A 9-billion-parameter open model, reinforcement-trained for about $500 of rented GPU time, outscored every frontier configuration a European consultancy tested on its own product-catalog review workflow. In an article published July 27, Fermisense reports that its trained specialist reached 87.3% of the achievable score on the task against 76.9% for the best frontier setup, at a cost of $0.50 per thousand listings versus $34 for the strongest commercial model.
Key facts
- The headline number: 87.3% of the achievable score versus 76.9% for the best of five frontier configurations, and roughly 68 times cheaper per reviewed listing.
- The training cost: two rented RTX PRO 6000 GPUs, 1,000 optimizer steps, about three and a half days, roughly $500 in GPU time -- with the frontier band crossed after about 250 steps.
- Who and when: Fermisense, July 27, 2026, in a write-up by Justinas Zaliaduonis, Joris Zilinskis, Fabian Hildesheim, Joel Hainzl and Gediminas Pazera.
- Primary source: The Rise of Intelligence Ownership.
The task is deliberately mundane, which is the point. An agent receives an e-commerce listing, searches a product taxonomy, checks whether the claimed brand is registered and protected, retrieves the attribute schema, and then either commits a structured decision or escalates to a human when evidence is thin. Fermisense built a scored simulator of that workflow -- listings flow in, the agent uses the same tools an analyst would, and a scorer grades every decision -- then trained the model against it with GRPO, a reinforcement-learning method that compares batches of attempts against each other rather than needing a separate value model. The infrastructure was the open-source prime-rl framework, and the released adapter names its base as Qwen3.5-9B.
Before training anything, they measured the ceiling. Five frontier models -- GPT-5.5, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Opus 4.8 and Claude Fable 5 -- ran 200 stratified validation episodes with identical tools, images, scorer and turn budget, each in two configurations: a plain prompt, and 2,800 characters of carefully tuned extraction conventions, lookup procedures and worked examples. The optimized versions all converged within a tenth of a point of each other. One of the strongest zero-shot models actually got worse with the extra instructions. And the instructions were not free: they inflated input-token bills by 28-55% on every single call, forever.
That is the sentence in the piece worth keeping: "Prompted task knowledge is rented per call; trained task knowledge is bought once and lives in the weights." A prompt is a briefing you re-read aloud to a contractor at the start of every job. Fine-tuning is hiring someone who already worked there for a year. The contractor may be smarter in general; the employee knows that this particular brand code means something odd, and that a specific product category is where the policy violations hide.
The reward design is where the local knowledge gets encoded. Fermisense weighted the score across category accuracy, attribute extraction and policy judgment, and made missed violations far more costly than false alarms -- a choice that encodes their client's risk appetite, not a universal truth. This is the pattern behind every credible version of this story, and it matches what our explainer on reinforcement learning with verifiable rewards describes: the model does not need to be generally better, it needs an automatic grader that reflects the actual job. Fermisense cites other companies converging on the same playbook, including a legal-AI firm whose reinforcement-trained open model beat two frontier models on its own rubrics, and a support platform that post-trained a vertical model to handle high-volume ticket resolution more cheaply.
At scale the arithmetic becomes the argument. At roughly 40 million decisions a day -- the volume Fermisense cites for large-catalog operations -- $0.50 per thousand versus $34 per thousand is the difference between about $7 million a year and about $500 million.
Now the caveats, which are real and which Fermisense partly names itself. The $500 buys the final training run and nothing else; the simulator, taxonomy, policy rubric, episode generator and scorer represent the actual investment, and none of them is published. The frontier models were graded by Fermisense's scorer against Fermisense's policy preferences -- exactly the asymmetry that makes a specialist look good -- and the article does not identify which frontier configuration was the best baseline. Neither Hugging Face release carries a model card, and the training recipe, prompts, validation split and evaluation harness are absent, so nobody outside the company can reproduce any of it. Read as "a small open-weight model now beats the frontier," this is unsupported. Read as "if you own a repeatable workflow and can score it automatically, buying the task knowledge is cheap," it is one of the better-documented data points of the year. The scorer is the moat, not the model.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)