Just published aggregate stats from AI Roundtable, where 200 models debate your question. 29,502 public sessions. 334,589 model responses. askroundtable.ai Three takeaways: @claudeai Opus 4.7 most influential, @GoogleDeepMind Gemini 3.1 Pro most used, @xai Grok 4.1 Fast highest conviction. Thread.

Welcoming Infercom to the Opper gateway. Sovereign inference, made in Europe.

Claude Cowork now works with 300 models via Opper. Route through EU-hosted inference, add fallbacks, or swap to a cheaper model mid-session — same Cowork window, different routing under the hood. Setup takes 3 fields. Guide: opper.ai/blog/claude-cowork-…

All the best agent frameworks can now run inference through Opper. Agents are just code that runs models. So they need what every production system needs: routing, observability, guardrails, fallbacks, and a model catalog that doesn't lock you in. • OpenClaw — the open-source personal agent running on millions of machines • pi — the terminal coding agent powering OpenClaw • Hermes by Nous Research — open-source agentic coding assistant • Vercel AI SDK — the de facto standard for AI in TypeScript apps • Continue.dev — the open-source coding assistant for VS Code and JetBrains • Cline — the autonomous coding agent built into VS Code • OpenCode — terminal-based AI coding for people who live in the shell One API key. 260 models. EU-hosted. See our integrations page for more details: docs.opper.ai/overview/integ…

@opperai just launched this fun page where you can get any llm to debate on a question. i particularly love this one where most them are just plain wrong but none change their answer!

“Car Wash” test with 53 models opper.ai/blog/car-wash-test (news.ycombinator.com/item?id…)

Reran every model 10 times (via @opperai gateway). Same prompt, no system prompt, no cache. The results got worse. Of the 11 that passed once, only 5 held up. GPT-5: 7/10 GPT-5.1, GPT-5.2, Claude Sonnet 4.5, every Llama, every Mistral: 0/10

Today we are introducing our new Agent SDKs: opper.ai/blog/new-opper-agen… We built these to offer a good starting point for building headless, reliable and extendable agents. SDKs are available for Python and Typescript and offers the following features: * Tool support (with MCP) * Hook system to extend the inner agent actions * Model interoperability with task completions * Observability and evaluations Only needs an Opper API key, which is available on our $10 free tier.

Wk 37, 2025 in #eutech: 🇸🇪EcoDataCenter (€600M), 🇸🇪@opperai acquires @FinetuneDB_com, 🇱🇹Kashimi ($1.4M), 🇸🇪@saltfish_ai ($730K). By @tech_eu #NordicMade #cphftw #helyes #siliconfjord #sthlmtech #estotech #startinLatvia #LTstartups tech.eu/2025/09/12/mistral-r…

Opper AI acquires FinetuneDB for AI model tuning tech.eu/2025/09/11/opper-ai-…

Join the conversation on Reddit about our GPT-OSS Benchmarks: How GPT-OSS-120B Performs in Real Tasks reddit.com/r/LocalLLaMA/comm…

Join the conversation on Reddit about our GPT-5 Benchmarks: How GPT-5, Mini, and Nano Perform in Real Tasks reddit.com/r/OpenAI/comments…

We at @opperai just published high level results and a leaderboard of task benchmarks for leading models Current leaderboard: Overall winner: xAI Grok 4 Grok 4 is the winner of agentic tasks (tied with o3) and normalization tasks. In the top 5 on all categories. Context usage: Claude Sonnet 4 This tests the models ability to correctly answer questions from supplied information. This tests "reading" context. Agent runtime: Open AI O3 and xAI Grok 4 This tests the models ability to plan, reflect and select appropriate actions to take. This tests "using" context. Normalization tasks: xAI Grok 4 This tests models ability to coherently produce output in a specific format from input. This basically tests "output" format consistency. SQL generation: Open AI GPT-4.1 This tests models ability to interact with a database with natural language goals. This tests a certain domain problem. Each category has around 30 tests of easy, medium and hard tasks. I think these evals mirrors the overall "vibes" of these models! What categories we should add? Coding? Multimodal? Drawing?

✨ Introducing Custom Evaluations — Test Model Responses and Build Real Feedback Loops Today, we're introducing `opper.evaluate()` — flexible scaffolding for evaluating model responses, built right into our SDKs. Because no matter how clearly we describe a task, models are still probabilistic. You can't just trust the output. You have to test it. ✅ Support custom evaluators — code, eval frameworks, or LLM-as-a-judge. ✅ Automatically upload and track eval results on the platform — filter, observe, fix. ✅ Act on evaluation results directly inside your code — close the loop, not just measure it. Pricing: $0.50 per 1,000 metrics

✨ Introducing Custom Evaluations — Test Model Responses and Build Real Feedback Loops Today, we're introducing `opper.evaluate()` — flexible scaffolding for evaluating model responses, built right into our SDKs. Because no matter how clearly we describe a task, models are still probabilistic. You can't just trust the output. You have to test it. ✅ Support custom evaluators — code, eval frameworks, or LLM-as-a-judge. ✅ Automatically upload and track eval results on the platform — filter, observe, fix. ✅ Act on evaluation results directly inside your code — close the loop, not just measure it. Pricing: $0.50 per 1,000 metrics