Industry research
Questions raised by the wider open-weight-vs-closed landscape — cost data, case studies and rankings, sourced and cited rather than asserted.
37 questions
Open-Weight Model Pricing Ranked: The 2026 Discount Table
In context →- Why do the two rankings put models in a different order?
- The rate-card ranking prices every token fresh. The agent ranking assumes 60% of input is a cached, repeated prefix — and models differ in how much their cached rate discounts their standard rate, so a model that is merely good on a one-shot request can be the best choice for a loop.
- Does a bigger discount mean a better model?
- No — it means a bigger gap between what we charge and what the model's own publisher charges for the same model. It says nothing about output quality, which we do not rank; see /press for what we do not claim.
- Why are 17 of the 24 catalogued models not callable?
- They are priced at their intended rate but no configured upstream currently serves them. We list them anyway, marked clearly, rather than removing them the moment they become servable and readding them later — the catalogue is the honest current state, not a promise.
Kimi K3 vs GLM-5.2 vs DeepSeek V4: The Real 2026 Cost Comparison
In context →- Which of the three is cheapest?
- DeepSeek V4 Pro is the least expensive of the three on both input and output at our rates; Kimi K3 is the most expensive; GLM-5.2 sits between them. See the rate figures above for the exact numbers, read from the same table our API bills from.
- Does a lower price mean a worse model for coding work?
- Not according to the reviews cited here — none of them rank the three strictly by price, and Kingy.ai's mid-2026 standings put GLM-5.2, not the cheapest of the three, at the top of its open-model field. We have not independently verified that ranking; it is attributed to Kingy.ai, not stated as our own finding.
- Are all three callable today?
- Yes. Kimi K3, GLM-5.2 and DeepSeek V4 Pro are three of the seven models in this catalogue currently served by a configured upstream, so every rate in this article is a rate you can actually call against, not a catalogued-but-unservable price.
Data Sovereignty and the Open-Weight Advantage
In context →- Does hosting an open-weight model automatically solve data sovereignty?
- No. It removes the vendor's server as a place your data has to pass through by default, which is the specific problem the sources above describe, but the deployment around it — where you host, who can access it, what logging you keep — still has to be built correctly. The licence also has to be checked: a publisher-specific text can carry its own conditions, and only a standard text like MIT or Apache 2.0 can be described from its name alone.
- Are most of the models in this catalogue easy to adopt from a licensing standpoint?
- Most sit under a standard permissive text with no user threshold and no field-of-use restriction, per the figure above. A smaller number carry a publisher-specific licence that has to be read before you rely on it — see /licences for the model-by-model reference, including which of each family is currently callable.
- Is this saying open-weight models are more secure or higher quality than closed ones?
- No. It is a claim about who controls where your data goes and what happens if a vendor changes course, not a claim about benchmark performance — see our other resources on the honest capability comparison between open and closed models.
The 2026 Open-Weight Landscape: What the Independent Reviews Actually Say
In context →- Do all the sources in this roundup agree with each other?
- No, and we say so rather than smoothing it over — Kingy.ai's own shortlist reports two different Artificial Analysis Intelligence Index scores for GLM-5.2 within the same piece. Where sources disagree, both numbers are stated and attributed rather than one being silently chosen.
- Does any of this mean open-weight models beat closed frontier models?
- No. Hakia's comparison, cited above, states that closed frontier models retain a measured lead on reasoning-heavy benchmarks as of September 2026. The case this roundup collects is about cost, control and sovereignty, not a claim that open weights are categorically the stronger models.
- Is the discount figure in this article the same as this site's headline savings claim?
- It is the rate-card discount for the seven models currently callable here, computed the same way as everywhere else on this site. It is not a claim from any of the outside sources above — those sources are cited for market context, not for our own pricing.
Video and Image Generation Models: What's Priced and What's Actually Callable
In context →- Can I call any video model on this platform today?
- No. Wan 2.2 T2V, Wan 2.2 I2V, LTX-2.5, HunyuanVideo 1.5, CogVideoX-5B and Mochi 1 are all catalogued at a real, published rate, and none currently have a configured upstream serving them. A request naming one returns a clear error.
- Which image model can I actually call today?
- Qwen-Image. It is the only one of the four catalogued image models — the others being FLUX.2 [schnell], Stable Diffusion 3.5 Large and HiDream-I1 — currently served by a configured upstream.
- Why publish a price for a model you can't serve?
- So the catalogue stays complete and comparable rather than silently shrinking to only what happens to be live this week. Every unservable model is marked plainly, and the price shown is what it will cost once served, computed the same way as every callable model's rate.
Why Enterprises Are Switching to Open-Weight Models in 2026
In context →- Is the 81% figure from Vercel typical of what routing saves?
- No -- Vercel presents it as one product's own bill, not an industry average, and attributes it to routing discipline layered on a model swap rather than to switching models by itself [1]. Broader research puts the more general figure at roughly 40% [2].
- If open-weight economics are this well documented, why hasn't adoption caught up?
- Research on the gap attributes it mainly to integration and MLOps overhead -- rebuilding evaluation and approval processes around a new model -- rather than to the cost case being weak [3].
The Hidden Cost of Vendor Lock-In: What Closed-API Deprecations Actually Cost You
In context →- Does using an open-weight model mean an application is never locked in again?
- No -- it removes one specific risk (the model itself disappearing on a provider's schedule) and not every dependency a production system has, including which host actually serves the weights. This API is built OpenAI-compatible specifically so a later host change doesn't require an application rewrite.
- Why do naive migrations cost so much more than the sticker price of the new model?
- The reported 40-85% increase [1] comes mostly from behaviour differences the old prompt and evaluation harness weren't built for -- a replacement model responds to the same prompt differently, and a migration done without rebuilding around that produces both a cost and a latency regression.
- Which of the models in the chart above can this API actually serve today?
- Kimi K3, DeepSeek V4 Pro and GLM-5.2 are callable now. Llama 4 Maverick, Mistral Large 3 and Qwen3 Max Instruct are catalogued at their intended price but not yet served by any configured upstream -- a request for one returns a clear error rather than a fabricated result.
Prompt Caching: Why It's the Biggest Lever in Your LLM Bill
In context →- Does caching happen automatically, or do you have to design for it?
- The cached rate applies automatically whenever a request's prefix exactly matches one already sent, but getting the benefit takes application-level discipline: keep the system prompt and tool schema first and unchanged, and append rather than edit earlier context [3].
- How much of a real bill can caching remove?
- Independent research on repeated-prefix workloads puts the range at 41-80% depending on how much of the prompt actually repeats and how consistently it's structured [1]; a workload with little repeated context has correspondingly less for caching to remove.
The rest, answered on their own pages
Every question below is answered in full where it belongs, beside the rate table and the specification it refers to.
Check it against your own numbers
$5 in free credits, no card required.
Every rate quoted above is the rate the API bills.