DEV Community: Titus Brett

Hyper3D Rodin Review 2026: The Free PBR Price Is Unverified

Titus Brett — Mon, 12 Oct 2026 15:48:02 +0000

This is an independent test. Nothing here came from vendor access: both tools were used on my own free and paid accounts.

I saw a knitted vase on Hyper3D's screen that I actually wanted to keep. The Gen-2.5 preview had the big crossing strands from my source picture. The page displayed 0.5 Credits. Then I tried to Download the confirmed GLB on the Free account and learned those strands were behind a subscription. To obtain an actual asset I went through a second, Gen-1.5 workflow. The smaller file came home, but its ribs no longer looked like that promising preview. I kept a ledger precisely so those two jobs would not be mistaken for one.

My verdict in 60 seconds — SupaVoxel gave me a traceable textured knit-design deliverable; Hyper3D's more faithful Gen-2.5 preview was not downloadable on the tested Free account. Hyper3D Gen-1.5 did export a smaller, cleaner-shell PBR vase, but its complete material-stage debit was not isolated. The SupaVoxel Hobbyist request linked one −3 transaction in its own credits to a textured file with raised cable crossings. Those are different plans and different units: no cash-price comparison follows. For an editable digital knit concept I would start in SupaVoxel, then repair its mesh before print work.

Eight side-by-side checks, Hyper3D then SupaVoxel:

Access to the best-looking observed result — Hyper3D Gen-2.5 was blocked at Free Download; the tested SupaVoxel result exported.
Attributable complete-PBR debit — Hyper3D material cost remains unknown; SupaVoxel linked one finished request to −3 of its own credits.
Verified charge scope — Hyper3D Gen-1.5 geometry Confirm used 0.5 of its own credits; SupaVoxel's full completed request used three of its own.
Material workflow — Hyper3D needed another Generate and Confirm; the completed SupaVoxel request delivered material.
Exported cable geometry — Hyper3D Gen-1.5 ribs versus raised SupaVoxel crossing strands.
Actual browser download — Hyper3D 15.87 MB ZIP versus SupaVoxel 35.05 MB Original size GLB.
Starting solid integrity — Hyper3D one welded watertight shell; SupaVoxel 744 non-manifold edges.
Repeat-job budgeting — Hyper3D's full-PBR debit and material duration remain blank; SupaVoxel offers one request-linked debit and lifecycle, not a production guarantee.

The only credit arithmetic supported here stays within each service's own stage and units. It cannot say a complete Hyper3D PBR costs 0.5 or that SupaVoxel is cheaper in dollars.

What was submitted—and how many jobs did it create?

One identical 1,746,019-byte knitted-vase PNG, not a dimensioned vessel design or a photograph of the unseen back.

Both products received the SHA-matched source image. Hyper3D then ran two separate tasks: Gen-2.5 for a screen preview and Gen-1.5 for a downloadable file. SupaVoxel ran one request on Hobbyist. This is a real consumer access test, not a same-stage GPU benchmark. Hyper3D's two automatic captions differed; Gen-2.5's omitted “knitted,” whereas Gen-1.5's mentioned it, so visual differences cannot be ascribed only to the version. The image was AI-generated with a third-party design cover used as visual inspiration. No rights over that underlying reference are established here. The offline comparison below shows what the two exported files looked like; the blocked Gen-2.5 preview is not substituted for Hyper3D's Gen-1.5 deliverable.

Hyper3D's downloadable Gen-1.5 textured PBR vase: smaller transfer and cleaner starting solid, but the source's large crossings simplify into ribs.

Did Hyper3D give me the Gen-2.5 vase it showed me?

Gen-2.5 visibly suggests bigger crossings; the interface reports a million faces, not an audited downloaded file.

After .glb selection, the Free-tier Download request demanded a subscription and said Gen-1.5 remains free to download.

No. The Gen-2.5 panel reported 1,000,000 faces and 864,347 vertices after geometry confirmation. The first screenshot showing a completed preview was captured no later than 140.6 seconds after task creation; that is a screenshot upper bound, not exact generation duration. Geometry confirmation took 17.7 seconds on its separate recorded interval. Then Download refused the file. We bought no subscription and obtained zero Gen-2.5 GLBs. Its visual promise is real as a preview, but its watertightness, exact material and downloadable byte size are unknown. I cannot borrow the Gen-1.5 file's measured numbers to certify a different gated mesh.

Why isn't the visible 0.5-credit tariff a bill for this task?

The Gen-2.5 UI says 0.5 Credits; that is a tariff for the selected control, not an isolated task debit for this vase.

The Hyper3D account was shared with another active case. Its header moved 5.0 → 4.5 between generation and the first finished preview; a later 4.0 snapshot also reflects account-wide activity. Neither delta can be uniquely attributed to this vase. The tariff and pilot charging rule are evidence of what the interface advertises, not of what this specific Gen-2.5 task alone deducted. Most damaging to any cheap-vase headline: even if the UI tariff is accurate in ordinary use, it did not buy a Free-tier downloadable Gen-2.5 object. A preview expense and a completed deliverable have different denominators.

Which half-credit can I actually isolate?

Gen-1.5 Geometry Confirm is the step bracketed by the observed 3.5 → 3.0 shared-header movement while this task was tracked.

A separate Hyper3D Gen-1.5 preview held the header at 3.5 before and after viewing. After its Geometry Confirm the header showed 3.0, isolating 0.5 Hyper3D credit for that geometry-confirm step in this case. Do not assign an earlier 4.0 → 3.5 account change to the vase; another case shared the login. Gen-1.5 task creation to first completed-preview screenshot had an upper bound of 64.8 seconds; confirm submission to its first visibly unlocked Pack state had an upper bound of 57.7 seconds. Neither is a complete material or export clock. The downloaded final mesh has 60,000 faces, not the blocked Gen-2.5 preview's million.

Was the first downloaded file a textured vase?

The first Gen-1.5 export is a 2,268,860-byte geometry GLB with UV coordinates and no embedded textures.

No. Geometry Confirm unlocked a 2.27 MB GLB with zero texture images. Its UVs are coordinates where paint could go; they do not make the file a textured product. To obtain the white knit-looking PBR asset I performed Material Generate and Material Confirm, then selected .glb and downloaded the archive. Those are four successful actions after geometry Confirm before ZIP extraction. SupaVoxel's observed path after completion was three menu selections: Export, GLB, Original size. Click totals are evidence of handling steps, not measured labor, reliability or pricing. I also cannot pretend Gen-1.5's material-painted small chevrons were the large geometric weave seen in the blocked Gen-2.5 preview.

What did the material stage cost?

Separate Material Generate is in progress: geometry's known 0.5-credit debit is not yet the complete textured-file price.

Material Confirm completed, but no independent per-stage ledger ties an exact material debit to this task.

The shared header showed 3.0 after geometry, during material preview and after confirmation. That sequence does not prove a zero charge for every material action: the account was shared and no authenticated task-specific material transaction was preserved. Nor does another case's material charge establish this vase's price. The Hyper3D full PBR credit cost remains unknown. A 0.326-second recorded material-confirm server interval does not include Material Generate and cannot be promoted to the time to deliver a textured object. If I were booking a hundred product assets, I would first repeat this one job on an isolated account, logging every stage. Otherwise the financial model rests on a missing receipt.

Can SupaVoxel identify the complete request's charge?

One uploaded image and four unused extra-angle slots; the action displays Generate (3 credits).

The saved SupaVoxel request's image hash matches the source and has exactly one linked −3 transaction. That is three SupaVoxel credits for this particular successful textured request, not three Hyper3D credits, and its 260-before/257-after shared balance display is not my proof when other cases used the account. The database lifecycle from request creation to update was 163.157 seconds, including waiting but excluding upload before and export after. No duplicate Generate click was recorded. It gives a real billing and timing anchor for one deliverable; it does not certify that its 744 non-manifold edges need no paid or manual repair before physical printing.

How large was the file that came home?

Hyper3D selected 2K PBR and Shaded, and the browser received both GLBs inside one ZIP.

Hyper3D delivered a 15,869,568-byte ZIP. Extracted: a 10,496,516-byte PBR GLB plus a 5,372,800-byte Shaded GLB. SupaVoxel delivered one 35,051,872-byte Original size GLB through its authenticated export; the exact browser Blob was assembled without geometry conversion. Website transfer is therefore 15.87 MB versus 35.05 MB, a legitimate Hyper3D win in these chosen branches. If hosting PBR alone, Hyper3D's member is 10.50 MB, not 15.87. SupaVoxel also offered Compressed in the menu, but it was not downloaded; I cannot imply all its outputs weigh 35.05 MB.

What happens if I multiply this by a hundred?

Original size was chosen; Compressed exists but has no measured size here.

A hundred repeats of the verified Hyper3D geometry Confirm alone would consume 50 Hyper3D credits, hypothetically, while one hundred repeated successful SupaVoxel requests would consume 300 of its own credits under this sample's price. Those figures cannot be divided into a cross-brand price ratio; Hyper3D's complete PBR cost is unknown, and neither 100-job batch was run. Its five-credit Free balance snapshot also does not reveal renewal or a right to make a hundred files. Repeating SupaVoxel's 163.157-second database lifecycle a hundred times serially sums to 4.53 hours, excluding upload, export, repairs and possible failed jobs. No corresponding Hyper3D full-file duration is known. That is planning arithmetic, not measured throughput.

Is the material cost the only missing cost?

SupaVoxel's downloaded file retains bigger knit crossings; its 744 non-manifold edges are an unpriced print-repair obligation.

No. If the deliverable is a digital turntable, SupaVoxel's raised cables are the stronger starting form. If the deliverable is a functional printed flower vase, the SupaVoxel welded mesh fails watertightness and has 497 face-edge-connected regions, though those form one vertex-connected cluster. Repair was not timed or run. Hyper3D's cleaner closed starting shell has largely straight ribs where the reference needed crossings, and sculpting that relief was also not timed. Both modeled cavities are shallow: Hyper3D 32.79–35.13 mm, SupaVoxel 22.8–24.0 mm with each longest bbox side scaled to 120 mm. Generation credits alone cannot price an approved, waterholding object when the main manual work was never measured.

Final verdict: what cost do we truly know?

The visible Gen-2.5 0.5 is a Hyper3D UI tariff, not a case-isolated debit, and the Free Download was blocked. The separate Gen-1.5 geometry Confirm demonstrably cost 0.5 Hyper3D credit; material stages were not independently reconciled, leaving the complete PBR total unknown. SupaVoxel's one successful request has an attributable −3 of its own credits and an actual textured GLB. Hyper3D wins on the tested browser download size and initial closed-shell topology. SupaVoxel wins downloadable cable geometry and billing traceability for this one design. My workflow preference favors the SupaVoxel design reference, not a claim that it is cheaper in dollars or ready to print.

Use SupaVoxel for the knitted concept—not an unpriced print

The SupaVoxel export is where I would start if a brief requires thick over-and-under knit in an editable digital model. I would budget a separate topology repair and an interior redesign before offering it as a water-holding print. If smaller bytes and a single clean starting shell are the brief instead, Hyper3D Gen-1.5 is a sensible alternate despite its simplified ribs. To estimate actual unit economics, run Hyper3D's whole material path in isolation and time repair on both; this case did neither.

How I separated evidence from arithmetic

The identical SHA-checked PNG went into both services. Hyper3D's tariff was read from its UI; the Gen-2.5 shared-account delta was explicitly not attributed to this task. Gen-1.5's 0.5 was isolated at Geometry Confirm only. SupaVoxel's −3 came from a request-linked transaction and its 163.157 seconds from one request row, not an end-to-end stopwatch. File sizes came from the browser ZIP and UI-exported GLB; 100-job credit/time scenarios and ideal transfer times are multiplication/division under stated no-failure/serial assumptions. No subscription price, Hyper3D complete PBR debit, repair hours, physical print, leak test or dollar exchange between credit systems was measured.

Hyper3D Vase Review 2026: A Watertight Shell, but No Big Cable Crossings

Titus Brett — Sat, 10 Oct 2026 16:05:16 +0000

Disclosure: this is an independent hands-on test. Both tools were run on ordinary customer accounts; neither company supplied review access or saw this piece before publication.

I wanted a vase whose thick knitted cables were there when the material was switched off. Not chevrons painted on a smooth pot: actual strands passing over one another. I gave Hyper3D and SupaVoxel the same image and took apart the downloaded meshes. Then I looked down the mouths. That second view made it impossible to mistake “vase-shaped” for “ready to hold flowers.”

My verdict in 60 seconds — Hyper3D Gen-1.5 has the cleaner starting solid, but SupaVoxel is the stronger knitted design. Hyper3D's downloaded model becomes one watertight face-connected shell with zero non-manifold edges after UV welding, while the SupaVoxel export keeps the photographed raised cable crossings but has 744 non-manifold edges and 497 face-edge-connected regions. If a file must go first to mesh repair and then slicing, I would inspect the Hyper3D shell first. If the brief is to preserve big physical-looking knit relief in a digital concept, I would start with SupaVoxel and repair it. Both cavities are shallow; neither has been printed, filled or proved to hold water.

Eight side-by-side checks, Hyper3D then SupaVoxel:

Rounded-vase silhouette — both retain mouth and belly; SupaVoxel shows fuller raised knit.
Raised cable crossings — the downloadable Hyper3D Gen-1.5 file simplifies them into ribs; SupaVoxel keeps over-and-under geometry without maps.
Watertight after UV seam welding — Hyper3D yes; SupaVoxel no, despite zero boundary edges.
Non-manifold edges — Hyper3D zero; SupaVoxel 744, not a solved slicer problem.
Face-edge-connected integrity — Hyper3D one shell versus SupaVoxel 497 regions after welding; the latter still forms one vertex-connected cluster.
Modeled cavity — Hyper3D 32.79–35.13 mm versus SupaVoxel 22.8–24.0 mm at 120 mm longest-side scale; both are shallow.
Slender-face hygiene — Hyper3D 1,416 slender triangles versus SupaVoxel just 17; both have zero degenerates.
Potential downward-facing surface — Hyper3D 4,655.3 versus SupaVoxel 5,704.4 mm² at one Y-up orientation; no support bill follows.

The decision changes with the intended use. Neither file is a measured physical vase.

Did the picture ask for real interwoven cables?

The same 1,746,019-byte generated PNG entered each product; the broad strands visibly cross, but the image supplies no dimensioned cavity or rear view.

Yes. The white source has thick, oversized knitted strands wrapping a round body. A strand running across another is a shape question before it is a color question. Both vendors got the exact SHA-checked input bytes, though no claim is made about identical internal image processing. Hyper3D offered a Gen-2.5 preview with larger crossings, but the tested Free account could not download it. The 60,000-triangle Gen-1.5 PBR file is the Hyper3D deliverable I measured against SupaVoxel's 971,714-triangle exported file. Those are separate workflows, not an export toggle on the same mesh. The input was made with a third-party cover as visual inspiration; this test does not establish rights to republish that underlying cover.

What did Hyper3D put in the downloadable shape?

Hyper3D Gen-1.5 front: narrow mouth and recognizable rounded vase, but mostly vertical ribs rather than the source's bold crossings.

From a distance, this is an attractive ribbed pot. Move in, and the strands from the photograph largely stop passing over one another. Fine diagonal stitches appear in the material, but the main downloaded geometry resolves into regular longitudinal ribs. The file contains 60,000 triangles and 59,614 raw vertices; the interface displayed 60,000 and 59,765 respectively, so I use the exported accessors for counts. This is not evidence that Hyper3D can never make cable relief—only that its accessible Gen-1.5 export did not preserve the defining large over-and-under form in this one image. If that motif is the commission, somebody would have to sculpt it back in, and I did not time that labor.

Do the crossings survive a clay render?

Remove color and normal maps: Hyper3D's raised longitudinal ribs remain, but most bold interweaving does not materialize.

Same no-texture view on SupaVoxel: thick cables still visibly travel over and under one another. Topology repair is a separate issue.

This is where SupaVoxel wins the design task. A bump suggested by a normal map changes lighting; a real mesh crossing can cast, silhouette and be inspected as geometry. The SupaVoxel file keeps the large bends in the unpainted view. I did not measure their relief depth in millimeters, and the extra 911,714 triangles do not by themselves prove greater likeness. The location of the triangles matters: SupaVoxel uses its mesh to make the source's big feature visible without material. If my deliverable were a digital sculpture to show a knit pattern, I would accept the larger file and make repair a separate task rather than settle for the Hyper3D ribbed substitute.

What happened when I welded the UV seams?

Hyper3D shows a Print-Ready label and 60,000 faces. The downloaded mesh, not the label, supplies the checks below.

Raw GLBs duplicate vertices across texture coordinates. Before welding, Hyper3D seemed to expose 49,964 boundary edges; treating UV splits as real holes would be wrong. Once coincident positions were merged independently of UVs/normals, Hyper3D's downloaded Gen-1.5 vase had one face-edge-connected shell, consistent winding, zero boundary edges and zero non-manifold edges, and was watertight. The raw vertex count went from 59,614 to 30,002 welded positions. These are unusually reassuring file checks for a print candidate, not an actual slicing, leak or load test. A closed shell can still be too solid in the wrong places, as the mouth sections show next.

Why is SupaVoxel's more faithful knit not print-ready?

After the same seam-weld method, SupaVoxel retains 744 non-manifold edges and fails watertightness. It yields 497 face-edge-connected regions. Important nuance: these are not 497 loose vases floating in space; at least one welded vertex connects those regions into one vertex-connected cluster. That distinction does not solve the manifold defect. Zero boundary edges and consistent winding coexist with the 744 problematic shared-edge junctions. I would not send this as an approved solid print simply because its renderer shows a coherent shape. SupaVoxel's menu offers Fix Mesh, but I did not run it; no repaired GLB, cleanup duration or proof of slicer acceptance is in this test. For a digital concept the knit wins; for an unchanged solid export the topology loses.

How deep is Hyper3D's apparently open vessel?

An actual mouth with a visible floor, not a view into a hollow lower belly.

I cut horizontal sections along the actual Y-up axis. When the file's longest bounding-box edge, not necessarily its height, is normalized to 120 mm, Hyper3D is about 117.11 mm tall. At 20%, 40%, 60%, 65% and 70% of that height, sections show one outer contour and no inner contour. A second inner loop appears between 70% and 72%, bracketing the modeled cavity at 32.79–35.13 mm deep from the top. The recess is real; a thin-walled, deep flower holder was not generated. A photograph cannot tell us the intended wall gauge, and these slices are threshold brackets rather than CAD tolerances. I would not promise full-length flower stems from that picture.

Is the SupaVoxel recess any deeper?

SupaVoxel also has an open mouth and an interior floor. This is not a tested water-holding container.

No. At the same 120 mm longest-side scale, the SupaVoxel body is 120 mm tall and its inner loop first appears between 80% and 81% of height. That brackets the modeled cavity at 22.8–24.0 mm, approximately 9–12 mm shallower than Hyper3D's bracket. More spectacular exterior knit did not buy a deeper vessel. Both models are mostly solid below their rims. These values are from mesh sections, not a water-volume or leak test. SupaVoxel's non-manifold geometry makes treating any signed volume as a trustworthy resin bill worse still. For a functional vase, both would need redesign: a wider/deeper interior, validated wall thickness, repairs and real testing.

What does solid resin arithmetic permit here?

Hyper3D's closed starting mesh allows conditional solid-fill volume arithmetic; its image still says nothing about liquid tightness.

Only the Hyper3D welded body passed the closed-solid checks needed for this estimate. At 120 mm maximum outer dimension, it computes to 817.04 cm³ of modeled solid; at a hypothetical $35 per liter standard resin, $28.60 before supports, waste, hollowing, washing or failed prints. That is not the cost of a successful vase. SupaVoxel has no validated solid-fill resin estimate in this case because its welded mesh is not watertight and carries non-manifold edges. I won't recycle a signed-triangle volume from an invalid solid and call it a bill. A real workshop would repair, revalidate, slice and only then estimate consumption for both, after choosing a functional cavity design.

Does “potential overhang” predict actual support?

The side view shows knit relief; it does not reveal actual slicer-generated support material.

With Y-up and normals within 45° of straight down, Hyper3D has 4,655.3 mm² (8.32%) of potentially downward-facing area and SupaVoxel 5,704.4 mm² (8.88%) at the same normalized largest dimension. That's 1,049.1 mm² more for SupaVoxel in this particular orientation. It does not mean 1,049.1 mm² more supports: bottom faces against the platform can count, orientation can change, and slicers use additional rules. This one normal-based measure favors Hyper3D, but implies no measured labor saving. If I rotate either vase or hollow its body, I would rerun the calculation instead of reusing this figure.

Does Hyper3D win every triangle-quality test?

The sparse Hyper3D wireframe is legible; its slender-face count is a separate measured file statistic.

No. Both files had zero degenerate faces under the common threshold, but Hyper3D had 1,416 slender triangles where SupaVoxel had 17. SupaVoxel wins that narrow hygiene row by 1,399 fewer slivers, despite losing the non-manifold and connected-shell rows. At the 120 mm scale their mean triangle edges are 1.5679 mm and 0.4410 mm. A 0.4 mm FDM nozzle offers context, not a guarantee that either strand is printable; mean edge length is not local groove width. You cannot hide Hyper3D's slivers behind its “watertight” label, just as you cannot hide SupaVoxel's 744 non-manifold edges behind attractive knit relief.

Was the better Hyper3D preview actually downloadable?

The Gen-2.5 screen shows larger crossings and a million reported faces; it supplied no audited GLB to this Free user.

Download on that confirmed Gen-2.5 task asked for a subscription and said Gen-1.5 downloads remain free.

The Gen-2.5 preview looked closer to the source than the Gen-1.5 model I took home. I would have liked to test its weave and topology. After confirmation the UI displayed 1,000,000 faces; the Free account could select GLB but not download it. Gen-2.5's automatic caption also differed from Gen-1.5's, omitting “knitted,” so I cannot assign their visual difference to version or triangle count alone. No Gen-2.5 file was measured; no claim about its watertightness, cavity or price of a downloadable PBR is warranted. A gated preview is a useful purchase illustration, not evidence for the file comparison on this page.

Final verdict: which problem would I rather repair?

For a first physical-prototype starting mesh, I favor the downloadable Hyper3D Gen-1.5 file's closed, single-shell topology and slightly deeper modeled recess. That is not an endorsement of it as a functioning vase. It largely misses the reference's defining cable crossings and has 1,416 slender faces. SupaVoxel nails the large geometric knit but its 497 face-edge regions and 744 non-manifold edges need repair before a solid-print claim, and its recess is only 22.8–24.0 mm deep. If the work is a digital relief study rather than a print, I choose SupaVoxel; if I need an untouched, cleaner starting shell, I choose Hyper3D. Waterholding remains untested for both.

Use SupaVoxel when the design brief is knit geometry

For a concept sculpt in which raised strands must remain when textures are hidden, SupaVoxel gave me the closer geometric starting point. I would repair its junctions and retest the corrected file before slicing; I would also redesign the shallow interior before calling it a flower vase. Hyper3D Gen-1.5 is my alternative for a simpler already-closed ribbed body. Neither recommendation substitutes for a print, leak test or vase-stability measurement.

How I tested and what remains open

Both files came from the same SHA-checked PNG. I compared actual exported Hyper3D Gen-1.5 PBR and SupaVoxel Original size GLBs, not the locked Gen-2.5 preview. Matched offline cameras, clay renders and Y-up horizontal sections were used; the sections bracket interior loops, not precise CAD measurements. UV-seam vertices were welded before topology checks. All 120 mm figures use the longest bounding-box side, which gives Hyper3D only 117.11 mm of vertical height. The downward-area rule and resin-dollar figure are arithmetic under explicit assumptions; neither measures printed supports or a watertight physical vessel. No Fix Mesh operation, slicer, print, fill or stem-stability test was performed.

Codex Pricing in 2026: Every Plan, Usage Limit and Credit Rate, Explained

Titus Brett — Thu, 08 Oct 2026 15:54:41 +0000

Codex pricing is not one number: Codex is bundled into ChatGPT Free, Go, Plus, Pro, Business and Enterprise, each with its own usage allowance, and then extended with token-based credits or an API key. I went through the Codex pricing page, the developer docs and the usage-limit tables to lay out what each tier actually buys, how the per-model limits work, and where the money goes once you hit them.

Ask “how much does Codex cost” and the honest answer is “it depends which ChatGPT plan you are on and which model you point it at” — which is why Codex pricing is one of the most-searched questions about the product. OpenAI’s own pricing page opens with “Codex is included in your ChatGPT Free, Go, Plus, Pro, Business, or Enterprise plan”, and everything after that is a matrix of usage estimates, model multipliers and credits. I spent a week with the Codex pricing page, the developer-docs pricing reference and the credit rate card to turn that matrix into something you can budget against.

Because Codex pricing rewards people who spend their allowance on the hard problems and not on scaffolding, I also kept a product builder that produces a deployed, working app from a sentence, Begin, open in the next tab — I explain why at the end.

Codex pricing starts at $0 and climbs to “from $100” for Pro; the interesting numbers are in the usage table underneath.

Codex pricing: the plans

The individual tiers on the Codex pricing page in 2026:

Free — $0/month. “Limited trial access to Codex”: limited messages and uploads, limited access to GPT-5.6, no image generation. Enough to test Codex on quick tasks.
Go — $8/month. Still described as trial access, with more messages, more uploads and access to the latest models. Codex for lightweight coding.
Plus — $20/month. “Power a few focused coding sessions each week.” Codex on the web, in the CLI, in the IDE extension and on iOS; the macOS and Windows Codex app; cloud integrations such as automatic code review and Slack; the full GPT-5.6 family (Sol, Terra, Luna); plugins for GitHub, Figma and more; and the ability to extend usage with ChatGPT credits.
Pro — from $100/month. Everything in Plus with 5x the Codex usage, or a higher 20x tier. “Designed for longer, higher-intensity sessions each day.” Unlimited GPT-5.6 subject to abuse guardrails.

Business is $20 per user per month (2+ users, billed annually; $25 monthly) with a dedicated workspace, SAML SSO, larger cloud VMs and no training on your data. Enterprise and Edu are custom, adding SCIM, EKM, audit logs, data residency and priority processing. And separate from all of these, an API key runs Codex in the CLI, SDK or IDE extension at ordinary API rates, with no cloud features.

Codex usage limits by model

This is the part of Codex pricing that matters day to day. Limits are quoted as estimated local messages per five-hour window, and they vary enormously by model:

GPT-6 Astra — Plus 5–45 · Pro 5x 25–225 · Pro 20x 100–900
GPT-5.6 Sol — Plus 10–100 · Pro 5x 50–500 · Pro 20x 200–2,000
GPT-5.6 Terra — Plus 25–200 · Pro 5x 125–1,000 · Pro 20x 500–4,000
GPT-5.6 Luna — Plus 250–2,000 · Pro 5x 1,250–10,000 · Pro 20x 5,000–40,000
GPT-5.5 — Plus 15–80 · Pro 5x 75–400 · Pro 20x 300–1,600

Standard Business seats match Plus; Business at $100 uses the Pro 5x numbers. The ranges are wide on purpose: a small script consumes a fraction of a “message”, a long agentic session holding a big repo in context consumes many. Local messages and cloud chats share the same allowance, cloud chats default to Sol and cost more, weekly limits can apply on top, and Fast mode and image generation both burn the allowance faster (images 3–5x).

OpenAI’s guidance on model choice maps directly to cost: Sol for the hardest reasoning and advanced coding, Terra as the everyday workhorse, Luna for fast, high-volume, focused tasks. On Plus, moving a routine task from Sol to Luna is the difference between 10 messages and 250 in the same five hours.

The docs version of Codex pricing adds the detail the marketing page leaves out — shared allowances, weekly caps and what each model is for.

Codex credits: what happens when you run out

Since the move to token-based pricing in April 2026, Codex pricing above the included allowance runs on credits. Plus and Pro users can buy extra credits instead of upgrading; Business, Edu and Enterprise workspaces on flexible pricing buy workspace credits. The published rate card, in credits per million tokens (input / cached input / output):

GPT-6 Astra — 250 / 25 / 1,250 (Fast mode is 2.5x)
GPT-5.6 Sol — 100 / 10 / 500 (promotional through at least 21 November 2026)
GPT-5.6 Terra — 50 / 5 / 300
GPT-5.6 Luna — 5 / 0.5 / 30
GPT-5.5–125 / 12.5 / 750
GPT-Image-2–200 / 50 / 750 for image tokens

OpenAI’s own rule of thumb is that GPT-5.6 usage averages 5–30 credits per message. If you hit the limit mid-turn, the agent finishes the turn it is on. Voice in the desktop app draws from the same budget at $0.05 a minute (1.25 credits a minute on credit-billed workspaces). And anyone can run extra local sessions on an API key at standard API rates, which is the escape hatch for CI and shared environments.

Two dates to know: GPT-5.5 retires from Codex on 14 October 2026 on every plan, and the API is unaffected.

Five-hour windows, weekly caps, and /status in the CLI for your live remaining allowance.

What Codex pricing gets right

A real free tier and an $8 entry point. You can evaluate Codex without a $20 commitment.
Model-tiered limits let a Plus subscriber stretch the plan a long way by defaulting to Luna and Terra.
Credits instead of forced upgrades when you overrun.
Finish-the-turn grace — hitting the limit does not abandon a running task.
An API-key path for automation, with usage-based billing and no plan required.
Transparent rate card with per-model credit costs published.

Where Codex pricing falls short

The ranges are enormous. “5–45 Astra messages” is not a budget; it is a warning.
Cloud chats cost more and share the allowance, so the convenient path is the expensive one.
Pro’s “from $100” hides a 20x tier priced separately, and the 5x/20x split is only explained in the docs.
Everything shares one pool — images, voice, Excel, code review through GitHub — so non-coding features quietly eat coding limits.
Weekly limits on top of five-hour windows mean a heavy Monday can cost you Friday.

Codex pricing vs the alternatives

Against per-token API billing, the ChatGPT plans are cheaper for most individuals until you run agents unattended for hours, at which point the API key wins on predictability. Against competing agent subscriptions, Codex pricing is competitive at Plus and expensive at Pro 20x. The real optimisation is not the plan — it is not spending Sol messages on work a template or a builder could have done.

Verdict: what should you pay for Codex in 2026?

Start on Free or Go to see whether Codex fits your workflow. Most working developers land on Plus at $20 and get the most from it by routing routine work to Luna and Terra and reserving Sol and Astra for hard problems. Move to Pro when you are hitting five-hour limits several times a week; use an API key for CI. Check the live tables on the official Codex pricing page before you commit, because the usage estimates move.

The alternative worth keeping next to it: Begin

The Codex pricing tables make one thing obvious: the expensive messages are the long agentic ones, and a surprising number of those are spent standing up a new project — auth, database, payments, deploy — before any interesting code exists. Begin removes that spend by producing a finished, deployed product from a sentence or a domain.

One brief, four targets. Describe an app, or paste a URL, and Begin builds a website, an iOS app, an Android app or a Chrome extension on ordinary, readable code you keep working in.
Batteries wired in. Sign-in, email, a database, storage and Stripe checkout are part of every project; each build goes live on Begin’s edge with TLS and your own domain.
Credits you can predict. Every plan grants 100 credits a month; a first build costs 5, every change afterwards costs 1, failed builds are refunded, and unused credits roll over on paid plans. Five free credits — one real build — need no card.
$49 and $99 tiers that differ on the workspace around the credits, not on what a credit buys.

Let Begin do the scaffolding at 5 credits. Spend your Codex Sol messages on the code that only you can specify — and keep Begin open for the next product.

Photo by Mohammad Rahmani on Unsplash

Tripo AI review: I ran the free tier through a Unity asset pipeline and here's what survived

Titus Brett — Mon, 21 Sep 2026 21:26:32 +0000

1,012,219 vertices. 1,911,270 faces. That is what Tripo handed me for a single character, from a single reference image, on a free account that had existed for about four minutes.

I'd been putting off a proper look at Tripo (the web app at tripo3d.ai, also branded Tripo Studio) because most "AI 3D" I've tried produces a hero-shot blob that turns into soup the moment you need it in an engine. Tripo keeps getting named as the exception, so I tested it the way I'd actually use it for a prop: reference in, retopologise, texture, export GLB, drop into a Unity URP project, see what breaks. I kept SupaVoxel open in the next tab as the image-to-3D control group, mostly to compare what each free tier really lets you walk away with, and I'll get to that at the end.

Everything below is from the live product on the free plan. Credit costs are what the UI showed me before I clicked, not what the pricing page promises.

Signup, and what "200 credits" buys

Account creation is the least interesting part, so quickly: Google, Microsoft, Apple, Discord or email, six-digit code from notification@tripo3d.ai, no card. A short survey asks what you build (I picked Game Assets, then Character) and you land in the workspace with 200 monthly credits.

Tripo's pricing page turns 200 credits into "about 13 models". That's true if a model is one generation and nothing else. It is not true for a game asset, and I'll show the real arithmetic once the numbers are in.

Plan	Monthly	Credits	Concurrent tasks	Licence
Free	$0	200	1	Public, non-commercial
Pro	$20 ($240/yr, ~$13/mo annual)	3,000	10	Private, commercial
Max	$90	25,000	100	Private, commercial
Team	$55/seat	90,000	200	Private, commercial

Two free-plan restrictions matter more than the credit count. Your models are public and the licence is non-commercial, so nothing you make on the free tier can ship. And "limited exports" means the download dialog is where the Pro badge shows up. Pro also raises the polygon ceiling to 2M and adds batch generation and a 7-day edit history.

Generation: 55 credits, about a minute, one enormous mesh

The home box defaults to image-to-3D and ships with sample subjects. I used the Dragon Archer sample so the result is reproducible for anyone who wants to check my numbers. Before generating, the panel showed the cost outright: 55 credits for Best Quality. There's a Clean Topology toggle next to it and a choice between the standard HD Model and the newer Smart Mesh, which can output quads directly. I left it on HD with Clean Topology on, because I wanted to see the raw output before Tripo's own cleanup.

Roughly sixty seconds later I had the model in Tripo's editor. The silhouette was faithful, the bow and quiver read correctly from the side, and the stats panel said what I quoted at the top: just over a million vertices, 1.9 million triangles.

To Tripo's credit, the number is right there at the top of the panel. A lot of tools bury it.

Inspecting the GLB before Unity ever sees it

I don't trust a viewer's count, so I pulled the raw GLB and checked it locally. Two lines of trimesh tell you everything you need to know before you waste an import:

import trimesh

scene = trimesh.load("dragon_archer_raw.glb")
mesh = scene.dump(concatenate=True) if isinstance(scene, trimesh.Scene) else scene

print(f"verts: {len(mesh.vertices):,}  faces: {len(mesh.faces):,}")
print("watertight:", mesh.is_watertight)
print("extents (units):", mesh.extents.round(3))
print("euler number:", mesh.euler_number)

The counts matched the UI. is_watertight came back False on the raw mesh, which for a game asset is irrelevant but would matter if you were heading to a printer. Extents were roughly 1 unit tall, so Tripo is exporting in metres, which saves a scale fight later.

If you'd rather stay in Blender, the same check in the Python console after import:

import bpy
objs = [o for o in bpy.context.selected_objects if o.type == 'MESH']
print(sum(len(o.data.polygons) for o in objs), "faces across", len(objs), "objects")

Either way: 1.9M triangles for one character is not a Unity asset. It's a sculpt.

Retopo is the feature, not the generation

This is where Tripo stopped being a demo for me. The Retopo tool gives you a Polygon Count slider from 500 to 50,000, a Triangles/Quads switch, and a Smart Low Poly toggle for aggressive reduction. I set 10,000, triangles, Smart Low Poly on. Cost: 10 credits. Time: comparable to the generation.

Quick guidance if you're new to this choice. Triangles for a real-time engine, because that's what the GPU draws anyway and you avoid non-planar quads getting split unpredictably at import. Quads if you're going to subdivide or sculpt further in Blender. Smart Low Poly is the right pick for a prop; for a hero character you'd want to crank the count and turn it off, because it collapses small detail on hands and faces.

The 10k result held the silhouette. The bowstring survived as geometry, which surprised me. Fingers became mitten-ish, which didn't.

Texture and export

Texturing is a separate pass with PBR maps up to 4K (8K is a members-only trial). The export dialog offers GLB, FBX, OBJ and STL with a texture-resolution dropdown, and a Send To button that pushes into Blender or Maya through Tripo's DCC Bridge.

On the free tier this is the point of friction: exporting and unlimited downloads sit behind the Pro upgrade. You can evaluate the mesh thoroughly inside Tripo for free; getting it out is where they want the $20.

Into Unity

For anyone following along: install com.unity.cloud.gltfast from the Package Manager (by name), drop the GLB into Assets/, and it imports as a prefab with a URP/Lit material if your project is on URP. Things I'd flag from this specific asset:

The 4K texture set is 4096² per map. For anything targeting mobile, set Max Size to 2048 in the texture importer or you'll ship 60MB of dragon.
Tripo exports Y-up, metres, so scale and orientation were correct without touching the import settings.
The 10k-tri mesh with one material is a single draw call. That's the whole point of the retopo step.
No rig on this pass. Tripo's Animate tab does auto-rigging, but I didn't spend credits on it here.

It sat in a test scene next to a Unity Asset Store character of similar budget and didn't look out of place, which is the bar I care about.

The real credit arithmetic

Here's the "13 models a month" claim measured against what I actually did:

Step	Credits
Best Quality generation	55
Retopo to 10k	10
Texture pass	separate charge
Total, one usable asset	65+

200 credits divided by 65-plus is three finished assets, not thirteen. Add a re-roll when the first generation misreads the reference and it's two. It's plenty to evaluate the tool. It's not enough to build anything with.

What I didn't like

Raw output is so dense that retopo isn't optional. It's a mandatory second step and second charge, and the pricing page's per-model estimate ignores it.
Free-tier licence is non-commercial and models are public. There's no path from "free" to "in my shipped game" without paying.
Exports are gated. Evaluate for free, download for money.
One concurrent task on free means batch iteration is slow by design.
Smart Low Poly is a blunt instrument on hands and faces.

Verdict

Tripo is one of maybe two AI 3D tools where I'd keep working on the output rather than close the tab. The editing stack (Segment, Retopo, Texture, Animate, Smart Mesh quads, DCC Bridge) is what makes it a pipeline tool rather than a toy, and upfront credit costs on every action are how it should be done. If you're evaluating, spend the 200 free credits at tripo3d.ai and judge the retopo yourself. If you're shipping, Pro at $20/month is the real entry price, and annual brings it to roughly $13.

Where SupaVoxel fits next to it

The two things that stung on Tripo's free plan were the non-commercial licence and getting three assets out of 200 credits. SupaVoxel is the image-to-3D tool I had open alongside it, and on those two axes it's a different deal:

3 free models a day, no card, versus a handful a month once retopo and textures eat Tripo's credits.
Credits don't expire and don't reset monthly, so there's no end-of-cycle race.
The annual plan is around $25/year against Tripo Pro's $240/year, and paid tiers include a private commercial licence.
Image to textured model in about 12 seconds, exported to GLB, GLTF, FBX, OBJ, STL, PLY or USDZ with textures baked.
One-click watertight repair, an auto flat base and colour 3MF, which Tripo leaves to you.

If you want the deepest editor, with segmentation, rigging and quad Smart Mesh, Tripo is the bigger workshop. If you mostly want a clean, commercially usable mesh from a reference image without rationing credits, try SupaVoxel first and spend Tripo's credits on the assets that need its extra tools.