@optio may be right on OpenAI subscription on pricing, I'm not a direct subscriber since rather give Microsoft or Anthropic money...but may actually be better deal after CoPilot's pricing changes. I have both Claude Code and CoPilot paid subscriptions.
What I used to do was Claude Code to make complex/long-running/single prompts and/or plans, which is not a lot of token... then run them in CoPilot - picking a model, including "premium" ones, using their now old per-request pricing which could run for minutes/hours. In CoPilot I often used GPT 5.5, which is quite good (at least when access via a paid CoPilot subscription). GPT 5.5 is "somewhat better" than Sonnet 4.6, but not as good as Opus 4.7/4.8...but Opus is 3x cost of GPT 5.5 but seeming 0.5-1x better... Sonnet is actually pretty good, but GPT 5.5 seems a bit better, with both being real close in costs. Again just my opinion. But I might have to consider an OpenAI subscription since CoPilot pricing for the model is not as subsidized. e.g. There are a lot funky arbitrations on model pricing, so hard to know.
But the initial benchmark does show have some TDB "good" tools helps all models. And even Opus with training alone could "lose" to a free-ish model with good tools in some cases... Basically a spin of "garbage-in, garbage-out" (e.g. a minimum prompt to frontier model, may be worse).
One thing that likely help any LLM with agents is point some user-level instruction (~/.../CLAUDE.md and/or ~/.../AGENTS.md etc. - most LLM hardness have some file that read for all prompts) which you mention checking MikroTik new's https://manual.mikrotik.com/llm.txt would likely offer immediate improvement to most models. Or at least, one sentance that "reminds" the agent there are docs to read when processing a prompt is unlikely to hurt. The new docs have a markdown format, so be more token-efficient. See:
...and if you prompt your agent to read the above post, and have it add manual.mikrotik.com to the "user-level instructions", it figure out how to write content into the user-level AGENTS.md/CLAUDE.md/copilot-instructions.md/etc file...then future agent can "reason" if it's useful to "read the manual" based on whatever you prompted, since it already know how to access it. And once MikroTik actually publishes MD, I'll add the "only manual.mikrotik.com" as a another test case in bench-routeros-tools — tikoci.github.io – so we can test the theory more exactly