent skills loader, and p99 gateway latency drops sharply.

Reasoning traces become a first-class field on OpenRouter via include_reasoning, starting with DeepSeek R1 and its derivatives. Sample code shows harvesting R1's reasoning and injecting it as context for a much weaker model to lift its answers.

Deep research becomes an API call rather than a product surface: Perplexity's Sonar Deep Research, Reasoning Pro and Pro arrive on OpenRouter with search and reasoning steps priced separately, joined by Cohere Command A, Jamba 1.6 and free Gemma 3 models.

The @openrouter/agent SDK collapses the agent loop into one callModel function: Zod-validated tools, validation of arguments the model produces, composable stop conditions such as stepCountIs and maxCost, concurrent text, tool-call and reasoning streams, and cost data per run.

OpenRouter replaces price-dominant default provider routing with a multi-factor score. A switchback experiment shows lower latency and fewer capacity errors at slightly higher cost. Price-only routing remains available via sorting and pinning.

OpenRouter adds retrieval to every model through an :online suffix or web plugin, retrieving via Exa at $4 per 1,000 results. The injected system message is published and configurable, along with the result count, so you can control cost and prompt shape.

OpenRouter details NVIDIA's Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters aimed at cheap, high-volume agent steps like tool calls and validation, contrasting it with the heavier Nemotron 3 Ultra.

OpenRouter makes provider selection controllable per request: :nitro sorts providers by tokens per second, :floor by price, and provider.sort exposes price, throughput or latency directly. Choosing a sort turns off the default load balancing.

OpenRouter's upgraded Auto Router classifies prompts into about 30 task types and ranks models by real community spend share over the prior 7 days, filtered by cost tier. It keeps multi-turn chats on one model to avoid cache rebuilds, with no extra fee.

Cloudflare's open-source Clef (27B) and Clef Flash (9B) decision models are now available through OpenRouter. They take text, JSON or images and return typed answers with probabilities. Input costs $0.24/M and $0.09/M, and output is free. Details come from OpenRouter's post only.

OpenRouter adds Cerebras as a provider, quoting 3,000+ tokens per second on Llama 4 Scout and 1,800+ on Llama 3.3 70B. The wafer-scale design keeps 40 GB of memory on chip, removing the external memory bottleneck that caps decode speed.

An explainer on the layer between an app and multiple model providers: unified request format, failover, routing, guardrails and cost attribution, plus where OpenRouter, LiteLLM, Portkey, Kong AI Gateway and Helicone differ on hosting, compliance and setup time. Written by OpenRouter.

OpenRouter extends zero completion token insurance to every model: a response with no output tokens and a blank, null or error finish reason is not billed, even when the upstream provider charges for prompt processing. Failed calls read as zero cost.

TypeSafe released Jev 1.14.0 with small gains across the board and the biggest improvement in multilingual use. It adds experimental image input, a box type, and opt-in long context behind headers. The model is now on OpenRouter, pricing is unchanged, and the post sets a deprecation policy.

Classification of LLM traffic against a taxonomy you define, applied by a small model after the request completes so inference latency is untouched. Sampling rates keep the tagging bill proportional, and the tags land in logs and usage analytics.

OpenRouter shipped a rebuilt provider router that ranks providers by quality relative to the actual cost of a request, not mostly by price. Better-performing providers win more traffic. The linked methodology post was not collected, so scoring details are unknown.