BackContext is live: one context layer for every agent you run.
Explore BackContextWe ran our four production models against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol on held-out tasks taken from our customers' own work. This is what we measured, how we measured it, and where the frontier models still win.
Four Applied Crew models are in production today: dfit-match-v3 and dfit-screen-v2 at DataFoldIT, and hytech-orders-v2 and hytech-catalog-v1 at Hytech Seeds. Each is an open-weight model in the 8B–32B parameter range, fine-tuned on that customer's verified outcomes. For every task, we ran our model and all four frontier models on the same held-out test set, with the same inputs, the same tools and the same scoring.
Every test set comes from verified outcomes recorded by BackContext. A matching task counts as correct when the candidate ranked first is the one the recruiter submitted and the client went on to interview. An order-extraction task counts as correct when every line item, quantity, variety and lot code matches what actually shipped. Nothing is labeled by a model, and nothing is relabeled by us after the fact.
Held out means held out. We split by time, not at random: the most recent weeks of verified records become the test set and never enter training. Random splits leak near-duplicates, such as the same distributor reordering the same seed variety, and inflate scores. A time split asks the harder question: does the model handle next month's work?
Averaged across held-out customer tasks:
Frontier cost is 4,000 input and 500 output tokens per task at list prices. The fine-tuned models are 46× cheaper per task than GPT-6 Astra and Fable 5.1, and they answer in under half a second.
Per successful task, the picture is the same. At 92.6% accuracy, GPT-6 Astra costs about $70.20 per 1,000 successful tasks. Opus 5.5 costs about $28.30 and GPT-6.1 Sol about $14.60. Our models cost about $1.44.
At DataFoldIT, dfit-match-v3 scored 98.1% on candidate-to-requisition matching, against 92.6% for GPT-6 Astra on the same held-out set, at 52× lower cost per task and 0.3s median latency. GPT-6 Astra's misses were rarely absurd. They were reasonable matches that broke the firm's own seniority and skills conventions. A "senior" role at DataFoldIT's clients means something specific, and 395K recorded recruiter and agent actions taught the model what.
At Hytech Seeds, hytech-orders-v2 reached 96.7% on order extraction, against 91.8% for Opus 5.5. Orders arrive as emails, PDFs and phone notes full of variety names, lot codes and regional abbreviations. Opus 5.5 reads them like a careful outsider. The fine-tuned model reads them like the order desk, because it learned from 206K order records verified against what shipped.
These are narrow, repeated tasks. On open-ended work, such as a new kind of request, a long chain of reasoning, or a document unlike anything in the training data, Opus 5.5 and Fable 5.1 are stronger, and we route to them. DataFoldIT's matching model falls back to a frontier model on 2.1% of tasks. That is by design: the fine-tuned model handles the volume, and the frontier model handles the exceptions.
Frontier list prices will keep falling. That narrows the cost gap. It does not close the accuracy gap on company-specific rules, because that gap comes from training data no general model has seen.
Averages are a starting point, not a promise. The only comparison that matters is on your data. Every engagement starts with the same benchmark on your held-out outcomes, side by side against all four frontier models. If the fine-tuned model does not win on your tasks, you should not switch, and the eval report will say so.
