Choose a model you can actually use.

“Frontier” is a capability judgment, not a hosting type

Here, frontier candidates are models documented for demanding general reasoning, coding or multimodal work, with published evaluation evidence. The label is an editorial shortlist checked on the stated date. We have not independently established which model leads the market, and a provider’s launch language alone does not make every product frontier.

Efficiency-oriented hosted models, compact local models and specialist image or speech models can be better fits for a particular job. Our older open-weight baselines remain useful setup examples; they are not automatically current frontier models. Compare like tasks before comparing scores.

A subscription and an API are different purchases

A web or desktop subscription provides an application, its tools and usage allowance. An API is an interface your software calls, usually with metered input and output. Paying for a chat plan does not provide unlimited API work. Some eligible plans offer separately claimed API credits, with expiry and account conditions.

A token is a chunk of text used for billing and context limits, not a fixed word count. Input tokens include the material you send; output tokens are what the model generates. A discounted cached-input rate applies only when a provider can reuse eligible earlier input under its caching rules.

Current eligible ChatGPT Pro/Business and Claude Max/Team subscriptions have documented API-credit programs. Check the linked terms and your account before counting those credits in a budget. Claude’s public pricing lists Pro at USD $20 monthly or $200 annually paid upfront, and Max from $100 monthly; model access, tax and usage limits still apply. Google’s current plan page confirms Argon access in eligible AI Pro/Ultra plans, but a reliably verified numeric subscription price is not included here.

Base API rates for the reviewed hosted models

USD per one million tokens, checked 29 October 2026. These are standard/base rates for the stated prompt tier, not all-in task costs. Open each profile for cache rates, long-prompt surcharges, tool charges and exact dated conditions. On a narrow screen, scroll the table sideways to see all columns; keyboard users can focus it and use the arrow keys.

For example, 10,000 input tokens at $2 per million plus 2,000 output tokens at $10 per million equals $0.04 in base token charges. This is arithmetic, not a measured task cost: reasoning output, repeated history, failed attempts, tools, caching and pricing tiers can change the bill.

Local fit starts with memory, not the model’s name

System RAM is your computer’s working memory; VRAM is dedicated graphics-card memory. A download’s file size tells you how much storage it occupies, not the memory required while it runs. A model also needs runtime buffers and a conversation cache, often called the KV cache. Images, audio, longer context and simultaneous requests add further pressure.

Quantization stores weights at lower precision to reduce memory. It can change quality and compatibility; “4-bit” is not a universal file size or fit guarantee. As a rough lower-bound calculation, 8 billion parameters at 4 bits are 4 GB before metadata and everything else. CPU offload can move some storage and work out of VRAM, often at a speed cost. Unified memory on some computers is shared rather than a separate RAM-plus-VRAM pool.

Start with the exact model file and runtime documented for your device, leave room for the operating system, and test a short input. Then increase context and input size while observing actual memory and response time. Do not buy a GPU solely from a parameter count or a vendor’s maximum context number.

Concrete examples, with their limits attached

Ollama’s qwen3:8b Q4_K_M package is about 5.2 GB to download. Google lists Gemma 4 E4B loading estimates of 17.9 GB BF16, 8.9 GB SFP8 and 4.5 GB Q4_0, explicitly excluding conversation context and supporting software. Those numbers measure different things and are not interchangeable hardware minimums.

OpenAI describes gpt-oss-20b in MXFP4 as fitting within 16 GB memory and gpt-oss-120b as fitting one 80 GB GPU. Those are publisher configuration claims. Runtime support, context and concurrent work still need testing. The 120b example usually moves ordinary-PC buyers toward rented GPUs or substantial local hardware.

Open weights give deployment control, not free operation

First read the exact checkpoint license. Open-weight means you can obtain the trained files; it does not imply unrestricted commercial use. A hosted API of an open model, a rented dedicated endpoint and a GPU you own are separate purchasing choices.

For a server deployment, choose a supported runtime and weight format, allocate enough GPU memory, load the model, then test one request before increasing concurrency. Add authentication, request limits and monitoring before exposing an endpoint. A notebook demonstration is not an operated service.

A managed API typically meters tokens. A dedicated GPU endpoint charges allocated capacity, including idle time under its billing rules. Your own machine has purchase, electricity and maintenance costs. As dated examples, Hugging Face lists one AWS A100 80 GB at $2.50 per hour and one GCP H100 80 GB at $10 per hour. Those are infrastructure rates, not a verified quote for a particular model or throughput. Provisioning, quotas, scale-to-zero behavior and compatibility must be checked.

Explore server-scale model profiles

Read the test conditions before the score

A benchmark measures performance on a particular task set under particular conditions. Math-exam accuracy, software-task completion, factual errors and memory use do not measure the same capability. Higher is not always better: an error rate is better when lower. A pass@1 estimate is also different from choosing a consensus among many attempts.

Check the exact model version, dataset variant, reasoning effort, tool access, harness and date. A coding harness is the surrounding software that gives a model files and tools; changing it can change a score. A publisher-modified dataset is not directly comparable with its original. Production safeguards and fallback models can also affect the reported result.

Each profile keeps selected publisher-reported rows next to their settings and limitations. Missing dates or settings are stated rather than filled in. PrisAIm has not reproduced these model evaluations, and does not combine them into a synthetic winner. Use the reports to choose a trial, not as a promise of accuracy on your own work.

Run a small trial you can judge

Choose a representative input you are allowed to use and write down the output you need. For document extraction, try several short documents with known answers, including one that lacks the requested fact. For coding, use a small change with checks that can actually run. For speech, compare a transcript with the recording; for images, inspect the output against the brief.

Keep the prompt and task constant when comparing options. Record the exact model, settings, tools, input length, errors, corrections, elapsed time and actual bill. Include a failure case. These are proposed trials, not results we have obtained. A model is useful when it completes your task with acceptable review effort, not simply when its response sounds confident.