MakeBoxExplore tools Jev can see.
Small questions.
Useful answers.
TypeSafe’s decision model can now work with images. Here is where it fits, what it costs, and why a bounding box is only the start of a good workflow.
Jev 1.14By TypeSafe AIExplore the visual demo ↗See the image. Ask a small question. Route the result.
Release: October 14 · This analysis: November 2
Sometimes you need
a decision.
Not an essay.
A photo arrives in an inbox. Before anyone writes a reply, a system needs to know what it shows, where the relevant detail is, and who should review it. That small decision is often the handoff between a messy input and useful software.
Jev is TypeSafe AI’s System One model. Its interface is built around typed questions. Noul returns a yes/no probability; Choice selects from supplied options; Score evaluates defined levels. Box adds a location in an image. Your code decides what happens next. Read the question types.
The current version is jev-1.14.0, released October 14. This is a November 2 assessment of image support, not a new launch announcement. Both 1.13 and 1.14 accept images; Box requires 1.14. Current models · Image documentation.
There is an important status distinction: Images and Box remain experimental. TypeSafe explicitly recommends them for prototypes and says they are not yet ready for production use. A polished video demo does not change that status. Read the release announcement.
From a picture to a next step.
SCROLL TO EXPLORE · SWIPE ON MOBILE
Give it
something to see.
A still image becomes part of the state your questions inspect. Start with a clear, upright input.
Ask for one
useful judgment.
Is the parcel visible? Where is the damaged corner? Keep each question focused.
Keep the next
action in code.
Choose a review path, collect better evidence or continue a low-risk process. A model answer is an input to that decision.
Decision.
Review.
Editorial illustration of a possible workflow. No Jev requests are executed on this page.
A rectangle tells you
where. Not whether.
Box returns one rectangle per question, with normalized edges and confidence. It has no “not found” answer. Pair it with a presence question and ignore a location when the target is absent. Official Box documentation.
A useful returns prototype could prepare a marked-up photo for a reviewer. It should not silently turn “a damaged-looking area is visible” into “approve a refund.” Product identity, order history and the actual policy still matter. This is our workflow recommendation, not a claim that Jev verifies a customer’s entire case.
And “80%” needs a label.
Choice confidence is a summary of its option distribution, not simply the selected option’s probability. For three options, a top probability of 0.80 gives confidence 0.70 under TypeSafe’s formula. Noul has no separate confidence field. How confidence is calculated.
Illustrative distribution, not a measured prediction. A threshold must be evaluated on your own task; a score is not a guarantee of correctness.
A small bill starts
with counted tokens.
TypeSafe lists $0.042 per million input tokens, with free output tokens. Images count as input tokens. There is no universal per-image price. Pricing · Image billing.
A video showing a hundred receipts processed for about a cent is a useful demonstration, not a quote for every receipt workflow. Image size, repetition, instructions and questions all influence billed input. Use the returned usage field to replace the assumptions in the calculator. Add the costs of preparation, storage, retries and human review when budgeting the whole system.
The speed claims
need their context.
TypeSafe advertises 193.6× faster and 444.6× cheaper from its workflow evaluations. Its launch article says these are likely toward the high end of real-world gains. They are not a promise that your image pipeline will achieve those ratios. Read the evaluation notes.
| Statement | What supports it | What remains to test |
|---|---|---|
| Very low token price | Published input-token tariff. | Your actual token usage and total workflow cost. |
| Large speedup | Vendor workflow evaluations and selected demos. | Your image sizes, concurrency, network and quality target. |
| Structured outputs | Defined answer types and response fields. | Whether the judgment matches reality. |
| Useful image localization | Experimental Box interface. | Accuracy on your targets, including absence and ambiguity. |
Type safety is valuable because software receives a predictable shape. It does not make every visual judgment true. A well-formed answer can still point to the wrong object, choose the wrong class or miss an important detail. Evaluate those errors before deciding which actions to automate.
Start with a bottleneck
you can actually measure.
These are workflow ideas, not eight ready-made MakeBox integrations. Choose one with a clear expected answer and an inexpensive way to catch mistakes.
Receipt triage
Separate document types and send unclear images to review. Keep accounting validation distinct.
Photo libraries
Sort shoot images by a defined set of categories. Sample-check ambiguous pictures.
Parcel review
Prepare a region of interest for a person inspecting a reported problem.
Delivery photos
Prototype checks for a visible parcel and usable framing before accepting the image.
Support screenshots
Identify a screen or issue category and route the ticket with its original context.
Video reframing
Use sampled still frames in a separate crop workflow. Test motion between samples.
Display checks
Compare shelf photos against a known layout; review uncertain or obscured slots.
Page sorting
Classify scanned pages before extraction. Routing a page is not reading every field.
A useful official example
TypeSafe’s OCR cookbook checks values that an OCR/LLM stage has already extracted. It compares each value with the page and a crop, then flags failures. That is a verification step—not evidence that Jev alone performs an entire bookkeeping workflow. Read the cookbook.
Prepare the image.
Define the decision.
Pin a model version when evaluating. Keep sample answers and reviewer labels. Measure mistakes as well as latency, then decide where a human should remain in the loop.
| Current documented constraint | Value |
|---|---|
| Image formats | PNG or JPEG; still images |
| File size | Up to 16 MiB per image |
| Dimensions | 280–1568 pixels on each side |
| Request body | 32 MiB; expanded-image limits also apply |
| Opt-in header | TypeSafe-Danger-Zone-Images: true |
Out-of-range images are rejected, not resized. Rotate the actual pixels upright; EXIF rotation is not applied. Prepare video frames separately. Full image limits and preparation rules.
Animated workflow illustration. No API connection, upload or model inference takes place here.
The opportunity is a
better handoff.
Jev’s interesting contribution is the boundary between a visual input and a small software decision. That can be useful even when a larger model, a specialist extractor or a person still does the rest of the job.
Start with a reversible task such as sorting a review queue. Test it on messy examples—not just the sharp, easy photos that make demos look good. Track missed cases, unnecessary escalations, token use and time saved. A cheap decision becomes valuable when the complete workflow behaves well.
Turn the idea into a working process.
Explore practical AI learning at MakeBox School or discuss a separately scoped automation project.
Independent editorial coverage. This article does not announce a Jev integration in MakeBox or a partnership with TypeSafe.
Sources & reading
- Jev 1.14 announcement — release date, experimental status and original artwork/chart.
- Models and pricing — model IDs and input-token tariff.
- Images documentation — formats, dimensions, limits, opt-in and billing.
- Box documentation — coordinates, one-box output and absence handling.
- Question types · Confidence — response semantics and calculation.
- System One launch and workflow evaluation notes — scope of the speed/cost claims.
- OCR confidence cookbook — separation of extraction and verification.
- TypeSafe AI — official brand mark. Logo and vendor graphics retain their attribution.
Prepared November 2, 2026. Inspired by the supplied video transcript; key product claims checked against TypeSafe’s own materials. MakeBox has not run a Jev benchmark. The parcel image is AI-generated; boxes and example distributions are editorial illustrations. Published November 2, 2026.