Releases·Agent Release
Reasonary AI's Auker 1.0 rewrites its own model-allocation policy, cutting MMMU-Pro benchmark cost ~6× at unchanged accuracy
Reasonary AI released Auker 1.0, an agent that decomposes tasks and learns to allocate models, agents, tools, and reasoning across subproblems rather than routing through one fixed model. After 116 optimization experiments, its best policy scored 80.35% on MMMU-Pro at $9.73 total cost — about 6× cheaper than baseline at essentially unchanged accuracy. The company published the methodology, cost assumptions, and Pareto-frontier results, and is soliciting feedback on benchmark clarity and real-world stress tests.
3 engagement1 sourceWed, Sep 30, 2026, 05:01 PM
AI Overview
- Reasonary AI says Auker 1.0 splits tasks into subproblems and learns to allocate models, agents, tools, and reasoning across them instead of routing through one fixed model
- After 116 optimization experiments, the best policy scored 80.35% on MMMU-Pro at a total cost of $9.73, about 6× below its initial baseline at essentially unchanged accuracy
- The company attributes the gains to changed intelligence allocation rather than a more powerful model, and published methodology, cost assumptions, and Pareto-frontier results
- Reasonary AI links its claims to a blog post that says the policy surpassed GPT-5.6 Luna on the benchmark's Pareto frontier
Related
Products·launched
Foreman V2 — leaner AI agent orchestration that cuts cost and token usage
Foreman V2 is an AI agent orchestration system rebuilt to strip out unnecessary overhead between LLM calls. The team reports benchmark results showing large reductions in cost, input tokens, and context history versus their previous version. It targets developers running multi-agent workflows who want lower inference costs.
Foreman V2 is an AI agent orchestration system rebuilt to strip out unnecessary overhead between LLM calls. The team reports benchmark results showing large reductions in cost, input tokens, and context history versus their previous version. It targets developers running multi-agent workflows who want lower inference costs.
2 engagement1 sourceredditFri, Oct 9, 2026, 11:06 PM
Social·Community
How a Two-Model AI Strategy Cut Costs by 35% and Halved Response Time
An article on Artificial Intelligence in Plain English describes a two-model AI strategy using selective retries, which cut model cost by 35% and halved median response time while approaching the flagship model's test scores.
An article on Artificial Intelligence in Plain English describes a two-model AI strategy using selective retries, which cut model cost by 35% and halved median response time while approaching the flagship model's test scores.
0 engagement1 sourcerssSat, Oct 10, 2026, 09:51 PM
Releases·Tool Release
Weave Router 2.0 matches GPT-6 Astra pass rates at ~52-54% cost on coding benchmarks
The Weave team announced Weave Router 2.0, an open-source model router that plugs into coding agents like Claude Code or Codex and switches between LLMs per task. Benchmarked against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas, the router achieved equivalent pass rates while costing ~52-54% as much and running 2.2-2.5x faster. Three improvements drove the gains: a new HMM-plus-classifier architecture, larger labeled training data, and smarter cache-eviction impact calculation.
The Weave team announced Weave Router 2.0, an open-source model router that plugs into coding agents like Claude Code or Codex and switches between LLMs per task. Benchmarked against GPT-6 Astra on Terminal Bench 4.0 and SWE Atlas, the router achieved equivalent pass rates while costing ~52-54% as much and running 2.2-2.5x faster. Three improvements drove the gains: a new HMM-plus-classifier architecture, larger labeled training data, and smarter cache-eviction impact calculation.
2 engagement1 sourcehackernewsWed, Sep 30, 2026, 04:58 PM
Social·Community
Jev-based model routing saved 33.2% vs premium in our pilot, but a fixed mid-priced model was better value
LLM Gateway published a self-run pilot benchmarking its Jev-based model routing against premium and fixed mid-priced model baselines across 640 measured requests. Routing cost 33.2% less than the premium baseline but passed 75/79 tasks versus 79/79 for premium. A fixed mid-priced model cost 72.9% less than routing while passing 76/79, leading the authors to conclude a cheaper fixed model remains a strong alternative.
LLM Gateway published a self-run pilot benchmarking its Jev-based model routing against premium and fixed mid-priced model baselines across 640 measured requests. Routing cost 33.2% less than the premium baseline but passed 75/79 tasks versus 79/79 for premium. A fixed mid-priced model cost 72.9% less than routing while passing 76/79, leading the authors to conclude a cheaper fixed model remains a strong alternative.
19 engagement1 sourceredditSat, Sep 26, 2026, 12:49 PM
Products·building
Pareta - OpenAI-compatible router that sends LLM requests to small specialist models
Pareta offers one OpenAI-compatible endpoint where model="auto" routes each request to a small specialist model, escalating to frontier models when quality checks fail. It targets developers running cheap, high-volume LLM tasks like classification and extraction who want frontier-level accuracy at a fraction of the cost. A benchmark on 60 labeled classification/extraction items showed near-parity accuracy versus gpt-5.5 at dramatically lower cost.
Pareta offers one OpenAI-compatible endpoint where model="auto" routes each request to a small specialist model, escalating to frontier models when quality checks fail. It targets developers running cheap, high-volume LLM tasks like classification and extraction who want frontier-level accuracy at a fraction of the cost. A benchmark on 60 labeled classification/extraction items showed near-parity accuracy versus gpt-5.5 at dramatically lower cost.
3 engagement1 sourceredditTue, Sep 15, 2026, 05:05 AM