Open models are now priced for long tasks, and reasoning is a dial you pay for

Reasoning becomes a setting

SiliconFlow has added DeepSeek V4 Pro with configurable reasoning levels, so developers can choose how much thinking a coding task gets instead of accepting a fixed amount. The host's Chat Completions API exposes a thinking_budget parameter with a default of 4,096 tokens and a range from 128 to 32,768 tokens, according to SiliconFlow's post. Reasoning levels do not change the model weights, widen the context window, or grant extra tool permissions, the post states, which means the dial controls effort, not capability. m:…8939

That distinction is the point for anyone running coding agents. A short budget keeps simple edits cheap and fast, while a large budget gives a hard debugging task more room to work. The cost follows the budget: thinking tokens are output tokens, and output is the most expensive part of the price list. m:…8939

DeepSeek V4 Pro on SiliconFlow supports a context window of about 1.049 million tokens and up to 393,000 output tokens, the post states. It also supports JSON mode, though schema-enforced structured outputs are not currently supported, and it does not currently accept image input on SiliconFlow. Those limits shape which tasks fit the model, regardless of price. m:…8939

The price list tells the story

SiliconFlow prices DeepSeek V4 Pro at $1.50162 per million uncached input tokens, $0.135 per million cached input tokens, and $3.135 per million output tokens. The gap between cached and uncached input is more than tenfold, which rewards agents that reuse context across turns instead of resending it. For long tasks with repeated tool calls, cache behavior may matter more than the headline input price. m:…8939

The contrast with GLM-5.3-Flash, also newly listed on SiliconFlow, is stark. Z.ai's model is priced at $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens, with a 1M-token context window. It uses a mixture-of-experts architecture with 18 billion active parameters out of 320 billion total, and it is Z.ai's first natively multimodal model in the GLM-5 series, according to SiliconFlow's post. m:…5772

Benchmark scores reported by Z.ai put GLM-5.3-Flash at 63.4 on DeepSWE v1.1, 84.3 on Terminal-Bench 2.1, 78.4 on Toolathlon Verified, and 48.8 on AutomationBench. DeepSeek V4 Pro, released August 13, 2026, is listed with 87.9 on Terminal-Bench 2.1, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, and 31.8 on AutomationBench Public. The two models trade places across benchmarks, so the choice is not simply the cheaper or the higher-scoring one; it depends on which task type dominates a workload. m:…5772 m:…8939

Both models are released under the MIT license, according to SiliconFlow, and GLM-5.3-Flash supports self-hosted deployment with SGLang, vLLM, Transformers, and KTransformers. That keeps the hosted price list honest: developers who find the hosted math unattractive can run the weights themselves, at the cost of operating the infrastructure. m:…8939 m:…5772

Cost tracking has to keep up

Arize Phoenix, the observability tool, now updates its built-in model prices automatically, according to Arize's post. The Phoenix server repeats its pricing catalog check about once a day, downloads the catalog from the Arize Phoenix repository on GitHub, and applies updated rates only to new spans after the refresh. No software upgrade is required to receive new or revised rates, and pricing checks run in the background without a server restart. m:…6666

The catalog is generated from LiteLLM's model price list, and custom pricing configurations take precedence over built-in catalog prices, Arize states. Costs already recorded on existing traces are not recalculated, and multi-server deployments use a database lock so one server applies an update at a time. Automatic updates are enabled by default, and setting an environment variable disables them. m:…6666

The practical effect is that cost attribution stops going stale as fast as hosts reprice. When a host changes a rate or adds a reasoning tier, traces recorded after the refresh reflect the new price, while older traces keep their original costs. For teams budgeting agent workloads, that is a small but real improvement in knowing what a long task actually cost. m:…6666

What to do with this

Start by measuring where tokens actually go in a real workload: cached versus uncached input, reasoning versus final output, and how often a task retries. The price lists above make those splits visible, and the cheapest model on paper can lose if it burns more reasoning on retries or resends context that a rival would have cached. m:…8939 m:…5772

Then set reasoning budgets per task type rather than one global default. Simple edits get a small budget, long debugging sessions get a large one, and the budget becomes a cost control rather than an afterthought. That is the shift these listings represent: developers are no longer buying tokens alone, they are buying a task profile. m:…8939