Check if an open-weights agent model beats frontier APIs: Occamy-1.0

An open-weights agent model is cheaper than frontier APIs only when active-parameter serving cost, trajectory efficiency, and independently replicated benchmark scores survive your own harness. Occamy-1.0, an open-weights agentic model from Alibaba's Accio Team, is a useful test case because its routed-expert design changes the serving math. The trade-off is agentic inference cost per capability, not capability per parameter.

Episode cost is calls times tokens per call times price per token. The arithmetic is simple enough to run in a spreadsheet: expected calls, expected tokens per call, and your serving price per token. Long tasks can call the model many times, sometimes dozens or hundreds, so tiny per-call differences add up. The first question is whether the episode stays inside your GPU budget, not whether the model is smart.

The active slice, not the total, sets the bill

Total parameters are a storage and weight-download number. Active parameters are the compute number. The 35B MoE agentic model packs 256 experts, and for each token it routes to 8 experts plus one shared expert, so about 3B weights do the work. That gap is why the model can look expensive in a model card and cheap in a serving profile. If your stack can keep the active slice resident and route efficiently, the per-token arithmetic can approach a much smaller dense model. If it cannot, the total weight footprint becomes the constraint.

Run this production check before you commit:

• Active parameters per token. Pull the active count from the config or model card, then multiply by your expected tokens per task.
• Tensor-parallel and context footprint. Record the GPU count, parallelism width, and max context your serving stack actually uses.
• Trajectory efficiency. Log tokens, tool calls, and wall time per task, not just final success rate.
• Independent harness replication. Rerun at least one public agent benchmark under your own harness and price the GPU-hours per task.

Each item catches a different trap: MoE accounting, serving footprint, agent-loop waste, and benchmark inflation.

Trajectory efficiency is where agents eat the budget

Agent tasks run as loops, not single prompts. That makes trajectory efficiency the real pricing variable. Against its starting checkpoint, Occamy-1.0 used 19.5% fewer tokens per trajectory, made 15.2% fewer tool calls, and cut trace wall time by 46.4%. Those are not cosmetic deltas. In a long episode, fewer tool calls can mean fewer retries, fewer context rebuilds, and fewer idle GPU seconds.

The paper also says the model gained a lot of capability while per-task inference cost moved only a little relative to the starting checkpoint. The result should make an engineer lean in. But it is still a relative claim, and the base checkpoint matters. If your production baseline is a different model, the percentage may not transfer.

Replication is the only score that matters

The Accio Team reported Occamy-1.0 scoring 82.20 on Claw-Eval, a co-work benchmark, with an average above two frontier models and below the highest reported score of 83.92. The paper places the model at the low-cost knee of a cost-performance curve for co-work agent inference under its own evaluation and pricing setup. It is a strong claim. It is also a self-reported claim. No independent third-party replication had been confirmed at the time of the coverage.

Benchmark inflation is a known risk. Training-data contamination and leaderboard tuning can push scores upward without improving production behavior. The independent harness check belongs in the same decision as the price check. If you can rerun a public agent benchmark under your own protocol, you get a number you can defend. If you cannot, treat the vendor score as a lead, not a line item. A replicated score also tells you whether the model behaves under your tooling, your context limits, and your failure modes. The check is cheap to run and expensive to skip.

The first public preprint was submitted on September 4, 2026, which gives you a dated artifact to audit. Dressage, an open-source training framework, was released in June 2026, but training infrastructure is less relevant to a serving decision than the inference profile. Start with the dated artifact, then run the check on your own hardware.