Short answer. Buy for the workload you can measure, not the model you saw in a demo. For 2027 planning, compare three things: the all-in cost at your peak, the evidence behind quality claims, and the contract's exit and change terms. Flat-rate plans remove per-token anxiety but not concurrency limits; token pricing is flexible but unpredictable. Run a 100–300 prompt evaluation on your own data before you commit annually.
Quick answers
This page is a planning guide with 2026 evidence. Every price and plan limit below was checked on 20 November 2026 and should be re-checked before signature.
| Buyer question | One-sentence answer |
|---|---|
| What changed in 2026? | Longer context became common, flat-rate plans with no token cap appeared at the top end, and procurement scrutiny moved from demos to contracts and evidence. |
| Is flat-rate cheaper? | Only if your token bill at the same workload exceeds the plan fee and your peak fits the concurrency limits. |
| What does no token cap mean? | No monthly token cap and no per-token overage; fair-use concurrency and RPM limits still apply. |
| Should we commit annually? | Only after a measured pilot; annual billing is 20% off, but the discount is not a reason to skip the exit terms. |
| How many prompts for an evaluation? | 100–300 well-chosen prompts per workflow is a practical start, and more if the task is high-risk. |
| What is the biggest hidden cost? | Retries, evaluation runs, retrieval, observability and the people who manage spend. |
| How do we compare providers? | Same prompts, same date, same scoring rubric, and published prices captured on the same day. |
| What belongs in the contract? | Price protection, model change notice, SLA credits, data terms, exit assistance and deletion. |
What changed by November 2026
The useful planning question is not “what will 2027 bring?” but “which 2026 changes are now structural enough to design around?”
What actually changed in AI API buying during 2026?
Four shifts are visible in the catalogue and contracts we checked. Context windows moved well past the 128K default, with Plugsky's main text models serving 262,144 tokens. Plugsky's upper tiers added flat-rate plans with no monthly token cap. For GCC buyers, regional deployment became a procurement line item. And the contract now matters as much as the demo: residency, subprocessors, model change notice and exit terms can decide the purchase. Treat these as observations from 20 November 2026, not a forecast.
Are context windows still a differentiator?
Less than they were. A 262,144-token window is now a baseline for Plugsky's main text models, with smaller windows for translation, vision and GPU-share routes. The useful question is not the maximum but the effective context: how well the model retrieves and reasons across a long document, and what it costs to send that document on every request. Long context can hide a retrieval problem. Test needle-in-a-haystack, multi-document synthesis and citation accuracy at the length you actually use.
Did flat-rate pricing become mainstream?
At the top of Plugsky's range, yes: Builder and Scale are priced at $60 and $120 per month with no monthly token cap and no per-token overage, subject to fair-use concurrency and RPM limits. That is not infinite capacity, and not every provider prices this way. The buying question is which risk you prefer: a predictable fee with a throughput ceiling, or a variable bill with elastic capacity. Read the limits, not the label.
What happened to model choice?
Catalogue size stopped being a useful proxy for capability. Plugsky publishes 39 model IDs, but some are compatibility IDs that route to a smaller set of serving engines; the API and the model ladder are authoritative. The practical consequence is that you should evaluate the engine that actually serves your request, not the legacy name on the ID. Ask for the mapping, the fallback chain and the change policy. A large catalogue with unclear routing is hard to govern. See the pricing cluster for the commercial detail.
Flat-rate versus token pricing
Plugsky's published plans are flat-rate; the comparison below is for buyers weighing that model against a token-billed provider.
| Plan | Monthly | Requests/day | Tokens/day | Concurrent | RPM | Input / output ceiling |
|---|---|---|---|---|---|---|
| Free | $0 | 100 (new) | 50K (new) | — | — | Model limits |
| Hobby | $8 | 1,000 | 2M | 4 | 120 | 64K / 8K |
| Starter | $20 | 2,000 | 5M | 6 | 240 | 128K / 16K |
| Builder | $60 | Not published | No monthly cap | 10 | 600 | 256K / 32K |
| Scale | $120 | Not published | No monthly cap | 16 | 1,200 | 512K / 32K |
New accounts from 2 November 2026 have the Free limits shown; older accounts retain 500 requests and 250,000 tokens per day until migration. Annual billing is 20% off, paid plans carry a 7-day money-back guarantee, and new monthly customers get 30% off the first month, capped at $20. The pricing page is the canonical source.
When is a flat-rate plan cheaper?
A flat-rate plan is cheaper when your workload is steady enough to fit the plan's daily token allowance and concurrency, and your token bill at the same volume would exceed the fee. Flat-rate is also cheaper in administrative terms: no spend dashboards per team and no surprise invoice after a batch job. Run your last three months of usage through the plan ceilings, including peak concurrency and RPM. If you are close to the ceiling, the operational risk may outweigh the savings.
When is token pricing cheaper?
Token pricing is cheaper when usage is spiky, experimental or genuinely small. It lets you pay for a pilot without committing to a plan, and it scales down when the workload disappears. The risk is that long prompts, retries, agent loops and evaluation runs multiply the bill in ways a demo does not show. Model the p95 request, not the average, and add a buffer for retries. If finance needs a fixed number, a capped plan is usually easier to approve.
What does “no token cap” not mean?
It does not mean infinite capacity. Plugsky defines “unlimited” as no monthly token cap and no per-token overage, not infinite throughput. Fair-use concurrency and requests-per-minute limits still apply, and individual model context limits still apply. A plan can also have an input ceiling per request, such as 512K on Scale. Read the plan table and the SLA schedule together. If your workload depends on bursts, ask what happens at the concurrency limit before you sign.
How do concurrency and RPM limits affect cost?
They determine whether you need a bigger plan, more retries or a queue. A plan can look cheap per token and still be expensive if peak traffic waits. Model concurrent requests, tokens per request, and the latency of your slowest step. Then compare the plan's concurrency and RPM with your p95, not your average. If you need more, the cost is either a higher tier or an architecture change: batching, caching, smaller models or asynchronous jobs.
What hidden costs should we model?
Include retries and failed calls, evaluation runs, embedding and re-indexing, retrieval infrastructure, observability, human review, security review, and the engineering time to switch models. Add the cost of a pilot that never ships. For flat-rate plans, add the cost of hitting a concurrency ceiling at peak. For token plans, add the cost of a runaway agent loop. The API fee is rarely the largest line item in year one. A total-cost model that ignores people and retries will flatter the cheapest option.
Evaluation criteria for 2027
A 2027 evaluation should be reproducible, dated and tied to your workload. A leaderboard is a shortlist tool, not a decision.
How should we score quality?
Start with the task, not a leaderboard. Define what a good answer looks like: correct facts, complete citations, acceptable tone, refusal when appropriate, and no harmful action. Build a rubric and have two reviewers score a sample. Track accuracy, consistency and failure severity. Use public benchmarks only to shortlist. A model that wins a general benchmark can lose on your documents, your language or your latency budget. Record the date, the model ID and the engine, because routing can change the result. See the 2027 comparison for the shortlist method.
How many prompts do we need for a credible evaluation?
For a single workflow, 100–300 prompts is a practical starting point: enough to cover common cases, edge cases and adversarial inputs without becoming a research project. Stratify by intent, document type, language and difficulty. Include at least 20 cases that should be refused or escalated. If the task is high-risk, or the decision affects a customer, increase the set and add expert review. The number matters less than the sampling: 100 carefully chosen prompts beat 1,000 random ones.
Should we test latency or throughput?
Both, at the percentile that matters. Time to first token drives perceived speed in chat; total latency drives completion; throughput drives batch and agent workloads. Measure p50, p95 and p99 under your expected concurrency, not on an idle endpoint. Test streaming, tool calls and long prompts separately, because they fail differently. If you have a regional requirement, measure from the region you will actually use. Publish the test method and date.
How do we evaluate security and legal fit?
Use the same evidence standard as the rest of procurement: a current SOC 2 report or readiness assessment with scope, a DPA and transfer terms, a subprocessor list, incident notification, deletion and audit rights. Then add AI-specific questions: model change notice, training use, prompt retention, key custody and the deployment's data path. Check that the region you need is live and that the support model does not create an unauthorized transfer. Legal fit is not a score you can average with quality; it is a gate. The security procurement guide has the checklist.
What should we ask about roadmap and deprecation?
Ask how much notice you get before a model or endpoint is retired, whether the public ID can change engines without notice, and what migration support is included. For pooled models, ask how the fallback chain is disclosed and whether the response metadata identifies the serving engine. Get the deprecation policy in writing and map it to your own release cycle. A vendor roadmap is not a commitment unless it is in the contract. Plan for at least one forced migration during a 12-month term.
Procurement and contract questions
The contract should make the commercial model legible to the people who operate it, not just to the lawyers who sign it.
What should be in an annual AI contract?
Cover scope and services, model and region commitments, SLA and credits, support and escalation, data processing and transfers, subprocessors, security obligations, incident notification, audit rights, price protection, renewal, termination assistance, deletion, liability and insurance. For AI, add model change and deprecation notice, evaluation data ownership, and the right to test a new engine before it serves production. Keep the order form short and put the operational detail in schedules you can actually compare.
Should we commit annually?
Commit annually only after a measured pilot and a security review. Annual billing is 20% off at Plugsky, and the 7-day money-back guarantee reduces the initial risk, but neither replaces an exit plan. If you commit, ask for price protection, a termination-for-convenience right after the first term, and a migration window if the model changes. If your usage is uncertain, start monthly and revisit after 60–90 days. The discount is valuable only if the workload and the contract both hold.
What price protection should we ask for?
Ask for a cap on renewal increases, notice before any change, and a written definition of the unit you are buying. For flat-rate plans, confirm that “no per-token overage” survives a plan change and that the concurrency and RPM limits are contractual, not just published. For token pricing, ask for rate-card notice and whether committed spend rolls over. If the vendor changes the model behind a public ID, ask whether the price can change with it.
What exit terms matter?
Require data export in an open format, a transition period with continued API access, deletion certification, log access during wind-down, and clear notice before termination. Check whether unused prepaid fees are refundable, whether you can terminate for cause after a security incident, and how disputes are resolved. For AI, add model deprecation as an exit trigger and require a migration window. Test the export before you sign: ask for a sample of your data in the format the contract promises. An exit plan you have never rehearsed is a hope. The residency guide covers deletion and transfer terms.
How do we avoid lock-in?
Keep the integration OpenAI-compatible where you can, isolate model-specific prompts and parameters behind a router, and store your evaluation set independently of the vendor. Avoid proprietary fine-tunes that cannot be exported, and keep a second provider or a self-hosted model for the critical path. Test a switch on one workflow before you need it. Lock-in is not only the API shape; it is the prompt library, the logs, the evals and the team's habits. Design for portability, but do not pretend that switching is free. See private AI for the deployment alternatives.
What to verify in a contract
These clauses turn a 2027 plan into an operable service with a known cost and a known exit.
- Unit of purchase. Plan, seats, tokens, requests, concurrency and RPM, with the overage rule stated explicitly.
- Price protection. Renewal cap, notice period and what happens if the model behind a public ID changes.
- Service levels. SLA percentage, measurement window, credits, exclusions and the plans that are out of scope.
- Model change. Engine, endpoint and fallback notice, evaluation rights and deprecation windows.
- Data terms. DPA, transfer mechanism, retention, training use, deletion and audit rights.
- Security. Report or readiness evidence, incident notification, subprocessors and key custody.
- Support. Severity definitions, response times, escalation path and named technical contact.
- Term and termination. Renewal, termination for cause, termination for convenience after the first term.
- Exit. Data return, transition assistance, log access, deletion certification and refunds.
- Remedies. Liability, indemnity, insurance and the dispute forum.
What we did not verify
We did not test any provider's latency, throughput or model quality for this page, and we did not obtain confidential price quotes. The plan limits and discounts are from Plugsky's published material on 20 November 2026; they can change, and the order form governs. Competitor prices are not quoted because they were not verified on the same date. We did not review a signed SLA, DPA or annual contract. The 2027 framing is a planning device, not a forecast: re-check the evidence before you commit budget or sign a term.
Frequently asked questions
Is this a prediction for 2027?
No. It is a planning guide built from evidence checked on 20 November 2026. Re-check prices, models and terms before you sign.
Does Plugsky charge per token?
Plugsky's published pricing is plan-based. Builder and Scale have no monthly token cap and no per-token overage; fair-use concurrency and RPM limits apply.
What does unlimited mean?
No monthly token cap and no per-token overage, not infinite capacity. Concurrency, RPM and individual model context limits still apply.
Which plan has no token cap?
Builder at $60/month and Scale at $120/month. Scale raises concurrency, RPM and the input ceiling; individual model limits still apply.
Is annual billing cheaper?
Annual billing is 20% off. The discount is worth taking only after a pilot, a security review and a check of the exit and price-protection terms.
How many prompts should we test?
100–300 per workflow is a practical start, stratified by intent, document type and difficulty, with at least 20 refusal or escalation cases.
Should we use public benchmarks?
Use them to shortlist, not to decide. They do not measure your documents, language, latency budget or failure costs.
What is the biggest hidden cost?
Retries, evaluation runs, retrieval, observability and human review. The API fee is rarely the largest line item in year one.
Can we switch providers later?
Only if you keep prompts, evals and logs portable. An OpenAI-compatible API helps, but the prompt library and evaluation set are the real lock-in.
What SLA applies?
A 99.9% headline uptime target, with plan-level terms and exclusions in the contract. Hobby is not in the SLA table.
Do we need an annual contract?
No. Monthly plans exist, and the 7-day money-back guarantee applies to paid plans. Commit annually only when the workload and terms are proven.
What should we re-check in 2027?
Prices, plan ceilings, model mappings, regions, SLA terms and the security evidence. Treat every number here as dated 20 November 2026.
Run the evaluation before the annual commit
Start on a plan that fits the pilot, measure the workload, then decide whether the annual term earns its discount.
Start free → See pricingSources
Primary documents and pricing we checked. The date matters more than the snapshot.
- Plugsky — pricing and plan limits — checked 20 November 2026.
- EUR-Lex — Regulation (EU) 2024/1689 (AI Act) — checked 20 November 2026.
- NIST — AI Risk Management Framework — checked 20 November 2026.
- ISO/IEC 42001 — AI management systems — checked 20 November 2026.
- European Commission — Standard Contractual Clauses — checked 20 November 2026.
- AICPA & CIMA — SOC 2 examination overview — checked 20 November 2026.
Last checked 20 November 2026. This is a 2027 planning guide, not a forecast or a price commitment.