OpenSwitch migration report

issue-digest · OpenAI / Anthropic → NVIDIA Nemotron on Nebius Token Factory

97%lower LLM cost (est. per 1K calls)

2/2 passedlive smoke tests written by OpenSwitch (repo had none)

1call sites migrated

$0.07spent by the agent to do it

Per call site

Function	Before	Now runs on	Score	$ per 1K calls (red before, green after)
`complete`	gpt-4o	Nano 30B fast	8.3	$1.8446 $0.0583

Every config, measured

Checklist score (0-10, share of pass/fail checks met) on inputs built from the app's own prompts; green = passes (100% functional, mean within 0.75 of the best config and ≥ 7.5, no output below 5.0); outlined = chosen (cheapest passing).

Function	Nano 30B fast	Nano 30B think	Super 120B fast	Super 120B think	Ultra 550B fast	Ultra 550B think
`complete`	8.3 $0.0583	8.6 $0.1365	8.1 $0.1862	6.7 $0.4527	8.1 $0.6150	8.6 $1.3145

Tests

Iteration 1: 2/2 passed: wrote live smoke tests (repo had none)

Which Nemotron did what

Role	Model	Calls	Tokens	Cost
judge	Ultra 550B	36	31,393	$0.0352
candidate	Ultra 550B	12	5,831	$0.0116
rewriter	Ultra 550B	1	3,080	$0.0055
test-writer	Ultra 550B	1	2,818	$0.0054
planner	Ultra 550B	1	2,460	$0.0044
candidate	Super 120B	12	6,231	$0.0038
researcher	Super 120B	3	10,904	$0.0036
eval-writer	Super 120B	1	3,034	$0.0023
candidate	Nano 30B	12	7,087	$0.0012

Failure cases we found

complete · Nano 30B fast · score 5: The output adds an assignee not present in the input, the Problem sentence exceeds 40 words, the Details section includes code and stack trace not present in the issue, and the Next step describes two actions instead of
complete · Nano 30B fast · score 8: All checks pass: the output contains exactly four ## headings in the correct order; the Problem section is a single sentence of 18 words; the Details section accurately notes that no error was reported and quotes the imp
complete · Nano 30B fast · score 8: Problem section contains 17 words, exceeding the 40-word limit by a small margin but the instruction says 'at most 40 words' and this is within that; however the instruction also says 'one sentence' and the Problem secti
complete · Nano 30B fast · score 8:

Baseline cost is an estimate: source-model token counts are approximated with Nemotron's tokenizer, and for o-series models hidden reasoning is approximated with Nemotron Ultra's thinking length.
The reference answers come from Nemotron 3 Ultra with thinking, not from the original provider (no OpenAI/Anthropic key is needed or used); the repo's own tests are the independent check.
Each function was evaluated on 6 generated inputs; rare inputs can still behave differently, so keep monitoring after merging.

Sources (live Tavily research)

gpt-4o: $2.5 / $10.0 per 1M · https://pricepertoken.com/pricing-page/model/openai-gpt-4o
Endpoint URL is https://api.studio.nebius.com/v1/ · source
Authentication uses Authorization: Bearer <NEBIUS_API_KEY> · source
Supported parameters include max_tokens, temperature, top_p, top_k, n, stream, stream_options, stop, presence_penalty, frequency_penalty, logit_bias, logprobs, top_logprobs, user, and extra_body · source
JSON output can be enabled via response_format: { type: "json_object" } or response_format: { type: "json_schema", json_schema: <schema> } · source
Nemotron thinking can be disabled by setting extra_body.chat_template_kwargs.enable_thinking: false · source
For Nemotron models, when enable_thinking is false, force_nonempty_content may be needed to avoid empty content due to parser behavior · source
Anthropic-specific parameters like thinking are not supported; use chat_template_kwargs instead for reasoning control · source
The OpenAI-compatible API forwards chat_template_kwargs to vLLM for Nemotron models when properly configured · source