Does a routed LLM subscription quietly serve you a cheaper model?
Only when you let it route. On Standard Compute an exact model ID such as anthropic/claude-sonnet-5.5 always goes to that model, and provider failover stays inside that model; the routed model standardcompute picks per request, and the dashboard's Logs page shows which model served each one.
How to check which model answered
| What you send | What runs, and where you see it |
|---|---|
| An exact model ID | That model on every request. The response's model field returns the ID, and Logs shows it |
The routed model standardcompute | A model from your selection, chosen per request. The response's model field says StandardCompute; Logs shows the served model, tokens, cost and latency |
| Claude Code's short names (sonnet, opus, haiku) | Smart routing, with the reply labelled by the name you sent. Check Logs, or pin with the full ID, e.g. anthropic/claude-sonnet-5.5 |
Who decides what is easy, and how to overrule it
- A small classifier model reads the last six messages and rates the request's difficulty. The router then picks from the models you selected on the Models page, up to 5, ordered from strongest to most efficient.
- A slider on the Models page leans that choice from Efficient to Max quality. Select a single model and the router is skipped.
- For one hard task, send the big model's exact ID. A one-line ask that is secretly hard is the case for pinning.
- Routing ignores how much budget is left, unless the key has Smart pacing, which new keys no longer get. At the budget, requests stop with HTTP 402.
- When a provider fails, a routed request with no model selection can move to a different model. Pinned and selected models only move to another provider of the same model.
Which routes run FP8
Claude, GPT and Gemini are closed-weight models that only their makers and licensed clouds serve, so a routed plan reaches the same weights you would call directly. Open-weight models can be hosted at any precision. Standard Compute pins those routes to named endpoints, and several are FP8: GLM 5.3, DeepSeek V4.1 Flash, DeepSeek V4 Flash and the MiMo V2.6 models. GLM 5.3 Flash runs on an FP8 and an FP4 endpoint. When precision matters, pin a closed model or test the open model on your own repository first.
Model quality on agentic coding, from public benchmarks
| Model | Intelligence Index | Terminal-Bench 4.0 |
|---|---|---|
| Claude Opus 5.5 | 57.6 | 59.6% |
| Claude Sonnet 5.5 | 56.0 | 63.6% |
| GPT-6 Astra | 52.7 | 59.1% |
| GLM 5.3 | 44.8 | 41.9% |
| DeepSeek V4.1 Flash | 39.5 | 26.8% |
| GPT-5.6 Luna | 37.3 | 11.6% |
| DeepSeek V4 Pro 0813 | 36.0 | 14.1% |
Highest reasoning setting each model offers, measured by Artificial Analysis and read on October 13, 2026. Source: Artificial Analysis (artificialanalysis.ai).
Run your own side-by-side
- Pick ten tasks from last month that you already know the right answer to, including two that took you a long time.
- Run them through the direct API and through the gateway with the same exact model ID, so routing is out of the comparison.
- Then run the same tasks with the routed model and compare what you keep: tests passing, diffs merged, cost per task from Logs.
- Keep both setups for a week before moving anything permanent.
The catch
Routing saves money by sending routine steps to efficient models, and on some steps an efficient model will do worse than Opus would have. What you get is visibility and control: pin when it matters, select the models you trust, and read Logs.
Where Standard Compute fits
Standard Compute is an LLM gateway for coding agents with monthly plans from $19 (USD) and a fixed compute budget: Standard is $89 with $95 of compute. Every plan serves every model, exact IDs pin, and the routed model standardcompute picks per request from the models you selected, with each choice shown in Logs.
FAQ
Flat-rate LLM subscriptions advertise smart routing. How do I verify which model answered my request?
On Standard Compute, open the dashboard's Logs page: every request shows the model that served it, with tokens, cost and latency. A request with an exact model ID never goes to another model, and its response names that ID. Claude Code's short names such as sonnet or opus are routed, so pass the full ID anthropic/claude-sonnet-5.5 when you want Sonnet itself.
If a subscription API routes 'easy' requests to smaller models, who decides what is easy, and can I force the big model?
On Standard Compute a small classifier model rates each request's difficulty from the last six messages, and the router picks among the models you selected, up to five. Yes, you can force the big model: send its exact ID, such as anthropic/claude-opus-5.5, and the router is skipped.
Is the model on a routed plan quantized?
It depends on the model. Claude, GPT and Gemini are closed-weight and served by their makers or licensed clouds. Open-weight models run on hosted endpoints, and on Standard Compute several are FP8 (GLM 5.3, DeepSeek V4.1 Flash, DeepSeek V4 Flash, MiMo V2.6), with GLM 5.3 Flash on FP8 and FP4. Pin a closed model when precision matters.
If I switch Aider from Claude Sonnet direct to a routed subscription, will a serious refactor get worse?
Not if you pin. On Standard Compute set OPENAI_API_BASE=https://api.stdcmpt.com/v1 and run aider --model openai/anthropic/claude-sonnet-5.5: every request goes to Claude Sonnet 5.5, with the same 1M-token context, and a refactor session never moves to another model. Routing (the model standardcompute) is what saves budget on routine edits, and that is where an efficient model can do worse; keep the refactor pinned and route the rest.
Can I A/B a direct Claude API setup against a routed subscription for a week before moving my workflow?
Yes. Run the same tasks on both with the same exact model ID, for example anthropic/claude-opus-5.5, so the only difference is the path; then run them again with the routed model to see what routing changes. Logs on Standard Compute gives the cost per request for the comparison.
Is DeepSeek V4, GLM 5.3 Flash or Qwen 3.8 27B good enough to drive a coding agent on a real codebase?
For routine edits often yes, for multi-file architecture changes test before you trust them. Standard Compute serves all three on every plan next to Claude and GPT, so you can pin one on a real task, read the diff, and keep a frontier model selected for the hard steps; with the model standardcompute the router does that split per request.