Product case study · August 31, 2026

Agnes 2.5 Pro Alpha resurfaced in August, but its release was July 24

Agnes 2.5 Pro Alpha’s hosted API is deprecated. Check the migration advice, open-weight status and evidence gaps before choosing a new integration.

Reading time
6 min
Checked
Oct 23, 2026
Cut-paper archive tray with an unplugged cable, an open box of blue model blocks and an ochre path toward a server
A documented API and model card are not independently measured workflow performance
Bottom line

Do not start a new integration on the deprecated Alpha hosted API. Agnes recommends agnes-2.5-pro; Alpha weights remain available, but neither route has a measured Superbash result here.

Sources checked October 23, 2026. The important change since this article’s August publication is API lifecycle status: the official Alpha guide now labels the hosted API deprecated and recommends agnes-2.5-pro for new integrations. It says the open-source weights are unaffected. The guide gives Alpha’s release date as July 24, 2026; it does not give a shutdown date or prove that every Alpha request already fails.

This refresh reviews documentation and the existing benchmark record. It does not report an authenticated API test or a successful migration.

What the public materials verify

The exact public weights repository is Agnes-AI/Agnes-2.5-Pro-Alpha; its card identifies a multimodal reasoning model with a 1,048,576-token context window, a 65,536-token maximum output, text and image input, text output, tool calling, and streaming.[1] The repository includes an Apache License 2.0 text with Agnes AI and Alibaba Cloud/Qwen copyright notices.[2]

The historical hosted route uses agnes-2.5-pro-alpha at https://apihub.agnes-ai.com/v1. The guide retains Chat Completions, Responses and Messages examples for reference, but its deprecation notice takes precedence over those examples.[3] Do not read an old request snippet as a recommendation to deploy Alpha today. We have removed the former price row from this buying guidance because it is not a current quote for a supported replacement.

For local serving, the card’s SGLang quickstart specifies tensor parallelism of eight and recommends eight NVIDIA H200 GPUs (141 GB) or equivalent plus about 1 TB of fast NVMe.[1] That is a vendor operating recipe, not an independently tested minimum or a statement about the cost of a production deployment.

A metadata inconsistency worth keeping visible

There is a real provenance conflict in the released materials.[1][3] The model card, repository author, API guide, and license identify Agnes AI.[1][2][3] The released chat_template.jinja, checked again for this refresh, injects a system identity naming Sapiens AI.[4] We do not resolve that conflict by inference. Treat it as an unresolved identity/evaluator-metadata caveat until the publisher provides a clear explanation.

The card also reproduces Artificial Analysis evaluation tables and calls them independent measurements, while the current API guide says third-party metadata can retain older entries.[1][3] Those figures may be useful external signals, but they are neither this site’s fixed harness nor proof of a live API route’s behavior, latency, price, or tool reliability.

Why there is no Superbash score yet

The benchmark packet at benchmarks/agnes-2-5-pro-alpha/ pins three public Superbash task families and repeats each three times. Each accepted run must pass its task test, git diff --check, output-scope rules, a human prompt-rubric review, and a complete redacted metrics record. A positive coverage recommendation needs at least two accepted task families and seven accepted runs out of nine.

The August 31 benchmark record says the worker had no AGNES_API_KEY; the unauthenticated /v1/models probe returned HTTP 401, and no spend authorization was supplied. No authenticated model request was made in that recorded evaluation, and this refresh adds none. The retained packet describes historical Alpha access; it is not a migration test for Pro. The result is not run, not zero quality. No latency, usage, accepted-result cost, or independent speed comparison is claimed.

Audit an existing integration before migrating

This is a documentation-backed checklist, not a tested migration recipe. It requires no API request to complete the first two steps.

  1. Search your agent configuration, environment templates and request fixtures for the exact string agnes-2.5-pro-alpha. Record the base URL, endpoint, configured model and owner of each integration. Inspect files locally; keep secrets out of the review record.
  2. Separate hosted Alpha dependencies from downloaded Alpha weights. The deprecation notice concerns the hosted API; it does not revoke the published weights. Self-hosting still requires a separate hardware and operations decision.
  3. For hosted use, review the provider’s current documentation and account access for agnes-2.5-pro, the recommended replacement. Verify its current price and supported parameters rather than copying Alpha’s historical price or assuming identical limits. A model-name substitution alone does not establish compatibility.
  4. In a separate configuration, run one synthetic task with a fixed output cap and approved budget. Keep the previous configuration available for review. Check the returned model, usage, finish reason and task artifact. If your agent uses tools or images, test those paths separately before enabling real work.
  5. If access fails, record the status code and redacted error once, then check model entitlement and endpoint configuration. If output is truncated, inspect the finish reason and documented output budget. If a tool call is malformed, keep external actions disabled until its arguments pass your application’s validation.

A useful acceptance record contains the old and candidate model IDs, endpoint, checked documentation date, exact synthetic input, response metadata and a pass/fail against the same task rubric. Treat this as a new comparison: Alpha leaderboard figures cannot establish Pro’s behavior.

Buying decision

For a new hosted integration, follow the provider’s supported-model guidance rather than starting with deprecated Alpha.[3] For self-managed Alpha, the weights and license remain relevant, but the hardware recipe and quality claims still need validation on your workload.[1][2] Superbash has no measured result here for either an Alpha deployment or a Pro migration.

Use the hosting decision lesson to separate model quality, access and operating effort. Choose only after you can show an accepted task result and its actual cost.

Sources

Official sources checked October 23, 2026. The benchmark status above comes from the repository’s existing August 31 record, not a new model run.

[1] https://huggingface.co/Agnes-AI/Agnes-2.5-Pro-Alpha — Agnes 2.5 Pro Alpha model card [2] https://huggingface.co/Agnes-AI/Agnes-2.5-Pro-Alpha/raw/main/LICENSE — Agnes 2.5 Pro Alpha Apache License 2.0 [3] https://wiki.agnes-ai.com/en/docs/agnes-25-pro-alpha — Agnes 2.5 Pro Alpha API documentation

[4] https://huggingface.co/Agnes-AI/Agnes-2.5-Pro-Alpha/blob/main/chat_template.jinja — Released chat template and forced identity

Put this to work

Keep a model-card specification, vendor leaderboard, and matched workflow result as separate evidence classes.

Try

Audit model IDs in your agent configuration, then plan a bounded comparison on a supported route.

Prove it worked

Record route, response usage, elapsed time, retries, artifact acceptance, scope checks, and correction time for every run.

Where it can pay

Compare cost per accepted result, not a model-card price row or a ranking screenshot.

Keep in view

  • Agnes documents the exact API model ID as `agnes-2.5-pro-alpha` and its weights as `Agnes-AI/Agnes-2.5-Pro-Alpha`.
  • The official Alpha guide now marks its hosted API deprecated and recommends agnes-2.5-pro for new integrations.
  • The release date is July 24, 2026—not an August new-model launch—and Superbash has no authorized measured result yet.
Learn the workflow: choosing a model by the job