PRODUCT UPDATES

Changelog

What we've shipped. Real milestones from the build log: no vaporware, no roadmap dressed as news.

Performance

Task-aware routing: the smart routing engine ships

Every API request is now classified by task type and routed to the model proven best at that specific work. Not a static map: a continuously evaluated routing engine that tests models against real tasks and updates routes automatically.

108+ task types covered
Blind-evaluated scoring
Automatic route updates

The routing engine updates itself: new model drops, it gets tested, scores get posted, routes update. Customers never touch a line of code.

Performance

Benchmark series: LiveBench results published

Autark scored 85.7% on LiveBench: 682 tasks across reasoning, math, language, data analysis, and instruction following. Beats the best single model on the public leaderboard. Processed 3.8 million tokens for $7.25 total. Equivalent frontier quality would cost ~$57.

This was the first external validation of the routing engine. Every question went to model: "autark." Under the hood, the classifier identified each task type and routed to the proven best model. No manual tuning. No hardcoded per-category maps. Automatic.

Performance

KWBench validation: routing doubles single-model coverage

Validated against KWBench: 223 expert knowledge-work tasks across 30 categories. The best single model scores 27.9%. Routing across models nearly doubles coverage to 50.7%. 44 tasks (39%) were passed by exactly one model: nobody dominates, and without routing, you miss nearly 40% of what's possible.

This isn't a one-off. The benchmark runner is automated: any new model gets scored against KWBench in a single run, no manual setup. As new models launch, the matrix updates.

Feature

Model lineup: Flash, Autark, Deep

Three capability levels, zero configuration. Set model: "autark" and the routing engine handles the rest. Flash for cost-optimized throughput. Autark for smart routing: best model per task. Deep for reasoning-heavy work with extended thinking.

autark-flash

Fast inference for simple queries, classification, and high-throughput workloads. $1/$2 per million tokens.

autark

Smart routing to the best model for each task. The default for most teams. $3/$6 per million tokens.

autark-deep

Extended reasoning for complex analysis and multi-step problems. $5/$10 per million tokens.

Platform

New website: product-first, real numbers, no fluff

Completely rewrote autark.ai. The old site positioned Autark as an enterprise gateway with four products. The new site tells the truth: one API, smart routing, real benchmark numbers. Product pages, pricing, benchmarks, trust center, docs: all rewritten from scratch.

Every number on the site is verifiable. Every benchmark result links to the methodology. No fabricated testimonials, no roadmap dressed as product, no features listed that don't exist yet.

Feature

Waitlist and early access open

Public waitlist launched. Contact form live with enterprise inquiry routing. Early access API keys rolling out to waitlist signups.

Infrastructure

First deploy: API live on dedicated infrastructure

Autark API deployed on European infrastructure with Zero Data Retention architecture. OpenAI-compatible endpoint. Drop-in replacement for any application using the OpenAI SDK. One base URL change, one API key swap.

More coming.

This changelog tracks what ships. No roadmap announcements, no "coming soon" dressed as delivered. Just the build log.