Analysis · OIYO Editorial
AI price and performance: what to check before choosing ChatGPT, Claude or Gemini
Subscriptions, API token prices and benchmark rankings measure different things. Compare cost per successful task, review time and the requirements of your work.
Putting a price list beside a performance ranking can make choosing an AI seem straightforward. But a monthly subscription describes access to an app, API prices describe usage-based billing, and benchmark scores describe results on a particular test. Combining all three into one ranking may leave you with a cheaper model and more work.
This guide draws on official documentation and evaluation methodologies. It is neither a ranking of the latest models nor a hands-on product performance test. The A/B cost comparison is hypothetical; the official API price table states its own verification date and conditions.
1. Separate the subscription, model and API
A chat app, coding agent and directly accessed API can use the same model name but offer different ways of working. Search or file-editing tools may change the outcome as much as the model’s ability to answer. Start by recording the plan, model version, permitted tools and verification date beside each candidate.
| Product | Conditions to check | What this alone cannot tell you |
|---|---|---|
| App subscription | Tools, usage limits, resets and team management | Whether every API call is included |
| API | Model, provider, input, output, caching and tool charges | The overall convenience of an app using that model |
| Agent | Model, runtime, permissions and retry budget | Whether its score reflects the model alone |
“Subscriptions and APIs are always separate” is not a reliable blanket rule either. Check app allowances, metered charges for direct API calls, and any API credits or authorized integrations explicitly included in the plan. For example, the official Claude pricing page lists API credits for some plans. Read the eligibility conditions to understand the actual benefit.
OpenAI’s Sign in with ChatGPT documentation also distinguishes login permissions from API access using a ChatGPT plan. A login button does not mean any service can use someone’s subscription without limits. Check app eligibility, user-authorized scope and applicable conditions.
2. An API price list contains more than one rate
OpenAI’s pricing documentation separates input, caching, output and tool charges. Claude’s API pricing documentation distinguishes cache writes from reads and standard requests from batch processing. Gemini’s pricing documentation details model-specific free and paid availability, output categories including reasoning tokens, and caching and search conditions.
Comparing only the price of a million input tokens therefore misses some charges. Align the model, context length, processing method, region and promotional period before calculating billable items. Apply cache terms only to the portion that qualifies. When cache writing or storage is charged, using only the discounted read rate understates the cost.
For each attempt, add uncached input + cache-related costs + billable output + tools and storage, then sum across attempts. Do not add retries again if those attempts are already included; add only costs not yet counted. Likewise, do not add reasoning tokens again if they are already included in billable output. Rather than assuming a fixed conversion from Korean characters to tokens, check actual usage records.
3. A cheap call is not necessarily a cheaply completed task
Our proposed comparison unit is total cost divided by the number of results meeting predefined acceptance criteria. This example shows how review time changes the choice. A and B are not real products. Each makes 100 attempts without retries. For simplicity, only accepted results are reviewed and failed results are discarded.
| Item | A | B |
|---|---|---|
| Cost of 100 calls | $2.00 | $5.00 |
| Accepted results | 60 | 95 |
| Review per accepted result | 3 minutes | 0.5 minutes |
| Review cost | $54.00 | $14.25 |
| Total cost | $56.00 | $19.25 |
| Total cost per accepted result | About $0.93 | About $0.20 |
A · $56.00
B · $19.25
A has cheaper calls, but under these assumptions B delivers accepted results at a lower total cost. The conclusion could change if review is barely needed or B’s results do not meet your criteria. This example is not evidence that expensive models are always better.
In practice, include the time spent identifying failures, retrying, editing, waiting and building integrations where relevant. State how you allocate a fixed monthly subscription across tasks. Local models also require electricity, hardware and operating time: no token bill does not mean zero total cost.
4. Benchmarks are a map, not a guarantee
Artificial Analysis’s intelligence evaluation methodology combines several tests to compare models. A high score alone does not establish that a model is best at Korean administrative documents, music generation or contract review. First check which skills and languages the tests measure.
SWE-bench Verified uses 500 human-reviewed cases. For coding results, examine the agent configuration, tools, attempt count and evaluation environment as well as the model name. Do not treat scores from different systems or test versions as a pure difference in model ability.
Speed needs similar care. Fast output tokens may not finish your task sooner if the answer starts slowly or needs repeated revision. Artificial Analysis’s performance methodology distinguishes time to first token from output speed and other measures. For your own comparison, also measure time from input to completed review.
5. Different work needs different acceptance criteria
| Task | Check first |
|---|---|
| Legal and administrative research | Existence of statutes, cases and notices; jurisdiction, currency, exceptions and source location |
| Reading disclosures and financial statements | Units, periods, consolidated versus standalone accounts, notes and verifiable calculations |
| Education and document writing | Misconceptions, omissions, reader level, editing effort, evidence and explanation |
| Music and social content | Fit for purpose, editability, usage rights and approval for external publishing |
This table does not identify winners in each field. It offers questions for designing small tests with public or your own materials. Do not use it as advice replacing individual legal or financial decisions. Before uploading sensitive contracts, account information or personal data, check permitted uses and data terms. A free tier is not permission to process sensitive information. Read data terms linked to pricing, such as the Gemini terms, too.
6. Before paying, narrow the choice to one recurring task
Write down a frequent task and its acceptance criteria in one sentence: “Extract this PDF’s table with correct units and totals, and let a person locate the original source.” Give the same material to a few accessible candidates and record errors, correction time and cost. If you publish the comparison, include product versions, tools, dates, attempt counts and failures.
“An existing free tool is enough” or “a person should do this directly” are valid conclusions. The criterion is not a higher ranking or a bigger subscription, but whether you can obtain results meeting your standards within an affordable budget and time.
Verification scope and sources
We checked the linked official documentation and evaluators’ methodologies on October 15, 2026. This is not a price list exhaustively verifying every plan and model. Recheck the product’s latest terms before purchasing. The hypothetical calculations and task questions are OIYO editorial examples, not substitutes for hands-on tests or professional suitability assessments.
Read official API prices under matching conditions
Checked 2026-10-23 · USD per million tokens · Standard paid APIs billed directly by providers. Not subscriptions or a quality ranking. Tax, currency conversion, tools and storage excluded.
| Model and official source | Ordinary input | Billable output | Cache reads |
|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 |
| Gemini 4 Argon | $2.00 | $10.00 | $0.10 |
Example: 100,000 uncached input tokens plus 10,000 billable output tokens cost $0.30 for each model. Not an invoice or a quality comparison; equal token counts do not imply equal text lengths.
Check cache, context and promotional conditions
- GPT-6.1 Sol: above 272K input tokens, long-context rates apply to the full request. Fast, Ultrafast, regional premiums and Batch/Flex discounts excluded. Cache-write prices follow. 30m: $2.50; 12h: $5.00
- Claude Sonnet 5.5: cache writes and reads are separate items. Batch, geography and tool charges excluded. Cache-write prices follow. 5m: $2.50; 1h: $4.00
- Gemini 4 Argon: these rates run through 2026-11-30. From December 1, input/output/cache are listed at $4/$20/$0.20. Implicit caching only; no separate explicit cache-write item. Output includes thinking tokens; search costs extra.
Do not count cache, reasoning output or tools twice. Recheck official terms before buying.
GPT-6.1 SolChange the value of review time
Fictional editorial example, not actual product prices or measured performance.
Only passed results are reviewed; failure detection and retries are excluded. A: $2 calls, 60 × 3 min. B: $5 calls, 95 × 0.5 min. Total = calls + passed count × review minutes × rate.
- Call charges
- + Review time × rate
- ÷ Divide by accepted results
Enter a number from 0 to 100. The last valid results remain visible.
A · Total cost:
Cost per passed result:
B · Total cost:
Cost per passed result:
With no review cost, A is about $0.033 and B $0.053 per accepted result. At $0.30/min, they become about $0.93 and $0.20. This fictional example holds the attempt and acceptance counts fixed; it is not a model recommendation.