October 13, 2026
Cost per Task: one GLM-5.3 task costs 9x more at some providers than others
The cost of a single agent task on GLM-5.3 varies 9x across providers, and the gap comes from caching, not speed.
Artificial Analysis
@artificialanlys
Cache hit rate and Cost per Task by provider are now available on Artificial Analysis: Cost per Task for GLM-5.3 (max) varies about 9x across providers, driven by the combination of cache hit rate and token prices Agentic tasks resend the whole conversation history on every turn, so most input tokens can be cached, and cache pricing and cache hit rate drive most of the gap in Cost per Task. We already calculated Cost per Task from the measured cache hit rate, and now we publish it for every endpoint of the models we track. What's new: ➤ Cache hit rate is now shown for all endpoints, first-party and third-party. We measure it 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. We calculate it from each provider's prices for input, output, cache hits and cache writes, applied to the measured hit rate. To compare providers on the same basis for a given model, we use the same input, output and cacheable token counts from that model's Intelligence Index run. Key takeaways: ➤ Cost per Task on GLM-5.3 (max) varies about 9x across providers: from $0.92 on Makora to $8.68 on Nebius. Both serve almost 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 gives no discount on cached input on its shared serverless endpoints and bills all input at the full $1.40. ➤ Cache hit rates across 21 providers (25 endpoints) range from 41% to 100%. 15 endpoints are at 98% or higher, and eight are at 100%: Modal, DeepInfra, Parasail, Together AI, DigitalOcean, Nebius, Novita and Z AI's own API. Lowest is Multiverse Computing at 41%, followed by Bitdeer at 50%.
· 5.3K views
The same agent run on GLM-5.3 costs $0.92 at Makora and $8.68 at Nebius.
On Oct 13, Artificial Analysis started showing cost per task and cache hit rate next to every endpoint. For open-weights models you can now see who charges more for the same run and why.
It's the cache. An agent resends the whole conversation history on every turn, so almost all of the input can be served from cache. Makora charges $0.10 per million cached input tokens, while Nebius gives no cache discount on its shared serverless endpoints and bills all input at the full $1.40.
Output costs the same at both, $4.40 per million tokens. Both serve almost all cacheable input from cache: 98% at Makora, 100% at Nebius. The ninefold gap comes from the input price alone.
Cache hit rate used to be hidden. Artificial Analysis already used it to calculate cost per task but didn't publish the number. Now it's shown for every endpoint: it's measured with multi-turn coding tasks, and cached tokens are read from the provider's response.
Across 25 endpoints from 21 providers, cache hit rates range from 41% to 100%. Eight endpoints hit a full 100%, including Together AI, DeepInfra and Z AI's own API. Lowest are Multiverse Computing at 41% and Bitdeer at 50%.
So far Artificial Analysis has opened per-provider costs only for open-weights models.