Cost Per Finished AI Task: GPT-6.1 Terra vs Mistral Large 4 vs Gemini 4 Carbon vs Claude Opus 5.5
Byline: Rana — Editor, Elite Era Trends
Checked against primary sources on November 10, 2026.
Table of contents
- The short answer
- The price list, in one table
- The three jobs we priced
- Job 1: a 1,500-word research brief
- Job 2: fixing a bug in a 400-line file
- Job 3: ten customer emails
- The agent session: where caching changes everything
- What 100 jobs a month costs on each model
- Why the cheapest model is not always the cheapest task
- The Carbon catch: you probably cannot buy it yet
- How to pick
- Frequently asked questions
- Sources
The short answer
Per-token prices are the wrong number to shop by. What you actually pay is the price of a finished job — and on the same three jobs, Mistral Large 4 costs about a tenth of Claude Opus 5.5, GPT-6.1 Terra costs about a quarter, and Gemini 4 Carbon costs the same as Opus 5.5 once its introductory discount ends. But price is only half the story: Carbon and Opus 5.5 are the strongest models in the comparison, and a cheap model that needs three attempts — or fails outright — is not cheap at all. Below is the rate card, the worked maths for three everyday jobs, and a rule for choosing.
The price list, in one table
All prices are standard API rates in US dollars per million tokens, as published by each vendor. Input is what you send the model; output is what it writes back — and output is where most of the money goes, because it costs 3–5 times as much as input on every model here.
| Model | Input / 1M tokens | Cached input / 1M | Output / 1M tokens | Released |
|---|---|---|---|---|
| GPT-6.1 Terra (OpenAI) | $1.00 | $0.05 | $5.00 | Oct 26, 2026 |
| Mistral Large 4 (Mistral AI) | $0.70 | $0.07 | $2.10 | Oct 29, 2026 (open weights) |
| Gemini 4 Carbon, introductory (Google) | $2.00 | $0.10 | $10.00 | Oct 28, 2026 |
| Gemini 4 Carbon, standard (Google) | $4.00 | $0.20 | $20.00 | After the intro period ends |
| Claude Opus 5.5 (Anthropic) | $4.00 | $0.20 | $20.00 | Sep 22, 2026 |
Three details from the fine print matter before you budget anything. First, Terra’s price doubles on long prompts: any request with more than 272,000 input tokens bills at twice the input rate and 1.5 times the output rate for the whole request. Second, batch processing halves the bill on Terra ($0.50/$2.50) and on Opus 5.5 ($2/$10) — useful if your jobs can wait minutes instead of seconds. Third, Carbon’s introductory price has no announced end date, so the $2/$10 rate is the price today and the $4/$20 rate is the price to plan around.
For context, Artificial Analysis — the independent benchmark tracker — also publishes a cost-per-task figure from its own heavy benchmark workload: $0.40 a task for Terra, $2.06 for Large 4, $7.76 for Carbon at standard prices, and $10.04 for Opus 5.5. Its workload is far bigger than the everyday jobs below, so the absolute numbers differ — but the ranking is the same, and that is the point: the cheapest model per token is the cheapest per task, and the most expensive is the most expensive, by a wide margin.
The three jobs we priced
To compare fairly, every model does the same job with the same token counts. That isolates the price difference. Real jobs vary — a model that writes more verbosely uses more output tokens, and a model that needs a retry doubles the bill — and we deal with that in section 9. Token counts below are our stated assumptions, not measurements: roughly 750 words per 1,000 tokens.
- Job 1 — a 1,500-word research brief. You paste a brief and some source material (3,000 input tokens) and the model writes the brief (2,500 output tokens).
- Job 2 — fixing a bug in a 400-line file. You paste the file plus an error log (12,000 input tokens) and the model writes the corrected code with an explanation (3,000 output tokens).
- Job 3 — ten customer emails. You paste tone guidelines and the customer details (2,000 input tokens) and the model drafts ten short replies (4,000 output tokens).
The formula for every price below is the same: (input tokens × input rate) + (output tokens × output rate), divided by one million.
Job 1: a 1,500-word research brief
3,000 input tokens, 2,500 output tokens.
| Model | Input cost | Output cost | Total per brief |
|---|---|---|---|
| Mistral Large 4 | $0.0021 | $0.0053 | $0.01 |
| GPT-6.1 Terra | $0.0030 | $0.0125 | $0.02 |
| Gemini 4 Carbon (intro) | $0.0060 | $0.0250 | $0.03 |
| Gemini 4 Carbon (standard) | $0.0120 | $0.0500 | $0.06 |
| Claude Opus 5.5 | $0.0120 | $0.0500 | $0.06 |
A research brief costs between one and six cents depending on the model. The spread is 6-to-1 between the cheapest and the most expensive, but the absolute money is trivial: even at Opus 5.5 prices, a hundred briefs a month costs about $6. For writing-heavy work at this size, choose on quality, not price — the model that writes the best first draft saves you more in editing time than the price difference costs.
Job 2: fixing a bug in a 400-line file
12,000 input tokens, 3,000 output tokens. This is the input-heavy job: most of the tokens are the code you paste in, not the answer.
| Model | Input cost | Output cost | Total per fix |
|---|---|---|---|
| Mistral Large 4 | $0.0084 | $0.0063 | $0.01 |
| GPT-6.1 Terra | $0.0120 | $0.0150 | $0.03 |
| Gemini 4 Carbon (intro) | $0.0240 | $0.0300 | $0.05 |
| Gemini 4 Carbon (standard) | $0.0480 | $0.0600 | $0.11 |
| Claude Opus 5.5 | $0.0480 | $0.0600 | $0.11 |
Notice what changes here: input is now the bigger cost on the expensive models, because the file you paste is four times the size of the answer. Large 4’s cheap input rate ($0.70) is doing the heavy lifting — a bug fix costs about a cent and a half. Opus 5.5 costs about eleven cents, roughly seven times as much. That still sounds small, but it compounds: a developer running 20 fixes a day pays about $0.30 a day on Large 4 and about $2.20 a day on Opus 5.5 — roughly $6 versus $44 a month.
Job 3: ten customer emails
2,000 input tokens, 4,000 output tokens. This is the output-heavy job: the answer is twice the size of the prompt.
| Model | Input cost | Output cost | Total per batch of 10 |
|---|---|---|---|
| Mistral Large 4 | $0.0014 | $0.0084 | $0.01 |
| GPT-6.1 Terra | $0.0020 | $0.0200 | $0.02 |
| Gemini 4 Carbon (intro) | $0.0040 | $0.0400 | $0.04 |
| Gemini 4 Carbon (standard) | $0.0080 | $0.0800 | $0.09 |
| Claude Opus 5.5 | $0.0080 | $0.0800 | $0.09 |
Output-heavy work punishes the expensive models hardest, because output is where the rate gap is widest: $2.10 versus $20 per million tokens is nearly a 10-to-1 difference. Ten emails cost about a cent on Large 4 and about nine cents on Opus 5.5. Per email, that is a tenth of a cent against just under a cent — either way, the model is not the expensive part of answering a customer. Your time is.
The agent session: where caching changes everything
So far every job assumed fresh input. Real coding agents do not work that way: they re-read the same codebase turn after turn, and vendors charge far less for input they have already seen. Cached input costs 95% less than fresh input on Terra ($0.05), 90% less on Large 4 ($0.07) and on Opus 5.5 ($0.20), and 95% less on Carbon ($0.10 introductory, $0.20 standard).
Take a realistic agent session: 200,000 input tokens, 90% of them cached, and 30,000 output tokens.
| Model | Fresh input (20,000) | Cached input (180,000) | Output (30,000) | Total per session |
|---|---|---|---|---|
| Mistral Large 4 | $0.0140 | $0.0126 | $0.0630 | $0.09 |
| GPT-6.1 Terra | $0.0200 | $0.0090 | $0.1500 | $0.18 |
| Gemini 4 Carbon (intro) | $0.0400 | $0.0180 | $0.3000 | $0.36 |
| Gemini 4 Carbon (standard) | $0.0800 | $0.0360 | $0.6000 | $0.72 |
| Claude Opus 5.5 | $0.0800 | $0.0360 | $0.6000 | $0.72 |
Two things stand out. First, caching collapses the input bill: on Terra the 180,000 cached tokens cost less than a cent, while the same tokens fresh would cost $0.18. Second, output now dominates everything — on every model, output is 70–83% of the session cost. The model that writes concisely is the model that bills less, whatever its rate card says.
What 100 jobs a month costs on each model
Scaling the per-job figures up makes the choice tangible. One hundred Job 2 bug fixes a month:
| Model | 100 bug fixes / month | 1,000 bug fixes / month |
|---|---|---|
| Mistral Large 4 | $1.47 | $14.70 |
| GPT-6.1 Terra | $2.70 | $27.00 |
| Gemini 4 Carbon (intro) | $5.40 | $54.00 |
| Gemini 4 Carbon (standard) | $10.80 | $108.00 |
| Claude Opus 5.5 | $10.80 | $108.00 |
And one hundred agent sessions a month at the section 7 size: $9 on Large 4, $18 on Terra, $36 on Carbon at introductory rates, $72 on Carbon at standard rates or Opus 5.5. Even the most expensive option here is cheaper than one hour of a developer’s time in the US — which is the right frame. The model bill is not the cost of the work; it is the cost of the assistant doing the work.
Why the cheapest model is not always the cheapest task
Everything above assumes each model finishes the job in one pass. That is the assumption to question, because capability is not equal across these four models.
On Artificial Analysis’s Intelligence Index v5.0, Carbon scores 59 — the top score in the index — and Opus 5.5 scores 58.41. Large 4 scores 26.02. That is a large capability gap: the two expensive models are roughly twice as capable on this composite benchmark as the cheap one. Terra sits between them as a fast mid-tier model, and its benchmark cost of $0.40 a task reflects that middle position.
Here is how that changes the maths. Take the bug fix at $0.015 on Large 4 and $0.108 on Opus 5.5. If Large 4 fixes the bug on the first attempt, it wins by a mile. If it takes three attempts — two failures and a success — the cost is 3 × $0.015 = $0.045, still less than half the Opus price. But if it cannot fix the bug at all, and you escalate to Opus 5.5 anyway, you have paid $0.045 plus $0.108 — more than using Opus first — and spent your own time on the failed attempts. For routine work where the cheap model succeeds first time, cheap wins. For the hardest 10% of jobs, the expensive model’s one-shot success is the cheaper outcome.
The practical rule: route by difficulty, not by brand. Use Large 4 or Terra for drafting, summarising, routine code and first passes — the jobs where a miss costs a retry, not a deadline. Use Carbon or Opus 5.5 for the jobs where failure is expensive: security-sensitive code, work you cannot easily check, and anything a customer will see unedited.
The Carbon catch: you probably cannot buy it yet
Carbon‘s price is published but its product is not, for most people. Google announced Carbon on October 28, 2026 and is giving it first to a subset of cyber defenders, then gradually to vetted partners through its Fairwind programme. There is no public API model ID yet, it is not on Google’s pricing page, and Google has given no date for general developer, business or consumer access. The $2/$10 introductory rate is therefore a price for a product you cannot yet buy — and the $4/$20 standard rate is the one to plan around, since the introductory period has no announced end date and could close at any time.
That also means our Carbon figures are arithmetic on announced rates, not a measured bill. Nobody outside Google’s early-access group has run a Carbon workload yet, so treat the per-task costs as the best available estimate, not a quote.
By contrast, the other three are buyable today: Terra through OpenAI’s API and ChatGPT, Large 4 through Mistral’s API — or self-hosted, since Mistral released the weights openly on October 29 — and Opus 5.5 through Anthropic’s API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Large 4’s open weights are a genuine cost option for teams with their own GPUs: self-hosting replaces the API bill with hardware and electricity, though running a 1-trillion-parameter model needs serious infrastructure — Mistral says one eight-GPU Blackwell node holds it at 8-bit precision.
How to pick
- You write a lot of routine content or code and check it yourself: Mistral Large 4. About a cent a job, and the open weights mean you are never locked in.
- You want the cheapest capable all-rounder with a big ecosystem: GPT-6.1 Terra. Two to three cents a job, deep tooling, and the batch tier halves the price if your work can wait.
- You do security work or the hardest reasoning tasks: Gemini 4 Carbon, when it opens up — or Claude Opus 5.5 today. Budget roughly ten cents a job and nine times the per-task cost of Large 4, and expect it to be worth it on the jobs where a wrong answer is expensive.
- You run an agent all day: watch output tokens, not input. Turn on prompt caching everywhere, keep prompts tight, and remember Terra’s 272,000-token surcharge before pasting a whole repository into one request.
Frequently asked questions
Mistral Large 4 is the cheapest of the four models compared: about $0.01 per research brief, bug fix or batch of ten emails, against about $0.02–$0.03 for GPT-6.1 Terra and $0.06–$0.11 for Claude Opus 5.5 and Gemini 4 Carbon at standard rates.
The gap comes from output pricing — Large 4 charges $2.10 per million output tokens, Opus 5.5 charges $20. Carbon at introductory rates sits in the middle at $0.03–$0.05 a job.
The vendors price for different things. Opus 5.5 is Anthropic’s flagship and scores 58.41 on Artificial Analysis’s Intelligence Index, against 26.02 for Large 4 — it is built for the hardest reasoning and agentic work, not for cheap tokens alone.
Large 4 is an open-weight model whose vendor monetises the hosted API cheaply and lets anyone self-host the weights. You are paying Opus prices for capability and reliability, not for tokens.
Almost certainly not. Google is rolling Carbon out first to a subset of cyber defenders and then to vetted partners through its Fairwind programme, with no announced date for general access — and there is no public API model ID yet.
It is not in the Gemini app, and it is not on Google’s pricing page. The $2/$10 introductory price is announced but not yet buyable for most developers, so plan around the $4/$20 standard rate.
Yes — it is the single biggest lever on input-heavy work. Cached input costs 90–95% less than fresh input on all four models: $0.05 per million tokens on Terra, $0.07 on Large 4, and $0.20 on Opus 5.5 and Carbon at standard rates.
In our example agent session, 180,000 cached tokens cost under four cents on every model, while the same tokens fresh would cost up to $0.72. Carbon’s introductory cached rate is $0.10 per million tokens.
No. Value is cost per accepted task, not cost per token. If the cheapest model needs three attempts to finish a job, it can still be cheaper — three Large 4 bug fixes cost $0.045 against one Opus 5.5 fix at $0.108.
But if it fails outright and you escalate to the expensive model anyway, you pay both bills plus your own time. Use cheap models for routine work you can check, and expensive models where a wrong answer is costly.
On our Job 2 assumptions — 12,000 input tokens and 3,000 output tokens per fix — 100 fixes a month cost $1.47 on Mistral Large 4, $2.70 on GPT-6.1 Terra, $5.40 on Gemini 4 Carbon at introductory rates, and $10.80 on Carbon at standard rates or Claude Opus 5.5.
At 1,000 fixes the figures are $14.70, $27, $54 and $108. Even the most expensive option is cheaper than one hour of a US developer’s time.
Sources
- OpenAI API pricing — GPT-6.1 Terra $1.00 input / $0.05 cached / $5.00 output per million tokens; batch and flex half rates; 272,000-token long-context rule. https://platform.openai.com/docs/pricing
- Mistral AI pricing — Mistral Large 4 $0.70 input / $0.07 cached / $2.10 output per million tokens. https://mistral.ai/pricing
- Anthropic pricing — Claude Opus 5.5 $4.00 input / $0.20 cached / $20.00 output per million tokens; batch API half rates; fast mode double rates. https://www.anthropic.com/pricing
- Google Gemini API pricing — Gemini model pricing page (Carbon not yet listed; introductory $2/$10 and standard $4/$20 rates announced October 28, 2026). https://ai.google.dev/gemini-api/docs/pricing
- Artificial Analysis Intelligence Index v5.0 — Carbon 59, Opus 5.5 58.41, Large 4 26.02; cost per index task Terra $0.40, Large 4 $2.06, Carbon $7.76 (standard), Opus 5.5 $10.04. https://artificialanalysis.ai/
Prices checked against the vendors’ own pricing pages on November 10, 2026. Token counts for the three jobs and the agent session are our stated assumptions for comparison, not measured bills; your costs depend on your prompts. Carbon figures are arithmetic on announced rates — the model is not yet generally available.
Rana is the editor of Elite Era Trends, where he explains Medicare, Social Security and AI news for US readers in plain language — costs, deadlines and plan choices, checked against official sources.
