Which AI model should you use, how hard should it think, and what will it cost? Start with the picker below, then explore how Claude, ChatGPT and Gemini count tokens, bill for them and reset your limits.
Answer three quick questions. You'll get a recommended model from each provider, the effort level to set, why it fits, and roughly what one request costs. Change any answer and the advice updates instantly.
Balanced picks. Go one tier cheaper if answers are already right; go one tier up (or raise effort) if the model misses steps or you keep re-asking.
When an AI "works" on something it doesn't answer once. It goes round a loop: re-read everything → think → call a tool → add the result, until the job is done. Each trip round the loop is a new, separately billed API call. Press Play and follow the moving dot.
Each button runs the current task several ways instantly and lines up the results below. Your task, model and settings are the starting point.
Input grows every step. Call 8 re-reads calls 1–7, so on agent work input tokens usually dwarf output.
Thinking is paid but mostly not kept. Reasoning tokens are billed as output, but most of them don't stay in later context.
Tool results are the heavy part. One web page or source file can be 10–30k tokens. Sub-agents and trimming keep them out of the main loop.
Steps, snippets, token sizes and the quality score are illustrative. Prices are the real per-token rates from the Models tab.
A benchmark is a standard exam given to every model in the same way, so their scores can be compared. We use one independent source for all 11 models: the Artificial Analysis Intelligence Index (v4.3, updated 7 Sep 2026). It combines 10 tests of agent work, coding, general knowledge and scientific reasoning, and it tests each model at every effort level. The higher the score, the more capable the model.
Source: Artificial Analysis Intelligence Index v4.3. "Best" is each model's top published score. Gemini 3.1 Pro, Gemini 3.5 Flash-Lite and Claude Haiku 4.5 were tested at one standard setting only. Claude Fable 5.1 was tested "with fallback".
Each dot is one model at one effort level; lines join the same model from low to max effort. Cost is Artificial Analysis's average cost per task to run their test suite at list prices. The horizontal axis is logarithmic: each gridline is 10× more expensive.
Score (cost per task). Darker cells score higher. "—" means that level wasn't tested or the model doesn't offer it.
Individual tests Artificial Analysis highlighted in its GPT-6 Astra review (Sep 2026).
Hallucination rate: how often the model gives a wrong answer instead of saying it doesn't know. Lower is better.
Four areas, weighted Agents 30%, General 30%, Coding 20%, Scientific reasoning 20%, built from 10 tests:
| GDPval-AA | Realistic office work tasks across many professions |
| AutomationBench-AA | Agents completing business workflows |
| Terminal-Bench 4.0 | Getting real work done in a command line (66 hard tasks) |
| SciCode | Writing code for scientific problems |
| Humanity's Last Exam | Expert-level questions across many fields |
| CritPt | Research-level physics reasoning |
| AA-Omniscience | Knowledge, and knowing when to say "I don't know" |
| AA-LCR | Reasoning over long documents |
| AA-Briefcase, GDP.pdf | Artificial Analysis's newer work-task tests (see their methodology page) |
A model that tops a leaderboard can still be worse at your job. Try the top 2–3 candidates on 10–20 real examples of your work before committing.
Launch posts show each company's own best-case numbers, often on different settings. Independent indexes run every model the same way, which is why this page uses one.
The index changed version twice in August–September 2026 and new models arrive monthly. Treat these as a September 2026 snapshot.
Models never see letters or words. Text is chopped into tokens: common words become one piece, rare words split into several, and every space, comma and emoji counts. In English, 1 token ≈ 4 characters ≈ ¾ of a word. Every provider bills and limits you by tokens, so this is the unit behind everything else on this page.
This is a teaching approximation of byte-pair encoding. Real tokenizers differ in detail; use each provider's token-counting endpoint for exact numbers.
| Thing | ≈ tokens |
|---|---|
| One English word | 1.3 |
| A short chat message | 30–80 |
| One page of text (500 words) | ≈ 670 |
| A 30-page PDF | 20k–25k |
| An average web page (10 kB) | ≈ 2,500 |
| A novel (90k words) | ≈ 120k |
| 1M-token context window | ≈ 750k words |
API bills are simple arithmetic: input tokens × input price + output tokens × output price, with prices quoted per million tokens (MTok). Three things change the answer a lot: thinking tokens (billed as output), cached input (about 90% off) and the Batch API (50% off if you can wait).
Thinking tokens are estimated from the effort level with illustrative multipliers (see Effort levels tab). Models that don't support the chosen level use their nearest one. Long-context surcharges are applied where a provider charges them (Gemini Pro above 200k, OpenAI above 272k).
Reasoning models "think" before answering. Those thinking tokens are usually hidden, but they are billed as output and count against your limits. The effort level is a dial for how much thinking (and how many tool calls) the model is allowed. Higher effort means better answers on hard problems, and slower, pricier answers on everything.
Summarised from each provider's documentation (September 2026). "Default" is what you get if you don't set anything.
Token counts are illustrative. Effort is a behavioural signal, not a hard budget: the model still thinks hard on genuinely difficult problems at low effort, just less than it would at high.
| Provider | Setting | Levels | Default |
|---|---|---|---|
| ClClaude API | output_config.effort | low · medium · high · xhigh · max xhigh on Fable, Opus 4.7+, Sonnet 5. max on Fable, Opus 4.6+, Sonnet 4.6+. Haiku 4.5 has no effort setting. | high |
| AIOpenAI API | reasoning.effort | none · minimal · low · medium · high · xhigh · max Model-dependent. GPT-6 Astra: low → max only ("none" returns an error). GPT-5.6 family: none → max. | model-dependent (GPT-5.5: medium) |
| AIChatGPT app | Model picker | Instant · Medium · High · Extra High · Pro Since 14 Sep 2026, Instant no longer switches to thinking on its own for Plus and Pro. | Instant |
| GGemini API | thinking_level | minimal · low · medium · high Model-dependent. 3.1 Pro: low → high. 3.8 Flash: low → high ("minimal" errors). Flash-Lite and 3.5 Flash also accept minimal ≈ no thinking. | 3.1 Pro: high · 3.8 / 3.5 Flash: medium |
| GGemini app | Thinking level | Standard · Extended · Deep Think (AI Ultra, Pro model only) | Standard |
Subscription apps don't give you a token number. All three now use the same two-layer idea: a rolling 5-hour window that starts with your first message, and a weekly ceiling on top. Heavy models, high effort, long chats and big files drain both faster. The API is different again: it bills per token and limits speed (requests and tokens per minute) instead.
Model × effort × chat length sets how many "units" each message costs; the plan sets units per window. Try Frontier + Max + big files on a Pro plan, then switch to Max 20× or Small model to see the difference.
Rolling (5-hour session): the clock starts at your first message, not on the hour. Wait 6 hours and you do not get a double allowance; unused capacity doesn't carry over.
Fixed (weekly): resets at one moment each week regardless of what you did. Hitting it on Wednesday can mean waiting until next week, even though your 5-hour window keeps "refreshing".
On the API you pay for every token, so there's no weekly allowance. Instead you get speed limits: RPM (requests per minute), ITPM / OTPM or TPM (tokens per minute) and sometimes requests per day. They grow as you move up usage tiers. Hitting one returns a 429 error for a few seconds, not hours.
Models have no memory between messages. Every turn re-sends the whole conversation so far as input. Turn 30 pays for turns 1–29 again, so total cost grows roughly with the square of the length. This is why limits drain faster in long threads, and why prompt caching exists.
With caching, the unchanged start of the prompt is stored for a few minutes (Claude: 5 min or 1 h). The next turn reads it at about 10% of the input price. Writing new content to the cache costs a little more than normal input (Claude and OpenAI: 1.25×). Here we assume each turn reuses the previous turn's cache.
Practical takeaways: start a fresh chat when the topic changes, summarise long threads, and keep big files at the top of the prompt so they stay cacheable.
Every provider sells a ladder: a small, fast model, a mid workhorse and a frontier model. Moving up the ladder buys fewer mistakes on hard, long or ambiguous work, at 3–50× the price. Each card says what the model is for, when to pick something else, and which effort level to start with.
"Benchmark score" dots come from the independent Artificial Analysis index (Benchmarks tab). Speed and value dots are relative judgements from price and positioning. Note: Gemini 3.1 Pro (preview) scored below Gemini 3.8 Flash in those tests; pick it mainly for video, audio and image understanding.
The vocabulary you'll meet on pricing pages and usage screens.