Through OrcaRouter you can reach the tuned build and its base on one key, which makes the Qwen3.8 Flash Next API worth reading for a specific row rather than the average. Of the five published scores, the one that predicts agent behaviour is Toolathlon Verified at 73.5 [VENDOR]. It sits between the ninety-ones and the sixty-twos, and it measures something neither of them does: whether the model can hold a plan together long enough to finish a task using other people's APIs.
Seventy-three and a half percent is a specific number with a specific meaning: roughly one task in four does not complete. For an agent, that is the difference between a tool that works and a tool that needs supervision.
What the score is really measuring
Toolathlon's tasks span many real APIs and require multi-step operation: authenticate, query, transform, write back. A model that scores 73.5 is doing several things right at once.
• It formats calls correctly often enough that the harness does not reject them.
• It reads responses and adjusts the next call rather than repeating the last one.
• It does not lose the objective across steps — the failure mode where an agent completes each call but drifts from the task.
The other 26.5% is distributed across those three. In practice, most agent failures come from the third: the calls are valid, the responses are read, and the model is no longer solving the original problem.
Why a refusal-free build has an advantage here
An agent loop is where refusal behaviour does the most damage, because a declined step does not just fail — it strands the plan. The model has a partial state, no result, and no instruction for what to do next.
A general-purpose model that declines an offensive-security tool call mid-chain produces exactly that: a stalled agent with no clean recovery path. The refusal-free build removes that failure class entirely, which is part of why an abliterated card is interesting for tool use and not only for chat.
It is not all upside. Removing refusals does not make a model better at planning, and the 73.5 is the same number either way. What changes is which failures remain.
The three integration details
The page's capability list is short and specific: tools, JSON, reasoning, vision [OURS]. Three details follow from it.
Function calling is native. You define the tools, the model emits calls. That is the standard shape, and it means an existing agent framework works without a translation layer.
Structured output is available. Ask for a schema and get JSON. For an agent that has to pass findings to a second stage, this is the difference between a parse and a regex.
Reasoning can be separated from the answer. Thinking and non-thinking modes let you trade latency for depth per request — useful when one step in a chain needs care and the others are mechanical.
What the page does not offer is a context window to match its ambitions: 262,144 tokens [OURS]. An agent that accumulates a long transcript, tool outputs and retrieved documents can reach that in a long session. Plan for compaction, or plan for a card with a wider window.
A four-step harness that predicts your number
Vendor task-completion figures rarely transfer. Build a small one instead.
1. Pick four tools you already call — a search, a database read, a file write, a message send.
2. Write ten tasks that each need at least three of them in sequence, with one task requiring a correction after a deliberate failure (a 404, an empty result).
3. Run them three times each. Agents are stochastic; a single run tells you about the sampler, not the model.
4. Score completion, not elegance. Did the final state match the objective?
Thirty runs is an afternoon and costs cents at this card's rates. The number you get will not equal 73.5, and that is fine — you want to know whether it is 90 for your tools or 40.
Include one task with a long tool output. That is where a 262,144-token window starts to matter, and where the compaction strategy you have not written yet will show up.
Where the score argues against the card
If your agent runs unsupervised overnight and a failed task is expensive, 73.5 is not enough on its own. Two mitigations make it workable:
Checkpoint the state between steps. A stalled chain should resume, not restart, and resuming needs the partial state written down.
Verify each tool result before proceeding. A cheap assertion after every call catches the drift before it compounds into a bad final state.
Both are engineering, not model selection. They are also cheaper than moving to a slower, more expensive card and finding that the failures simply take longer.
The agent loop nobody writes about
There is a fourth failure mode that no benchmark captures: token growth. Every tool result enters the context, and a long-running agent accumulates them. On a million-token window, that is a background process. On a quarter-million window, it is a countdown.
The mitigation is a summarisation step every N turns, which costs an extra call and loses detail. Both are worth budgeting before you build, because retrofitting compaction into an agent that has already been tuned on full history is a rewrite.
The tool schema is part of the score
One detail is easy to miss when reading a tool-use benchmark: the harness defines the tools, and the schema it uses is not yours.
A model that scores 73.5 on Toolathlon has been measured against a particular set of tool descriptions, with particular parameter names, at a particular level of detail. Change those and the number moves — sometimes by ten points, because the model has to infer from the description what the parameter means.
Two practical consequences. First, invest in the descriptions: a tool described precisely is worth more than a clever prompt. Second, treat the benchmark number as a prior about the model's planning, not a prediction about your schema. The only way to know the second is to run your own tools.
The takeaway
Seventy-three point five on Toolathlon means three tasks in four complete, and the missing quarter is mostly plan drift rather than bad calls. For an agent that verifies its steps, that is workable; for one that does not, it is not.
The card brings native function calling, JSON output, separable reasoning and no refusals — a combination that removes an entire failure class from agent loops. It brings a quarter-million-token window, which is a tighter budget than an accumulating transcript would like.
Sourcing note: the Toolathlon figure and the other four benchmark rows are OrcaCyber's own claims, evaluated 2026-08-25 and not independently reproduced [VENDOR]. Capabilities, context window and pricing are from the OrcaRouter model page, checked 2026-10-27 [OURS]. The four-step harness is a suggestion for the reader's own evaluation, not a measurement.
