Streaming tool-call deltas turn a base-URL swap into a per-model parser project

Start with the parser, because that is where the migration budget goes. In an OpenAI-style stream, tool-call arguments arrive as incremental deltas that you accumulate per call index until the buffer parses as JSON. The comparison's finding is that some providers instead send a whole JSON object per chunk, which the author flags as having a large impact on your parser ^[9]. Kimi K2.6 is the named case: the standard payload is accepted, the deltas have historically been non-standard in some versions, and teams end up maintaining version-specific parsers ^[14].

Arity is the second divergence. Some providers emit only one tool call at a time rather than several in a single assistant turn ^[8]. An orchestrator written to fan out three calls from one turn and join the results does not fail loudly against such a provider. It serializes. Every tool call is another round trip in the loop ^[12], so that regression shows up in wall clock and nowhere else.

Now the table, which is a claim about someone else's workload. 70.1 against 67.8 is a 2.3-point margin, 3.4 percent relative ^[18]. For that to transfer, your tool schemas, your parallel-call pattern and your error-recovery paths have to resemble the harness: identical schemas and prompts across all ten variants, mock downstream services, GPT-5.5 as baseline, and recovery cases limited to malformed JSON, downstream exceptions and empty results ^[2]. This is one harness, self-reported in a dev.to comparison ^[3]. A 3.4 percent gap is the size that prompt wording and retry policy move on their own.

The Alibaba demonstration will get quoted more than it deserves. Qwen3.7-Max ran a single task unattended for 35 hours with 1,158 cumulative tool calls ^[17]. That is one call every 109 seconds ^[19]. The source presents it as evidence of context retention across long agent loops ^[17], which is the honest reading of it, and it is not a throughput result.

The cost side is the thinnest part of the report. Teams that budget for per-model adapters are said to reach tool-use accuracy at or above GPT-5.5 levels at a fraction of the cost ^[16], with no price published to check that against. What is specified is the work. MCP adapter quality varies enough between providers that you test each implementation ^[15]. MiniMax M3 introduces custom extension fields that require additional parsing before production ^[10]. DeepSeek, GLM and Qwen hand you the easy half, accepting OpenAI-format functions and tools schemas including tool_choice and multi-turn tool-result messages ^[7].

In my context that makes the base-URL swap worth an afternoon as a spike and worth nothing as a plan. It is the cheapest line in the diff, and the only one that works first try. The six dimensions the report actually scores, protocol compatibility through total integration effort ^[2], are the project.