Sarashina Free Beta: My First Test Would Preserve What Is Still Undecided
Checked October 16, 2026. This guide reviews official information and proposes an evaluation task. We did not run a Sarashina generation or translation benchmark.
A free AI trial is an invitation to choose a small test, not to upload the whole project. For MSL, the first question is whether a model preserves the difference between a decision and a possibility. Here is the exercise I would use before relying on Sarashina for Japanese announcements or bilingual publishing.
What was released?
SB Intuitions launched the free Sarashina beta on October 13, 2026. It offers web chat and translation, plus APIs for mini, nano, embedding, rerank, guard and translate. Google sign-in is supported without an advance application. Ledge.ai covered it on October 15. Those are different dates. See the official announcement and service page.
These are different kinds of tools: generating a draft, finding relevant material and checking language are separate jobs. I would select one publishing task before deciding whether an API integration is useful.
My first test: keep “undecided” undecided
MSL created the following fictional brief. It is neither client information nor a model response.
A small exhibition is planned for sometime in November. The exact date is undecided. Free admission is being considered but is not confirmed. The venue is an eight-minute walk from the station. Permission to use the photos is pending.
Use only the supplied brief to draft a short social announcement. Keep confirmed facts, plans, undecided items and pending permissions separate. Do not present the date or free admission as confirmed. Do not invent event names, addresses, dates or past results. Finish with a list of checks needed before publication. [Insert the fictional brief.]
| Check | Expected treatment |
|---|---|
| Date | November is a plan; no exact date is known. |
| Admission | Free entry is a proposal, not a decision. |
| Access | Keep the eight-minute walk unchanged. |
| Photos | Permission still needs checking. |
| Missing details | Do not invent names or achievements. |
This table is an answer key derived from our brief, not Sarashina output. I would mark factual changes separately from stylistic edits. A fluent paragraph can still fail the exercise if it turns a tentative plan into a promise.
A short example of what needs changing
MSL wrote these sentences to illustrate an editorial distinction. They are not Sarashina outputs or measured test results.
| Does not preserve the brief | A revision that preserves the conditions |
|---|---|
| Join our free exhibition this November. | We are preparing an exhibition for November. Dates and admission details will follow once confirmed. |
| We will share photographs from the venue. | We will decide whether to publish the photographs after checking permission. |
The first column turns a proposal and a pending permission into promises. The second is a suggested correction, not a finished announcement. Permission to publish the announcement itself remains a separate check.
Sort the draft into three decisions
- Keep: facts traceable to the supplied brief, such as the eight-minute walk.
- Rewrite: a detail present in the source whose uncertainty was lost. Restore the original condition.
- Hold: unsupported dates, prices or image permissions. Better prose cannot supply missing evidence.
Record the date, model, prompt, changed commitments, revised wording and unresolved questions. Zero corrections on this exercise does not establish accuracy on other material.
A bilingual check that stays manageable
For Japanese–English work, I would begin with this paragraph rather than a whole article. Place the source and translation beside each other. Mark numbers, names, tentative statements and negative statements. Ask whether “free admission is being considered” became a confident offer of free entry.
A second translation can help locate differences worth reviewing, but agreement between two outputs does not prove accuracy. Return to the source whenever the wording changes the commitment. Our translation workflow guide provides another place to start; the check here remains a proposed method.
Free does not mean unlimited or suitable for private material
The beta has no fixed announced end date, and usage data may be used to improve the service or models. I would use public or fictional material first and review the terms and privacy policy before introducing confidential work.
The official API limits, checked October 16, list six requests per minute for mini, guard and translate, and 60 for nano, embedding and rerank. Excess requests receive HTTP 429. These are API limits; do not assume the chat interface has the same allowance. An integration needs a deliberate waiting and retry strategy rather than a rapid loop.
Where I would adopt it—and where I would wait
I would include Sarashina in a small comparison using publishable sample text. I would record the model, date, prompt, factual corrections and review effort. The useful outcome is an announcement that preserves its conditions with manageable editing, not simply an attractive first draft.
I would defer automatic client publishing, integrations that depend on uninterrupted beta availability, and confidential processing until their requirements have been checked. This article does not establish a quality ranking, a time saving or a successful deployment. It supplies a repeatable question to investigate.
Sources and review date
- SB Intuitions launch announcement, October 13, 2026
- Official Sarashina service
- Official API documentation and rate limits
- Ledge.ai coverage, October 15; used for discovery
Sources checked October 16, 2026. Official facts and MSL proposals are distinguished above.
