In June 2023 a federal judge fined two New York lawyers $5,000 for a brief full of cases ChatGPT had made up. At the prices verified for this article, $5,000 buys about 67,000 runs of a ten-page document summary on the priciest OpenAI or Anthropic tier, and about 6.7 million runs on the cheapest.
Most pricing comparisons stop at the per-token rate. This one prices whole tasks, shows where the per-token rate misleads, and attaches real dollar figures to what goes wrong when outputs go unchecked or keys go unprotected.
How to read this. Every price comes from a first-party page (OpenAI, Anthropic, Google), read on 12 October 2026. Vendors reprice every few weeks, so the date is part of the data. The Task A, B and C token counts below are my assumptions, not measurements, so run your own prompts through each vendor’s tokenizer before trusting any dollar figure. Confidence labels: High = first-party page, court or tribunal record; Medium = reputable secondary reporting; Low = single secondary source.
If you only remember three things
- Price per token is not cost per task. In Artificial Analysis’s own data, Claude Opus 5.5 costs twice what GPT-6.1 Sol does per token but about 6× as much per task, because it writes about 2.5× as much. Measure how much a model writes for your job.
- At list price, OpenAI and Anthropic are identical at three tiers. Google undercuts them at the top and middle and costs more at the bottom. Choose on caching, token use and your own error rate, not the sticker.
- The expensive failures here cost CA$812 to $82,314, and none came from a model’s price. They came from unchecked output and an unprotected key. Verify what matters and cap your spend.
Cost per task, October 2026
Cost per task is (input tokens × input price + output tokens × output price) ÷ 1,000,000. The token counts are my assumptions, not measurements. Replace them with counts from your own prompts.
- Task A: summarise a ten-page document. 5,000 tokens in, 500 out.
- Task B: analyse a long report. 50,000 in, 2,000 out.
- Task C: one agentic run. 500,000 tokens in and 50,000 out across all turns, with every individual prompt under the long-context thresholds. The last column assumes 450,000 of the input tokens are cache reads and 50,000 are fresh input. It ignores the one-time cache-write premium and Google’s hourly storage fee.
| Model (input / output per 1M tokens) | Task A | Task B | Task C | Task C, cached |
|---|---|---|---|---|
| GPT-6 Astra ($10 / $50) | $0.075 | $0.60 | $7.50 | $3.45 |
| Claude Fable 5.1 ($10 / $50) | $0.075 | $0.60 | $7.50 | $3.11 |
| Claude Opus 5.5 ($4 / $20) | $0.030 | $0.24 | $3.00 | $1.29 |
| Gemini 3.1 Pro Preview ($2 / $12, prompts ≤200K) | $0.016 | $0.124 | $1.60 | $0.79 |
| GPT-6.1 Sol ($2 / $10) | $0.015 | $0.12 | $1.50 | $0.645 |
| Claude Sonnet 5.5 ($2 / $10) | $0.015 | $0.12 | $1.50 | $0.645 |
| Gemini 3.8 Flash ($0.75 / $3.75, introductory) | $0.0056 | $0.045 | $0.56 | $0.26 |
| Gemini 3.5 Flash-Lite ($0.30 / $2.50) | $0.0028 | $0.020 | $0.275 | $0.15 |
| GPT-6 Luna ($0.10 / $0.50) | $0.00075 | $0.006 | $0.075 | $0.035 |
| Claude Haiku 5.5 ($0.10 / $0.50, prompts ≤100K) | $0.00075 | $0.006 | $0.075 | $0.035 |
Confidence: High for prices (first-party pages in Sources). The arithmetic is mine and shown above.
Four things in that table matter more than the rankings.
At list price, OpenAI and Anthropic are identical. Astra and Fable 5.1 are both $10 / $50. GPT-6.1 Sol and Sonnet 5.5 are both $2 / $10. Luna and Haiku 5.5 are both $0.10 / $0.50. If you are choosing between those two vendors, the sticker price will not decide it.
Google is cheapest at the top and in the middle, and dearest at the bottom. Gemini 3.1 Pro’s input price is a fifth of Astra’s, and Task A costs about 4.7× less. Gemini 3.8 Flash runs about 2.7× cheaper than Sol or Sonnet until its introductory price ends on 31 December 2026, then about 1.3× cheaper. At the budget tier the order flips: Flash-Lite costs about 3.7× what Luna or Haiku do on Task A.
The widest gap is between tiers, not vendors. Per token, the top tier costs 100× the budget tier.
Caching roughly halves the agent run, and then output dominates. Cached, Sonnet 5.5’s Task C costs $0.645, of which $0.50 (78%) is output. Once input is cheap, the lever left is how much the model writes.
The price book
| Model | Input | Cached read | Output |
|---|---|---|---|
| GPT-6 Astra | $10 | $1.00 | $50 |
| GPT-6.1 Sol | $2 | $0.10 | $10 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
| Claude Fable 5.1 | $10 | $0.25 | $50 |
| Claude Opus 5.5 | $4 | $0.20 | $20 |
| Claude Sonnet 5.5 | $2 | $0.10 | $10 |
| Claude Haiku 5.5 (≤100K prompts) | $0.10 | $0.01 | $0.50 |
| Gemini 3.1 Pro Preview (≤200K prompts) | $2 | $0.20 + storage | $12 |
| Gemini 3.8 Flash (to 31 Dec 2026) | $0.75 | $0.075 + storage | $3.75 |
| Gemini 3.5 Flash-Lite | $0.30 | $0.03 + storage | $2.50 |
Sources: OpenAI pricing and GPT-6.1 Sol page (Astra and Luna cached rates from OpenAI’s GPT-6.1 Sol announcement); Anthropic pricing; Google Gemini API pricing. Confidence: High.
Footnotes you need before relying on it:
- Batch processing is 50% off at all three vendors. OpenAI’s table lists GPT-6.1 Sol at $1 / $5 in Batch, and Anthropic’s and Google’s pages state the same discount.
- Long prompts reprice the whole request. Past 272K input tokens OpenAI charges 2× input and 1.5× output. Gemini 3.1 Pro moves to $4 / $18 past 200K. Haiku 5.5 moves to $0.50 / $2.50 past 100K.
- Gemini 3.8 Flash doubles to $1.50 / $7.50 on 1 January 2027. Google’s page also lists Gemini 3.1 Pro as a Preview model, which carries more risk of changed terms than a general-availability one.
- Cache write premiums differ. OpenAI charges 1.25× the input rate for writes. Anthropic charges 1.25× for a five-minute write and 2× for one hour. Google charges hourly storage instead ($4.50 per million tokens per hour on Pro).
- Cached reads are not equal at the flagship tier: $1.00 per million on Astra against $0.25 on Fable 5.1. That is why Fable edges Astra on cached Task C even at identical list prices.
- Tools bill on top of tokens. Anthropic charges $10 per 1,000 web searches and $0.08 per Managed Agents session-hour. Google gives 5,000 grounded searches free per month, then $14 per 1,000.
Why the per-token price misleads
Output costs about five times input almost everywhere, and more at the edges. Eight of the ten models above sit at exactly 5×. Gemini 3.1 Pro is 6×. Gemini 3.5 Flash-Lite is 8.3×. A task that writes a lot is priced by its output.
The same job takes a different number of tokens on different models, and list prices hide this. Artificial Analysis runs the same ten-evaluation suite on every model and publishes cost per task. In its comparison at medium effort, GPT-6.1 Sol scored 48 at $0.21 per index task, using about 15 million output tokens for the whole suite. Claude Opus 5.5 scored 51 at $1.34 per task, using about 38 million. Opus costs twice as much per token and about 6.4× as much per task, because it wrote about 2.5× as much (Artificial Analysis comparison; High).
The gap widens further up the effort ladder. Sol at high effort against Opus 5.5 at xhigh: $0.32 against $3.46 per task (about 11×), 25 million against 100 million output tokens, and index scores of 50 and 56 (comparison; High). Artificial Analysis’s own write-up puts Opus 5.5 at max effort at about 119,000 output tokens per index task, against about 27,000 for GPT-6 Astra (Opus 5.5 analysis; High).
Opus scores higher at each matched setting, so this is not a like-for-like quality comparison, and the index averages ten evaluations that may look nothing like your workload. The point is not that one model is wasteful. A per-token price tells you little until you know how much the model writes for your task. The pattern is not new: when GPT-5.5 launched, Artificial Analysis reported a doubled per-token price offset by roughly 40% fewer tokens, for a net rise of about 20% (April 2026 analysis; High).
Thinking is billed as output. Google’s page labels its output price as “including thinking tokens,” so reasoning-heavy settings multiply the output you pay for. Check each vendor’s reasoning documentation before estimating.
Token counts differ by tokenizer. Anthropic states that Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text” than earlier Claude models. Tokenizers also differ across vendors, so per-token is not per-word. Run each vendor’s token-counting endpoint on your real prompts.
Three failures with a price tag
Each case has a tribunal, court or reputable reporting trail. Cases I could not source that way are left out. The first two are failures of output, and the third is a failure of spend control.
1. Air Canada’s chatbot: CA$812.02
A customer asked Air Canada’s website chatbot about bereavement fares. It told him he could apply retroactively, which contradicted the policy page it linked to. The BC Civil Resolution Tribunal found negligent misrepresentation and awarded CA$650.88 in damages, CA$812.02 in total with interest and fees. Air Canada had argued the chatbot was a separate entity responsible for its own actions. The tribunal rejected that, saying the chatbot is part of the website and the airline answers for what its website says (Moffatt v. Air Canada, 2024 BCCRT 149).
Confidence: High (tribunal decision). It is a small-claims ruling, not appellate precedent, so the reasoning matters more than the amount.
2. Invented case law: $5,000 and a dismissed suit
In Mata v. Avianca, lawyers filed a brief citing non-existent opinions generated by ChatGPT. Judge P. Kevin Castel fined the firm and two lawyers $5,000 and dismissed the underlying suit as time-barred. He said they had abandoned their responsibilities by standing by the fake opinions after the court questioned them (Bloomberg Law; opinion at 678 F. Supp. 3d 443, S.D.N.Y. 2023, copy).
Confidence: High. The lawyers also had to send the opinion to each judge falsely named as an author of the fake rulings.
3. A stolen Gemini key: $82,314.44 in 48 hours
On 11–12 February 2026, a small team’s Google Cloud API key was abused to run up $82,314.44 in charges, mostly on Gemini 3 Pro image and text endpoints. The team’s normal bill was about $180 a month, roughly 455 times lower. According to The Register’s report of the developer’s account, Google support cited its shared-responsibility model, and Google declined to say whether it would waive the charges (The Register, 3 March 2026).
Confidence: Medium-High (one developer’s account, independently reported; I could not find how the dispute ended). Reports also note that Google API keys can default to “unrestricted,” so a key created for another service can gain Gemini access once the API is enabled.
What the three share
In none of them did the token cost of the bad output matter. Two came from AI-generated text that nobody checked before it was relied on. The third came from a credential that could spend without limit. The cases involve OpenAI and Google products because those are the ones I found with documentation. That says nothing about the safety of any other vendor.
The rules now
The three failures above happened under existing law: negligence, professional-conduct rules and ordinary contracts. New AI-specific rules are arriving as well, and they add a compliance cost that no vendor’s price list includes.
European Union. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It pushed back the high-risk deadlines and left the rest of the AI Act’s timeline in place:
| Date | What applies |
|---|---|
| 2 Feb 2025 | Prohibited practices and AI-literacy duties |
| 2 Aug 2025 | General-purpose AI model obligations |
| 2 Aug 2026 | Article 50 transparency duties: telling people they are dealing with AI, labelling generated content |
| 2 Dec 2026 | Grace period ends for machine-readable marking on generative systems already on the market |
| 2 Dec 2027 | Stand-alone high-risk systems (Annex III: hiring, credit scoring, biometrics, education, critical infrastructure), moved from 2 Aug 2026 |
| 2 Aug 2028 | High-risk AI embedded in regulated products (Annex I), moved from 2 Aug 2027 |
| 2 Aug 2030 | AI systems intended for use by public authorities |
Sources: Orrick and the Cloud Security Alliance. Secondary sources cite fines of up to €15 million or 3% of global turnover for transparency violations, and up to €35 million or 7% for prohibited practices.
If you build a chatbot or content tool on a vendor’s API, you will often count as the provider or deployer of that system, so these duties can land on you and not only on the model vendor. One law-firm analysis adds that a deployer who uses an AI agent in a way its provider did not intend may become the provider of a high-risk system (Formalize). Which role you hold is a question for counsel.
Confidence: Medium-High. Confirm article numbers and dates against the Official Journal text on EUR-Lex.
United States. There is no single federal AI statute. Colorado’s AI Act, first due 1 February 2026, was delayed to 30 June 2026. By May 2026 it was effectively frozen: a magistrate judge stayed enforcement on 27 April in a suit brought by xAI, the Department of Justice intervened, and the state attorney general committed not to enforce it while replacement legislation was considered (Squire Patton Boggs).
Confidence: Medium. This is a May 2026 snapshot and I did not verify later developments.
In practice. The Moffatt reasoning and the Mata sanctions both rest on existing rules, and in both, liability followed the party that relied on the output. Rules are moving, but that principle is the one already being enforced. Which brings the question back to money.
Is the cheaper model worth it?
If a human has to check the output, review time can dwarf token cost. Take Task A on Astra against Luna: the difference in API cost is $0.074. At an assumed $40 an hour, one minute of review costs $0.67.
That gives a break-even you can compute without inventing any pass rates. The cheap model has to pass review at least (c_cheap + r) ÷ (c_top + r) as often as the top model, where c is API cost per attempt and r is review cost per attempt.
| Review time per output | r (at $40/h) | Luna must pass at least this share as often as Astra |
|---|---|---|
| None | $0 | 1% |
| 6 seconds | $0.067 | 48% |
| 1 minute | $0.67 | 90% |
| 10 minutes | $6.67 | 99% |
The more a human has to read, the more the choice depends on how often the model is right, and the less on its price. Where nobody reviews the output at all, the cheap model wins easily on cost, and the exposure falls on whoever relies on the result.
Exposure has its own break-even. If a bad output can cost you $5,000 (the Mata fine), paying $0.074 extra per task is justified if it cuts the chance of that error by more than about 1 in 67,000 per task (0.074 ÷ 5,000). Whether any model does that is something only your own measured error rates can say. No vendor price list can.
What I could not verify
- Vendor contract terms. Indemnities, liability caps and data-use terms were not read. Read them before building a regulated product on any vendor.
- Hard spend caps. I did not verify how each vendor’s spend limits behave, and whether they stop usage or only alert. Test that before launch.
- Real-world pass rates by model. No neutral, current, task-level error-rate dataset is cited here. Artificial Analysis’s cost per task is a weighted average over its own ten evaluations, not your workload, and I read only its medium and high/xhigh comparisons directly. Max-effort figures that circulate on other sites were not checked against its pages.
- The Deloitte-style cases I left out. Large consulting and government cases involving AI errors exist, but I could not confirm refund amounts from primary documents, so I did not use them.
FAQ
What is the cheapest major AI API per task? Of the models priced here, GPT-6 Luna and Claude Haiku 5.5, both $0.10 / $0.50 per million tokens. Gemini 3.5 Flash-Lite is the cheapest Google option at $0.30 / $2.50. Cheapest per token is not cheapest per usable result.
Is Gemini cheaper than GPT or Claude? At the top and mid tiers, yes at list price. Gemini 3.8 Flash costs less than GPT-6.1 Sol or Claude Sonnet 5.5 until its introductory rate ends on 31 December 2026. At the budget tier, Luna and Haiku are cheaper.
Does prompt caching change the ranking? It lowers every bill by a similar share, but flagship cached reads differ: $0.25 on Fable 5.1 against $1.00 on Astra.
Who is liable when an AI chatbot gives wrong information? In Moffatt v. Air Canada, the company was held responsible. That is a tribunal ruling, so check the law where you operate.
Correction ledger
| Item | Earlier state | Correction | Basis |
|---|---|---|---|
| Output-to-input ratio | v1 said output costs “4–6×” input everywhere | 5× on eight of ten models, 6× on Gemini 3.1 Pro, 8.3× on Flash-Lite | Price pages |
| Vendors “close within a tier” | v1 said the three vendors are close within each tier | OpenAI and Anthropic are identical at three tiers. Google differs by roughly 2.7× to 5× | Price pages |
| Pass-rate example | v1 used assumed 80% / 95% pass rates | Removed; replaced with a break-even computed from review cost | n/a |
| OpenAI batch discount | v1 flagged as secondary-only | Verified: Sol Batch is $1 / $5 | OpenAI pricing page |
| Caching | v1’s agent run assumed no caching | Cached scenario added with first-party cache prices | Price pages |
| Deloitte case | v1 included it | Removed: refund amount unverified | Wire reports differ on contract size, refund undisclosed |
| Stolen-key reports | Boing Boing described a solo developer in Vietnam | Used The Register’s account (a small team); other reports say Mexico. The Boing Boing piece also misdates its context | Cross-check |
| Opus vs Sol cost-per-task gap | v2 first cited “about 8×” from a secondary summary of max-effort figures | Replaced with Artificial Analysis’s own pages: about 6.4× at medium effort ($1.34 vs $0.21) and about 11× for Sol high vs Opus xhigh. Secondary sites disagree on max-effort Sol figures ($0.72 vs $1.04), so those are no longer used | Artificial Analysis comparison pages |
| Regulatory placement | v2 put the rules after the break-even maths | Moved directly after the failures, with a dated EU timeline table | n/a |
| EU article numbers | Some secondary sources place GPAI duties in Articles 50–55 | GPAI is Articles 51–56; Article 50 is transparency. Verify on EUR-Lex | Cross-check |
Sources
- OpenAI: Pricing, GPT-6.1 Sol, Models, GPT-6.1 Sol announcement
- Anthropic: Pricing
- Google: Gemini Developer API pricing
- Artificial Analysis: Sol (medium) vs Opus 5.5 (medium), Sol (high) vs Opus 5.5 (xhigh), Opus 5.5 analysis, GPT-5.5 analysis
- Moffatt v. Air Canada, 2024 BCCRT 149: decision, CanLII
- Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023): Bloomberg Law
- Stolen Gemini key: The Register
- EU AI Act / Digital Omnibus: Orrick, Cloud Security Alliance, Formalize
- Colorado AI Act: Squire Patton Boggs