If your experiment tracker belongs to a company that just got bought, the tracker is easy to replace and the run history is the part you can lose. Neptune's hosted users found that out in March 2026: after OpenAI agreed to acquire Neptune, Neptune's transition hub set the hosted shutdown for March 5 and said remaining data would be irreversibly deleted. Run MLflow (Apache 2.0) on a database and bucket you own, managed on SageMaker or Databricks if you already pay for either, on Postgres if you don't. Stay on Weights & Biases if your researchers live in its UI, but put artefacts in your own bucket and export history on a schedule. Comet loses: at our 10-person reference workload, its training-hour meter makes it the priciest.
| Option | Owner | Licence | Price shape (checked October 19, 2026) | Reference workload, per month | Do not pick it if |
|---|---|---|---|---|---|
| MLflow, self-hosted | You (open-source project) | Apache 2.0 | Free; you pay for Postgres, object storage and a server | Your infra bill plus the engineer who runs it | Nobody on the team will own an upgrade or a backup |
| MLflow, managed (SageMaker serverless, Databricks) | AWS / Databricks | Apache 2.0 client and data model | SageMaker: no additional charge, service limits apply. Databricks: no separate MLflow price | Close to $0 on SageMaker beyond storage | You are not already on that cloud |
| Weights & Biases (now inside CoreWeave Forge) | CoreWeave | Proprietary; self-managed needs a licence | Pro from $60/month, up to 10 seats, 100 GB then $0.03/GB; Enterprise is contact sales | From about $66 on Pro, if you qualify | You have 50+ employees and won't negotiate Enterprise |
| Comet | Comet (venture-backed) | Proprietary | Pro $19/user/month, 1,500 training hours then $1/hour, 500 GB then $3/100 GB | About $690 | Your runs are long and many |
| ClearML | ClearML | Server under SSPL, SDK Apache 2.0 | Pro $15/user/month plus usage, 120 GB then $0.10/GB | About $168 plus metric and API usage | Your legal team treats SSPL as not open source |
| ZenML, Lightning AI | ZenML / Lightning AI | ZenML core Apache 2.0 | ZenML Scale $999/month per workspace; Lightning tracker priced with its platform | Not a like-for-like tracker | You only need a tracker |
The reference workload is a team of 10 ML engineers logging 2,000 training hours and holding 300 GB of artefacts a month. Prices come from each vendor's live pricing page on October 19, 2026, linked in the sections below.
What Happened to Neptune, and Why It Is the Case Study
Neptune is the clearest recent example of an acquired tracker's hosted service closing with a fixed deletion date. InfoWorld reported on December 4, 2025 that OpenAI had agreed to buy Neptune, had been a customer for more than a year, and would keep the platform for internal use. Neptune withdrew its products from the market. Hosted users got a few months of security fixes and no new features.
The dates moved during the window. InfoWorld quoted an initial switch-off of March 4, 2026 at 10 a.m. PST; Neptune's transition hub later gave March 5, 2026 at 10 a.m. PST, with self-hosted customers' Helm and container image repositories deleted on March 8. Neptune's final service shutdown notice states that data stored in the hosted service was permanently deleted and cannot be recovered.
Three months was enough for teams that started in December and too short for teams that assumed someone else had it in hand. The run history (the logged metrics, parameters, artefacts and lineage behind every model you trained) was what disappeared.
The exporter Neptune shipped shows exactly what survives a forced move. Neptune Exporter writes runs to Parquet plus files, and can load them into MLflow, W&B, Comet, ZenML, Lightning AI and others. Its README lists the losses: MLflow and W&B accept only integer steps, so decimal steps need a multiplier applied identically on every load; string series and histograms become artefacts in MLflow; fork relationships are mapped best-effort or dropped; attribute names get sanitised and truncated at 250 characters. Model registry export worked only for Neptune 2.x. Plan for that much loss in any tracker migration.
Who Owns Each Tracker Now?
Of the trackers on this page, only MLflow lacks a single company that can switch the hosted service off. That ownership is the risk you are underwriting when you pick one.
CoreWeave closed its acquisition of Weights & Biases on May 5, 2025, and since September 30, 2026 Free, Pro and Enterprise SaaS accounts sign in through CoreWeave Forge. CoreWeave says accounts, projects, teams and API keys stay the same, and that self-hosted and dedicated-instance accounts are unchanged. The wandb SDK and WANDB_API_KEY did not change either, per CoreWeave's docs migration note. The same note says the public wandb/docs repository is archived and documentation now lives in a private coreweave/docs repo that takes no community contributions. Nothing there signals a shutdown. It does show the product is now one module of a GPU cloud's platform, which we covered in GPU Clouds Compared: Nebius on Price, CoreWeave at 3AM.
MLflow is Apache 2.0, and the hosted versions are run by clouds you already have contracts with. Comet and ClearML are independent companies. Comet has raised venture funding, including a $50 million Series B, and an independent venture-backed tool is exactly the profile Neptune had the week before its deal.
MLflow: The Default, Managed or Self-Hosted
MLflow is the right default because every hosted version and your own server share one open data model, so leaving any of them is an export between MLflow servers. That is the property Neptune's customers did not have.
Managed MLflow is now close to free on AWS. Amazon announced on December 2, 2025 that serverless MLflow on SageMaker AI "is offered at no additional cost," with service limits that the post does not enumerate. AWS also published a migration guide from self-managed tracking servers built on the open-source export tool. Databricks describes its version as "a fully managed and hosted version of MLflow" with customer-managed key encryption that open-source MLflow lacks; its docs list no separate MLflow price, so you pay for the Databricks Mosaic AI compute that logs to it.
Self-hosting is a database, a bucket and a process. MLflow's self-hosting guide recommends moving the backend store to PostgreSQL and artefacts to S3, GCS or Azure Blob as you scale, and ships basic auth, SSO, a Docker Compose bundle and an official Helm chart. The failure mode practitioners hit is the default file store. MLflow's own troubleshooting page names the file-based backend, which has no indexing, as the most common cause of a slow UI, and a 2022 GitHub issue documented the run table slowing as rows grow. Start on Postgres on day one.
Moving between MLflow servers uses mlflow-export-import, which copies experiments, runs, registered models, traces and prompts and keeps references intact in bulk mode. Two caveats from its README: the PyPI package is marked legacy (install from GitHub), and bulk mode cannot rename objects at the destination.
Who should not pick MLflow: a research team that compares hundreds of curves a day and will judge the tracker by its charts. W&B's UI is better for that work, and forcing MLflow on those users produces a shadow W&B account within a quarter.
Weights & Biases: Keep It, With Two Conditions
W&B is still the strongest product for interactive experiment analysis, and the conditions for keeping it are that artefacts sit in your bucket and history gets exported on a schedule.
Pricing moved to CoreWeave's Forge pricing page, where W&B's old URL now redirects. As checked on October 19, 2026, Pro "starts at $60/month, billed monthly," with up to 10 model seats, 100 GB of storage a month and additional storage at $0.03/GB. The page does not say whether $60 is per seat. Pro is "for early-stage teams fewer than 50 employees," and customers above that "will be required to transition to CoreWeave Forge Enterprise," which is contact sales. Most readers of this page are Enterprise buyers by that rule. Bring-your-own-bucket is listed as an Enterprise feature, and it is the one to insist on.
Self-managing W&B does not remove the vendor. The self-managed requirements say "you must obtain a license before installation," and require Kubernetes, MySQL, Redis and S3-compatible storage. On that path you own the data, and the software still runs on a licence the vendor can decline to renew.
The export trap is documented and easy to miss. In the W&B public API guide, run.history() returns 500 sampled points by default, and only run.scan_history() returns the complete, unsampled history. An export script written with history() looks like it worked and keeps a fraction of your curves.
Who should not pick W&B: a company over 50 employees that wants published pricing, or a regulated team that cannot accept a licence-gated self-managed install.
Comet: The Meter That Costs Most
Comet is a capable tracker priced on training hours, and at our reference workload that meter makes it the most expensive hosted option. Comet's pricing page lists Pro at $19 per user per month for up to 10 users, 1,500 training hours included and then $1 per training hour, and 500 GB of storage included and then $3 per 100 GB per month. Enterprise is contact sales, and on-premise deployment is Enterprise-only.
At 10 users and 2,000 training hours, that is $190 in seats plus $500 in overage, about $690 a month, before storage passes 500 GB. Double the training hours and the bill passes $2,000. The seat price is low; the hour meter grows with the very activity you bought the tool for.
Portability is the second problem. Comet's cometx toolkit copies experiments from one Comet installation or workspace to another and can download them to local files, but documents no path into another tracker. The Comet-for-MLflow bridge runs the other direction, importing MLflow runs into Comet. Leaving means writing your own loader for those files or your own extractor against Comet's API.
Who should not pick Comet: anyone running many long jobs, or anyone choosing a tracker after Neptune specifically to reduce exit cost. Its LLM evaluation product, Opik, is a separate decision.
ClearML: Cheap, With a Licence Footnote
ClearML is the cheapest per-seat hosted option and bundles orchestration, but its server is SSPL, which many legal teams do not count as open source. ClearML's pricing lists Pro at $15 per user per month plus usage for teams up to 10, with 120 GB of artefact storage, 1.2 GB of metric events and 1.2 million API calls a month included, then $0.10 per GB of storage, $0.01 per MB of metric events and $1 per 100,000 API calls. At our workload, seats plus 180 GB of extra storage come to about $168 a month before metric and API usage.
The ClearML server repository states it is under the Server Side Public License v1.0 and relies on MongoDB and Elasticsearch. You can run it yourself for free. An SSPL dependency still goes through legal review at many companies, and a MongoDB-plus-Elasticsearch backend is heavier to operate and back up than MLflow's single Postgres database.
Who should not pick ClearML: a team that wants only a tracker, or one whose open-source policy excludes SSPL.
ZenML and Lightning: Orchestrators That Delegate Tracking
ZenML and Lightning AI were both Neptune migration targets, but ZenML delegates experiment tracking to another tool, and Lightning's tracker ties you to its platform. ZenML's experiment tracker components are integrations with MLflow, W&B, Comet, Neptune and Trackio, so choosing ZenML still means choosing one of the trackers above. ZenML's pricing lists the open-source core free and Scale at $999 a month per workspace with 2,000 runs by default. Buy it if you need pipelines.
Lightning's path is LitLogger, which logs to Lightning's Experiments dashboard using Lightning.ai credentials; Neptune's Lightning migration guide creates one Lightning teamspace per Neptune project. That is a reasonable home if your training already runs in Lightning Studios. Otherwise you are moving your history into another single-vendor platform, which is the situation this page is about.
What Predicts Regret
The decision that teams regret is the one made on charts and seat price while the history export went unexamined. These are the criteria that predict it, in order:
- Can you get full-fidelity history out without the vendor's help? MLflow: database dump plus bucket. W&B: yes, with
scan_history()and artefact downloads. Comet and ClearML: through their APIs, with code you write. - Where do artefacts physically live? Your bucket survives any acquisition. A vendor's bucket survives only as long as the vendor's hosted service.
- What grows with use? Comet meters hours, ClearML meters events and calls, W&B meters storage. Model the meter at twice this year's volume.
- Who owns the vendor, and could they want the product for internal use? That was OpenAI's stated plan for Neptune.
The answer changes in three cases. If your team runs on Databricks already, use its managed MLflow and stop there. If your researchers will revolt without W&B, keep W&B on Enterprise with your own bucket. If you need pipelines, scheduling and data versioning in one product, ClearML's bundle is worth the SSPL review.
For the related problem of keeping model weights out of a single vendor's hands, see Hugging Face Hired Bankers. Go Mirror Your Weights.
Schedule the Export, Whatever You Pick
The tracker is replaceable in a sprint; years of run history are not, so export it weekly to storage you own. Neptune's exporter existed, worked, and still left teams without their data because nobody ran it before the deadline.
For W&B, a weekly job is a few lines against the public API:
import json, os, wandb, pandas as pd
api = wandb.Api()
os.makedirs("export", exist_ok=True)
for run in api.runs("my-team/my-project"):
pd.DataFrame(list(run.scan_history())).to_parquet(f"export/{run.id}.parquet") # never run.history(): it samples 500 points
with open(f"export/{run.id}.json", "w") as f:
json.dump({"config": run.config, "summary": run.summary._json_dict}, f, default=str)
For self-hosted MLflow, the export is a nightly pg_dump of the backend database plus versioning on the artefact bucket. For managed MLflow, run mlflow-export-import against a second server or bucket on a schedule. For Comet and ClearML, write the extractor now while you have time to test it.
This Week: Find out where your tracker stores artefacts and who owns the vendor. Run one full-fidelity export of your most important project and load it somewhere you control.
This Month: Price your actual meter (hours, events or storage) at twice current volume. If you are on W&B Pro with more than 50 employees, open the Enterprise conversation and ask for bring-your-own-bucket in the order form.
Before Q1 Close: Put the export on a schedule with an alert when it fails, and add a change-of-control and data-return clause to the tracker contract at renewal, the same clause we argued for in Prime Intellect Buys Moondream and Shuts Its Cloud on Nov 13.
The Bottom Line
Experiment tracking has followed the path of CI servers and log platforms: the tools converge on similar features, and the switching cost settles into the data you have accumulated. MLflow on your own Postgres and bucket keeps that data in a format several clouds already host. If you choose anything else, the weekly export is what makes the choice reversible.
Run the first export before you read the next acquisition headline.
Continue Reading
- Best MLOps for Regulated AI: Domino, Then a Sign-Off Layer
- GPU Clouds Compared: Nebius on Price, CoreWeave at 3AM
- Hugging Face Hired Bankers. Go Mirror Your Weights.
- Databricks Alternatives: Move Serving, Not the Lakehouse
- Cloudflare's Deno Deal Gives Deno Deploy Six Months to Live
- Model Deprecation: Bedrock Keeps Claude 4 Months Past Anthropic