Qwen3.5 9B Beats Frontier Models After $500 RL Fine-Tune

Fermisense published a reinforcement-learning case study on July 27 showing a fine-tuned Qwen3.5-9B model outperforming five frontier models on an e-commerce catalog-review workflow.

The result is unusually interesting because the reported training bill was small and the model weights are public. It is also unusually easy to overstate. This is a specialist beating general models on the specialist’s own narrow task—not evidence that a 9B model is broadly smarter than the frontier.

The live chart, calculator, and pricing table above use current hosted API rates from our dataset. They are useful deployment proxies, but they are not Fermisense’s self-hosting estimate.

What happened

Fermisense built a simulated catalog-integrity environment from the Amazon Berkeley Objects dataset. Its agent had to inspect listings, use taxonomy and brand tools, extract attributes, and decide whether to approve or escalate each item.

The company created 177,767 training episodes and a 200-episode stratified validation set. It then compared Qwen3.5-9B trained with group relative policy optimization (GRPO) against GPT-5.5, GPT-5.6 Sol, Gemini 3.1 Pro, Claude Opus 4.8, and Claude Fable 5 under the same tools, images, scorer, and turn budget.

Source-reported result	Fine-tuned Qwen3.5-9B	Best frontier configuration
Share of achievable task score	87.3%	76.9%
Relative improvement	+13.5%	Baseline
Cost per 1,000 listing reviews	$0.50	$34
Training required	1,000 optimizer steps	None

These are Fermisense’s measurements, not independently reproduced AI Pricing Guru results. The cost is specific to the tested prompts, tool calls, token usage, and deployment assumptions.

Why the $500 training run matters

The full run used two rented RTX PRO 6000 GPUs: one generated rollouts while the other applied gradient updates through the open-source prime-rl framework. Fermisense says 1,000 steps took about 3.5 days and roughly $500 in GPU time.

More important than the final bill is the learning curve. The 9B model reportedly crossed the frontier performance band near step 250, after approximately one day. The remaining training improved the specialist further.

That changes the build-versus-buy calculation for frequent, scoreable work. A frontier API is still the fastest way to establish a baseline and collect traces. Once traffic is large and correct outcomes can be graded automatically, a compact specialist may repay its training cost through lower inference cost and fewer errors.

What the benchmark does—and does not—prove

The strongest evidence is the controlled comparison: identical workflow, tools, scorer, images, and turn budget. Optimized instructions helped the frontier models only modestly, while the fine-tuned model learned task conventions in its weights.

The main limitation is external validity. The scorer encodes Fermisense’s catalog policies, including a seven-times larger penalty for missed violations than false alarms. Change the policy, data distribution, or tool behavior and the ranking may change.

Reproducibility is also incomplete. Fermisense has published the full model and adapter weights on Hugging Face, but the derivative repositories had no README, model card, evaluation instructions, or license declaration when we checked. The Qwen3.5-9B base model is Apache-2.0; that does not automatically settle the licensing and documentation gap for the published derivative. The public write-up explains the environment and training setup, but an independent team has not yet reproduced the score.

Pricing impact: who wins and who loses

Fermisense reports its specialist at 40× lower cost than the cheapest frontier configuration it tested and about 340× lower than the most expensive. Against the strongest frontier baseline, it reports a 68× per-listing advantage.

Do not apply those multipliers directly to a token invoice. A real deployment must include GPUs or hosted inference, idle capacity, monitoring, retraining, evaluation, storage, engineering time, and frontier fallbacks. Compare cost per accepted decision, not cost per token alone.

High-volume teams with stable workflows, abundant examples, and machine-checkable outcomes stand to gain most. Frontier API vendors and general-purpose models lose routine traffic only after a specialist clears a task-specific holdout; teams with low volume, changing policies, or hard-to-score outcomes will usually spend more building the evaluator than they save on inference.

Teams that want to test an open-model route before operating GPUs can compare hosted open models on Novita, then benchmark them against Together AI pricing, OpenAI pricing, and Anthropic pricing.

Affiliate disclosure: we may earn a commission from the sponsored link above.

Labs availability

The catalog specialist is not in the AI Pricing Guru Labs leaderboard. Labs requires a stable callable model ID, a public endpoint, and an official token rate. The Hugging Face repositories publish weights and an adapter, but no hosted API or per-token price. Our canonical pricing files therefore keep the already verified Together-hosted Qwen3.5-9B base row and add no fabricated Fermisense SKU.

What developers should do

Start with a task that runs often and has a deterministic test, rubric, or agreed expert label. Measure a frontier baseline first, log corrections, and preserve a holdout set that training never sees.

Run a short RL pilot and check whether the specialist crosses the baseline early. Include the AI token calculator and infrastructure costs in the decision. Keep a frontier fallback for rare cases that require broad reasoning.

Our Together AI vs OpenAI cost comparison covers the hosted route. The recent World Model Optimizer analysis shows a complementary approach: route easy tasks to smaller models before paying to fine-tune one.

Bottom line

The Fermisense result is a strong demonstration of specialization economics: a small open model, a scoreable workflow, and a few days of reinforcement learning can beat much larger rented models on that exact job.

The next milestone is independent replication. Until then, use the reported result as a pilot design and cost target—not as a general Qwen-versus-frontier leaderboard.

Sources: Fermisense case study, public Qwen3.5-9B catalog weights, public adapter weights, the Qwen3.5-9B base model card, Together AI’s official model pricing, Hacker News discussion, and AI Pricing Guru’s live pricing dataset.