Replacing TensorZero or legacy OpenPipe: what to use, by job

What happened, with dates

• TensorZero: the GitHub repository was archived by its owner on June 12, 2026 and is read-only. The project site says it remains available on GitHub but is no longer maintained. Self-hosters can keep running it; there will be no fixes or updates.
• OpenPipe: announced on May 18, 2026 that the legacy fine-tuning platform stops supporting new training and inference on July 30, 2026, as workloads move to Weights & Biases (announcement). OpenPipe itself continues, refocused on reinforcement learning for agents. Existing customers must migrate before the cutoff.
• Related deadlines: OpenAI stops accepting new fine-tuning jobs on January 6, 2027 (deprecation record), and GitHub Models retires on July 30, 2026. If you are consolidating, the window for an unhurried migration is now.

First: when Apprentice is the wrong pick

Choose something else if your main need is one of these:

• An LLM gateway. TensorZero's gateway handled routing, retries, and observability across providers. Apprentice has no gateway. LiteLLM (open source) is the common self-hosted route.
• Hosted inference for your fine-tuned model. Apprentice trains locally (MLX on Apple silicon) or hands you an open-weights adapter to serve yourself, for example with vLLM. If you want someone else to host the model, look at Fireworks, Together, or W&B Inference.
• Open-ended chat or agent tasks. Apprentice is built for repeatable structured tasks, JSON extraction and classification first. Broad conversational fine-tuning is not the fit.

The jobs Apprentice does take over

• Turn production traffic into a verified dataset. Capture requests with one callback, verify rows into gold and silver tiers. This was the heart of the legacy OpenPipe workflow; your exported JSONL imports directly.
• Prompt optimization, scored honestly. DSPy GEPA rewrites the prompt against your verified rows and both versions are scored on a held-out slice. Measured on public data: GPT-4o-mini went 72.9 to 84.2 on receipt extraction. How it works.
• Fine-tune a small model you own. Local training on Apple silicon ships today (apprentice train --local, MLX). On the same public benchmark, a fine-tuned Qwen3.5-4B scored 89.2 against its optimized teacher's 84.2. Full write-up. Hosted training is being built.
• Eval-gated takeover. The small model replaces frontier traffic only after it passes your eval gate on gold rows, with instant rollback. This part is being built; until it ships, you decide manually with both scores in front of you.

Migrating from legacy OpenPipe before July 30

• Export your request logs and training datasets as JSONL from the OpenPipe console.
• Your fine-tuned OpenPipe models stop serving at the cutoff; the training data is the durable asset. Import it, verify a gold slice, and retrain on an open model you keep.
• The OpenAI migration guide covers the same verify-then-retrain path; the mechanics are identical once your data is exported.

Related

The receipt-extraction benchmark behind the numbers above, the prompt-optimization step, and migrating off OpenAI fine-tuning. Or see every guide.