Prices and benchmarks · checked on 26 September 2026
AI cost compared: the same task costs $0.55 or $5.98
How thoroughly the AI works can be set in 5 levels. On the Opus 5.5 model the same test task costs $0.55 at the lowest level and $5.98 at the most thorough one — 11 times as much. The result only rises from 42 to 58 points. Even the “medium” level reaches 51 points: as many as the previous model, Opus 5, at its highest level, for 77% less.
- Independently measured
- 10 dated references
- As of 26/09/2026
- DE / EN
In short: AI is billed by the amount of text, in so-called tokens. The price per token says little about the bill, though. What counts is how many tokens a task uses — and that is set mainly by the effort level, not by the price list. All figures come from the official price list, from independent benchmarks and from our own logs, retrieved on 26 September 2026.
This page is for anyone who uses AI at work and wants to know what they pay for. What does a token cost? What does a task cost? And which setting really moves the bill? The page does not replace your own costing. It shows the orders of magnitude.
Free and without sign-up. Your inputs never leave your browser.
- Token
- A text fragment of a few characters. This is what you pay for: per million tokens, separately for text read and text written.
- Model
- The version of the AI. Claude comes as Fable, Opus, Sonnet and Haiku; the number is the release. Fable is the most expensive, Haiku the cheapest.
- Effort level
- How thoroughly the model works: 5 levels from “low” to “max”. Higher levels think longer and write more tokens.
- Points and tasks
- The independent benchmark firm Artificial Analysis gives every model the same test tasks. Points show the result; more is better. Cost per task is the average per test task.
The biggest lever
The effort level decides: the same AI costs up to 11 times as much
Every recent Claude model has an effort setting that decides how thoroughly it works — 5 levels from “low” to “max”. It affects everything the model writes: the answer, the thinking before it and work steps such as searching or reading files. Out of the box Opus 5.5 is set to level 2 (“medium”), all other models to level 3 (“high”). If you never change it, this default decides your bill.
| Model and effort | Points | Cost per task | Cost compared | Extra cost vs. level below | Verdict |
|---|---|---|---|---|---|
| Opus 5.5 — recommended by Anthropic, widest range | |||||
| low | 42 | $0.55 | — | baseline | |
| medium (default) | 51 | $1.34 | +144% | worth it | |
| high | 54 | $1.82 | +36% | worth it | |
| xhigh | 56 | $3.46 | +90% | weigh up | |
| max | 58 | $5.98 | +73% | weigh up | |
| Fable 5.1 — for long, hard tasks | |||||
| low | 47 | $2.37 | — | baseline | |
| medium | 49 | $2.98 | +26% | worth it | |
| high (default) | 51 | $3.91 | +31% | worth it | |
| xhigh | 53 | $5.98 | +53% | weigh up | |
| max | 53 | $7.63 | +28% | skip | |
| Opus 5 — the predecessor, still available | |||||
| low | 39 | $1.10 | — | baseline | |
| medium | 45 | $2.19 | +99% | worth it | |
| high (default) | 48 | $3.61 | +65% | weigh up | |
| xhigh | 50 | $4.88 | +35% | worth it | |
| max | 51 | $5.86 | +20% | weigh up | |
| Sonnet 5 — the low-cost all-rounder | |||||
| low | 24 | $0.51 | — | baseline | |
| medium | 28 | $1.00 | +96% | weigh up | |
| high (default) | 32 | $1.79 | +79% | worth it | |
| xhigh | 34 | $2.87 | +60% | weigh up | |
| max | 38 | $5.09 | +77% | worth it | |
| Haiku 4.5 — no effort levels | |||||
| fixed setting | 17 | $0.21 | — | baseline | |
Fable 5.1 reaches the same 53 points at “xhigh” and at “max”. The top level costs 28% more. This is the only place in the table where an extra charge demonstrably buys nothing.
Opus 5.5 at its lowest level reaches 42 points for $0.55. Sonnet 5 at its highest reaches 38 points for $5.09. The higher-priced model is better and 89% cheaper here. It writes about 10,000 tokens per task, Sonnet 5 about 118,000.
Opus 5.5 starts at “medium”, all other models at “high”. Switch from Opus 5 to Opus 5.5 without setting it, and you work one level lower. You get 51 instead of 48 points at 63% less cost per task.
How the verdict is set: Up to 20% extra cost per point gained reads “worth it”, above that “weigh up”. If a level gains no point, it reads “skip”. The threshold is our choice, not a measurement — the figures next to it are measured.
Performance
From $1.34 per task, no tested AI scores higher than Opus 5.5
A low price is no use if the result falls short. The independent benchmark firm Artificial Analysis therefore measures score, cost and time on the same tasks. The result: from $1.34 per task, no measured model from any provider scores higher than Opus 5.5. At its default level it needs 3.3 minutes per task and leads in all 6 measured fields.
- Opus 5.5
- Fable 5.1
- Sonnet 5
- Haiku 4.5 (no effort levels)
Opus 5.5: 42 points for $0.55 up to 58 points for $5.98. Fable 5.1: 47 to 53 points for $2.37 to $7.63. Sonnet 5: 24 to 38 points for $0.51 to $5.09. Haiku 4.5: 17 points for $0.21. All values are listed in the table in the section “The effort level decides”.
Source: Artificial Analysis, Intelligence Index v4.3.2, retrieved 26/09/2026.
Show values as a table
| Model and effort | Time per task | Tokens per second |
|---|---|---|
| Haiku 4.5 (fixed setting) | 2.9 min | 86.2 |
| Opus 5.5 (medium) | 3.3 min | 81.0 |
| Sonnet 5 (high) | 6.6 min | 70.5 |
| Fable 5.1 (high) | 6.9 min | 58.8 |
| Opus 5 (high) | 8.5 min | 58.0 |
Source: Artificial Analysis, Intelligence Index v4.3.2, retrieved 26/09/2026.
- Opus 5.5
- Fable 5.1
- Sonnet 5
Show values as a table
| Field | Opus 5.5 | Fable 5.1 | Sonnet 5 |
|---|---|---|---|
| Finance and accounting | 60.7 | 56.4 | 40.4 |
| Strategy and operations | 63.7 | 59.7 | 39.2 |
| Law | 63.3 | 60.5 | 42.3 |
| Health and medicine | 60.5 | 58.0 | 42.8 |
| Engineering | 60.4 | 56.5 | 40.2 |
| Economics | 65.6 | 62.7 | 49.4 |
Source: Artificial Analysis, field scores for Intelligence Index v4.3.2, retrieved 26/09/2026.
Show or hide providers (15 of 15)
- Line of best-value models
- Best-value model
- Other models
Show values as a table
| Provider | Model and effort | Points | Cost per task | Time per task | Tokens per second |
|---|---|---|---|---|---|
| Anthropic | Opus 5.5 · max | 57.6 | $5.98 | — | — |
| Anthropic | Opus 5.5 · xhigh | 56.0 | $3.46 | 7.5 min | 91.8 |
| Anthropic | Opus 5.5 · high | 53.6 | $1.82 | 4.3 min | 87.4 |
| Anthropic | Fable 5.1 · max | 53.4 | $7.63 | 11.9 min | 69.6 |
| Anthropic | Fable 5.1 · xhigh | 53.2 | $5.98 | 10.6 min | 65.8 |
| OpenAI | GPT-6 Astra (max) | 52.7 | $3.26 | 7.8 min | 58.2 |
| Anthropic | Opus 5.5 · medium | 51.2 | $1.34 | 3.3 min | 81.0 |
| Anthropic | Fable 5.1 · high | 51.2 | $3.91 | 6.9 min | 58.8 |
| Anthropic | Opus 5 · max | 50.8 | $5.86 | 12.6 min | 61.0 |
| Anthropic | Opus 5 · xhigh | 49.7 | $4.88 | 10.8 min | 58.2 |
| Anthropic | Fable 5.1 · medium | 48.9 | $2.98 | 5.1 min | 58.1 |
| Meta | Muse Spark 1.3 (max) | 48.1 | $1.60 | 4.6 min | 226.8 |
| Anthropic | Opus 5 · high | 48.1 | $3.61 | 8.5 min | 58.0 |
| OpenAI | GPT-6 Sol (max) | 47.5 | $1.06 | 6.1 min | 86.2 |
| Anthropic | Fable 5.1 · low | 46.8 | $2.37 | 4.0 min | 59.2 |
| SpaceXAI | Grok 4.7 (xhigh) | 46.4 | $3.74 | 15.9 min | 47.5 |
| Xiaomi | MiMo-V2.6-Pro (open weights) | 46.3 | $0.13 | 23.4 min | 44.7 |
| Alibaba | Qwen3.8 Max (0902) | 45.4 | $5.41 | 33.3 min | 39.7 |
| Z AI | GLM-5.3 (max) (open weights) | 44.8 | $2.01 | 13.6 min | 69.0 |
| Anthropic | Opus 5 · medium | 44.8 | $2.19 | 5.5 min | 58.2 |
| SpaceXAI | Grok 4.6 (high) | 44.3 | $1.86 | 8.9 min | 68.5 |
| StepFun | Step 5 Preview | 43.7 | $0.72 | 12.2 min | 81.4 |
| Kimi | Kimi K3 (max) (open weights) | 43.6 | $2.00 | 21.0 min | 35.2 |
| Anthropic | Opus 5.5 · low | 42.3 | $0.55 | 1.3 min | 80.9 |
| Z AI | GLM-5.3-Flash (open weights) | 41.8 | $0.25 | 19.7 min | 44.8 |
| Gemini 3.8 Flash (high) | 40.9 | $1.24 | 3.5 min | 338.2 | |
| DeepSeek | DeepSeek V4.1 Flash (max) (open weights) | 39.5 | $0.27 | 4.4 min | 236.1 |
| Anthropic | Opus 5 · low | 39.4 | $1.10 | 2.5 min | 59.6 |
| Anthropic | Sonnet 5 · max | 38.2 | $5.09 | 14.3 min | 82.1 |
| OpenAI | GPT-6 Luna (max) | 37.3 | $0.07 | 5.7 min | 146.2 |
| Anthropic | Sonnet 5 · xhigh | 34.4 | $2.87 | 9.3 min | 74.1 |
| Alibaba | Qwen3.8 27B (xhigh) (open weights) | 33.7 | $1.01 | 18.4 min | 43.9 |
| Anthropic | Sonnet 5 · high | 31.7 | $1.79 | 6.6 min | 70.5 |
| MiniMax | MiniMax-M3 (open weights) | 29.2 | $0.51 | 4.4 min | 179.4 |
| Anthropic | Sonnet 5 · medium | 28.1 | $1.00 | 4.2 min | 70.5 |
| Thinking Machines | Inkling (open weights) | 25.0 | $0.61 | 2.9 min | 167.1 |
| Anthropic | Sonnet 5 · low | 24.3 | $0.51 | 2.5 min | 64.3 |
| NVIDIA | Nemotron 3 Ultra (open weights) | 22.9 | $0.55 | 3.1 min | 145.3 |
| Gemini 3.5 Flash-Lite | 22.2 | $0.12 | 0.8 min | 355.9 | |
| Meta | Muse Glimmer (high) (open weights) | 17.5 | $0.06 | 2.3 min | 104.4 |
| Anthropic | Haiku 4.5 | 16.9 | $0.21 | 2.9 min | 86.2 |
| Mistral | Mistral Medium 3.5 (open weights) | 14.2 | $0.44 | 2.9 min | 144.1 |
Source: Artificial Analysis, Intelligence Index v4.3.2, default selection and model pages, retrieved 26/09/2026. Time per task = pure writing time.
Worked example: coding
Having an app built: a solved task costs $15 with Opus 5.5, $82 with Fable 5
The points above are an average across many fields. In coding the differences are especially large. If you let AI build an app, you also pay for attempts that fail. What counts is the cost per solved task. In the independent coding benchmark Terminal-Bench 4.0, Opus 5.5 at “xhigh” solves 59.6% of the tasks. That costs $14.73 per solved task on average. Opus 5 needs $39.37, Fable 5 $82.28.
How it was measured: Terminal-Bench 4.0 consists of 66 tasks, 18 of them from software development, plus science, machine learning, IT operations, hardware, security and media. Example: speed up the start of a payment service so that overdraft alerts arrive complete and within 5 seconds. The AI works on its own in an isolated test environment, through text commands only, for up to 8 hours per task. A task counts as solved only if all automated tests pass. Artificial Analysis runs each task 3 times and measures cost, tokens and time.
- Opus 5.5 (xhigh)
- Opus 5 (max)
- Fable 5 (max)
Show values as a table
| Model and effort | Solved | Cost per task | Cost per solved task | Tokens written | Time |
|---|---|---|---|---|---|
| Opus 5.5 (xhigh) | 59.6% | $8.78 | $14.73 | 169,497 | 19.5 min |
| Opus 5 (max) | 49.0% | $19.29 | $39.37 | 205,110 | 35.5 min |
| Fable 5 (max) | 42.4% | $34.91 | $82.28 | 216,734 | 32.2 min |
Source: Artificial Analysis, Terminal-Bench 4.0 in Intelligence Index v4.3.2, values per task, retrieved 26/09/2026. Cost per solved task calculated by us.
- Opus 5.5
- Opus 5
- Fable 5 (only measured at max)
Opus 5.5 “medium”: 52.5% solved for $4.04 per task. Opus 5 “max”: 49.0% for $19.29. Fable 5 “max”: 42.4% for $34.91.
Show values as a table
| Model and effort | Solved | Cost per task |
|---|---|---|
| Opus 5.5 (low) | 31.3% | $2.08 |
| Opus 5.5 (medium) | 52.5% | $4.04 |
| Opus 5.5 (high) | 56.6% | $5.12 |
| Opus 5.5 (xhigh) | 59.6% | $8.78 |
| Opus 5.5 (max) | 59.6% | $13.11 |
| Opus 5 (low) | 26.3% | $5.18 |
| Opus 5 (medium) | 34.3% | $9.18 |
| Opus 5 (high) | 46.0% | $13.14 |
| Opus 5 (xhigh) | 46.5% | $17.35 |
| Opus 5 (max) | 49.0% | $19.29 |
| Fable 5 (max) | 42.4% | $34.91 |
Source: Artificial Analysis, Terminal-Bench 4.0, retrieved 26/09/2026. Vendor claim: Anthropic, Opus 5.5 announcement of 22/09/2026.
Opus 5.5 solves 59.6% at both “xhigh” and “max”. “Max” costs $13.11 instead of $8.78 per task and writes 253,000 instead of 169,000 tokens.
Fable 5 writes more tokens per task than Opus 5.5 at “xhigh” (217,000 instead of 169,000) and pays 2.5 times the token price. It solves fewer: 42.4% instead of 59.6%.
Requests the provider flags as sensitive, such as IT security tasks, are answered by an older model on Opus 5.5 and Fable 5. The measurements include this. At Vals.ai it affected 30 of 198 attempts; counting them as failures, Opus 5.5 still leads Opus 5 by 53.5% to 45.5%.
4 measurements, the same ranking: Opus 5.5 ahead of Opus 5 ahead of Fable 5
| Source | AI working program | Opus 5.5 | Opus 5 | Fable 5 |
|---|---|---|---|---|
| Artificial Analysis (independent) | mini-swe-agent | 59.6% | 49.0% | 42.4% |
| Vals.ai (independent) | Terminus 2 | 61.6% | 45.5% | — |
| Snorkel (independent) | Claude Code | — | 53.9% | 44.5% |
| Anthropic (vendor) | not stated | 66.4% | 52.3% | — |
“—” means: not measured there. Snorkel measures Opus 5 at “xhigh”, Anthropic Opus 5.5 at “xhigh”; the other values apply to “max” or to the source's standard setting.
Base prices
What the AI writes costs 5 times as much as what it reads
All the costs above come from a handful of prices. Billing is per token — a piece of text, around 4 characters in English. Text the AI reads (your question, attached documents, the conversation so far) is much cheaper than text it writes. All prices per million tokens in US dollars, retrieved on 26 September 2026.
| Model | Newly entered | Written by the AI | Stored for later | Reused | Max. text |
|---|---|---|---|---|---|
| Fable 5.1 | $10.00 | $50.00 | $12.50 | $0.25 | 1M |
| Opus 5.5 | $4.00 | $20.00 | $5.00 | $0.20 | 1M |
| Opus 5 | $5.00 | $25.00 | $6.25 | $0.50 | 1M |
| Sonnet 5 | $2.00 | $10.00 | $2.50 | $0.20 | 1M |
| Haiku 4.5 | $1.00 | $5.00 | $1.25 | $0.10 | 200K |
| Opus 5.5, fast mode | $8.00 | $40.00 | $10.00 | $0.40 | 1M |
2 discounts come on top: if you collect tasks and have them processed as a batch, you pay half. If you need the same text several times, reading it again costs only 10% of the input price. On Opus 5.5 it is 5%, on Fable 5.1 just 2.5%.
Cost traps
4 cost traps no price list mentions
Besides price and effort level, 4 things push up the bill without anyone noticing.
Models from Claude 4.7 onwards split the same text into about 30% more tokens than their predecessors. An identical price per token therefore does not mean the same price per page of text. Anyone comparing old and new models has to allow for this.
Without a setting, Opus 5.5 works one level lower than Opus 5. Switch without setting effort, and you get a different bill and a different result. Set the effort level explicitly whenever you change models.
Fast mode delivers answers up to 2.5 times faster — at twice the token price. It does not make the model smarter, only faster. For work that runs in the background it is money thrown away.
In long sessions the AI reads known text from a store instead of processing it again. Without that, our own work would have cost 6.3 times as much. Important: changing the effort for the whole request or switching to fast mode empties this store, and you pay full price again. Changing the effort for single messages only keeps it — available on Fable 5.1, Opus 5.5 and Opus 5.
Our own figures
Our own AI use: 6.3 times cheaper by reusing known text
How big these levers are in practice is shown by our own use. Teddynews uses Claude every day. We analysed all our work sessions from 20 April to 25 September 2026. At today's price list this use would have cost $30,584 — that is what you pay if every request is billed by usage. We ourselves pay a subscription with a fixed monthly price.
The main saving comes from reuse. In long conversations the program resends the whole history with every new question. The AI then reads text it already knows from a store instead of processing it again — for 2.5 to 10% of the normal price. Without this reuse the same work would have cost $192,754, 6.3 times as much.
- Reused (read from store)
- Stored for later
- Written by the AI
- Newly entered
Show values as a table
| Item | Tokens | Share of tokens | Cost | Share of cost |
|---|---|---|---|---|
| Reused (read from store) | 32,159.3M | 96.77% | $16,503 | 53.96% |
| Stored for later | 956.5M | 2.88% | $10,866 | 35.53% |
| Written by the AI | 113.8M | 0.34% | $3,197 | 10.45% |
| Newly entered | 3.6M | 0.01% | $19 | 0.06% |
| Total | 33,233.2M | 100% | $30,584 | 100% |
Source: our own Claude Code logs, 20/04–25/09/2026, valued with the provider's price list of 26/09/2026.
Total 20 April to 25 September: $30,584 · 72,192 AI responses · 33.2 bn tokens
Show values as a table
| Month | AI responses | Tokens | At list price |
|---|---|---|---|
| April | 3,564 | 1.24 bn | $1,093 |
| May | 8,898 | 1.22 bn | $1,023 |
| June | 13,637 | 6.90 bn | $6,577 |
| July | 11,551 | 5.94 bn | $6,748 |
| August | 12,646 | 6.57 bn | $5,924 |
| September | 21,896 | 11.37 bn | $9,219 |
| Total | 72,192 | 33.23 bn | $30,584 |
Source: our own Claude Code logs, 20/04–25/09/2026, valued with the provider's price list of 26/09/2026.
How we counted: The basis is the Claude Code logs on the Teddynews work computer; each separate session log counts as one session. Every AI response counts once, even if the log writes it several times; the final value applies. Each response is valued at today's price list of its model; storing text at the rate for the storage time shown in the log. Not included: sessions before 20 April 2026 (no logs exist for them), other tools and other computers. For 0.36% of the amount the log does not say whether the fast mode at twice the price was used; we counted the normal price.
Calculator
Calculator: what your AI use costs on 5 models
Estimate your own use with the same 4 items as in our own bill. The example is a longer working session. Everything is calculated in your browser — no input is transmitted.
Questions
7 answers on subscriptions, effort levels, tokens and sources
Does this also apply to subscriptions such as ChatGPT Plus or Claude Pro?
No. A subscription has a fixed price; you do not pay per task. The figures on this page apply to usage-based billing, as it arises when AI is built into your own software. For subscribers they are still a yardstick: they show how quickly a usage allowance is used up.
Why is the most expensive level not always the best?
Because extra thinking stops adding value at some point. For Fable 5.1 this point lies between “xhigh” and “max”: same score, 28% more cost. For Opus 5.5 the top level still adds 2 points but costs 73% more. For very long tasks it can still make sense — then for stamina, not for quality.
What changes with Opus 5.5 compared with Opus 5?
3 things. The token price is 20% lower: $4 instead of $5 per million input tokens, $20 instead of $25 for output. Reused text costs 5% instead of 10% of the input price. And without a setting, Opus 5.5 works at “medium” instead of “high”. Opus 5 remains available for now.
What is a token, in one sentence?
A token is a piece of text; in English it is about 4 characters long, according to the provider. Models from Claude 4.7 onwards split the same text into about 30% more tokens.
How do I get my own usage figures?
Every bill lists the 4 items from the calculator separately: newly entered, reused, stored for later, written by the AI. Developer tools usually have their own command for it, web interfaces a usage page in the account.
Are the scores from Anthropic itself?
No. They come from Artificial Analysis, an independent benchmark firm that rates models using a fixed method of 10 evaluations. The prices come from the official price list. Both are linked below with their retrieval date.
Why does Teddynews publish this?
Because we use AI in our own tools and had to do the same calculations. Who is behind it and what they sell is stated openly in the section Who writes this page.
Disclosure
Who writes this page — and what they sell
Teddynews builds AI tools — including AI agents for businesses (German) — and calculates with the same prices. Our own figures in the section “Our own AI use” come from this work. The page costs nothing, collects no address and recommends no provider.
Anyone who publishes something like this usually has something to sell. That is true here too. So it is stated openly, with its real status:
Your documents become a tool that pre-sorts enquiries — instead of a PDF to download.
In preparation. €149 is a proposal, not yet billable.
10 Instagram cards a month, built from your reader's own words. With a forecast of which ones get shared.
In preparation. €79 is a proposal, not yet billable.
Made the same way as this page: electricity and AI, the training market, the reasons for AI scepticism. All with sources.
Free, without sign-up.
No newsletter, no call-back, no deadline. If you have a question, write to office@teddynews.at. A person answers, not a queue.
Where the figures come from
Sources: 10 references with origin and retrieval date
All prices are official — they come from the provider. Scores, cost, tokens and time per task are independently measured — measured by independent benchmark firms, not confirmed by the provider; vendor claims are named as such. The figures in the section “Our own AI use” are our own count — counted from our logs. No value on this page is estimated; calculated values such as the cost per solved task are marked as such.
- Price list and billing rules, retrieved 26/09/2026 — provider documentation · pricing overview (both checked against each other)
- Maximum text per request (context window) and default effort per model, retrieved 26/09/2026 — provider model overview
- Scores and cost per task, Intelligence Index v4.3.2 (10 evaluations), retrieved 26/09/2026 — Artificial Analysis. “Cost per task” there is the weighted average of the tokens actually used.
- Effect of the effort setting on token use, level names, defaults and per-message effort changes — provider documentation on effort
- Fast mode: up to 2.5 times the output speed, twice the price, same capabilities — provider documentation on fast mode
- Coding benchmark Terminal-Bench 4.0: structure and time limit — tbench.ai · method and values per task — Artificial Analysis · cross-check — Vals.ai and Snorkel, all retrieved 26/09/2026
- Vendor claims on Opus 5.5 (coding benchmarks, forwarding of sensitive requests to an older model), published 22/09/2026 — Anthropic
- Market comparison of 42 models: default selection and model pages, retrieved 26/09/2026 — Artificial Analysis
- Our own figures: Claude Code logs on the Teddynews work computer, 20/04–25/09/2026, counted on 26/09/2026
- New way of splitting text from Claude 4.7 (about 30% more tokens for the same text) and discounts — provider documentation, sections price list and prompt caching
How this page was made: Research and text were created with AI assistance and checked by Teddynews (disclosure under Art. 50 of the EU AI Act). Data as of 26/09/2026; prices and measurements were retrieved the same day.
For the cost per task and effort level there is currently no second independent measurement; Artificial Analysis is the only benchmark firm measuring this way. Prices change. This page therefore gives the retrieval date for every figure. If you use it as the basis for a costing, please check the current state at the linked source.