Cache hit rates and Cost per Task by provider are now live on Artificial Analysis: Cost per Task for GLM-5.3 (max) varies ~9x across providers, driven by a combination of cache hit rate and token pricing Agentic workloads resend the full conversation history every turn, so most input tokens are cacheable, and cache pricing and cache hit rates drive most of the gap in Cost per Task. We already use measured cache hit rates to calculate Cost per Task, and are now publishing them for each model endpoint we track. What's new: ➤ Cache hit rates are now visible for all first-party and third-party endpoints, measured by running multi-turn agentic coding tasks and reading cached tokens from each provider's usage response. Cache hit rate is the share of cacheable input tokens served from cache. ➤ Cost per Task is now available by provider for open weights models, calculated from each provider's input, output, cache hit and cache write prices, applied at its measured cache hit rate. To standardize across providers for a given model, we use the same input, output and cacheable token counts from the model's Intelligence Index run. Key takeaways: ➤ Cost per Task on GLM-5.3 (max) spans ~9x across providers, from $0.92 on Makora to $8.68 on Nebius. Both serve nearly all cacheable input from cache (98% and 100%) and charge the same output price ($4.40 per 1M tokens), so the gap comes from input pricing: Makora charges $0.10 per 1M cached input tokens, while Nebius offers no cached input discount on its shared serverless endpoints and bills all input at the full $1.40. ➤ Cache hit rate ranges from 41% to 100% across 21 providers (25 endpoints). 15 endpoints are at 98% or above, and eight are at 100%: Modal, DeepInfra, Parasail, Together AI, DigitalOcean, Nebius, Novita and Z AI's first-party API. Multiverse Computing is lowest at 41%, followed by Bitdeer at 50%.

We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves. We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them: "/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started" I imagine we'll have some backend services that could be replaced entirely by this new architecture.