Same $2 Token Price, Ten Times the Bill: Budget the New Mid-Tier Models by Finished Job

The mid-tier API price tag stopped being a decision on 2 October 2026. Igor Rotor’s note that day put three new models on the same line: Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon, each listed at $2 per million input tokens and $10 per million output tokens. Anthropic shipped Sonnet 5.5 on 28 September. OpenAI followed with GPT-6.1 Sol on 29 September. Google announced Argon on 30 September, still limited to trusted cyber defenders in its Fairwind programme, with an introductory rate Google has said will rise to $4 and $20 later.

For a one-person SaaS, that sticker is a trap. Agents do not buy a million tokens. They buy a finished job: a pull request reviewed, a support thread drafted, a changelog summarised. Rotor, citing Artificial Analysis, reported an average task cost of $0.72 on GPT-6.1 Sol, $1.99 on Gemini 4 Argon and $7.67 on Sonnet 5.5. Same menu price. More than ten times the bill.

Why the same rate produces different invoices

Modern model calls are loops. The model plans, calls a tool, reads the result, and tries again. Every token in that loop is billed, including tokens written before a user sees a sentence. A longer scratchpad can cost several times more for the same customer-visible outcome.

Sonnet 5.5 is the clean example. Decrypt, cited in the 2 October note, reported that at maximum effort the model wrote about 193,000 tokens per Artificial Analysis test task, roughly 60 percent more than Opus 5.5. Sonnet 5.5 then cost more per task than Opus 5.5, about $6, even though its tokens are half the price. It landed near Fable 5.1, whose tokens list at five times the Sonnet rate. Anthropic has said Sonnet 5.5 can cost up to 30 percent less per task than its predecessor. The independent maximum-effort run found the opposite, $7.67 against $5.09 for Sonnet 5. Artificial Analysis points founders at the High effort setting for better value.

Quality moved with the extra work, so “pick the cheap one” also fails. Artificial Analysis ranked Sonnet 5.5 second on its Intelligence Index at 56, behind Opus 5.5 at 58. On Terminal-Bench 4.0, Anthropic reported 70.6 percent for Sonnet 5.5 against 66.4 percent for Opus 5.5. The lab’s own run kept that order, 63.6 against 59.6. Argon scored 53, level with GPT-6 Astra and Fable 5.1. GPT-6.1 Sol scored 52, up from 48 for GPT-6 Sol a week earlier.

Pawel Huryn’s bug hunt, summarised in the same note, makes the loop cost concrete. Models search two repositories for 105 hidden bugs. GPT-6 Sol fixed 29.3 at maximum effort. GPT-6.1 Sol fixed 44.3, close to GPT-6 Astra at 45. The run cost $5.88, against $95.35 for GPT-5.6 Sol. Huryn credited cheaper cache reads, a discount he warned may be temporary, and fewer steps: about 200 at the xhigh setting, against 485 before.

Speed is the other missing invoice. Artificial Analysis measured Sol at 64 output tokens per second. At maximum effort it can think for almost five minutes before the first word. Sonnet 5.5 ran at 139 tokens per second. On Huryn’s board, maximum-effort Sonnet 5.5 fixed 51.3 bugs, then needed 1,497 steps and almost five hours, against 270 steps and 90 minutes for Astra. At the xhigh setting both Sonnet 5.5 and Opus 5.5 fixed 36 bugs, at $42.86 against $34.98. Long sessions are mostly cached input, priced at $0.20 per million on both Anthropic models.

Argon is not a switch you can flip. Google is releasing it first to cyber defenders, including a version without cyber guardrails, and has not dated paid API access. The top of the catalogue also skipped the sale: Astra and Fable 5.1 still list at $10 and $50. Open-weight models are the floor. Exponential View estimates they already serve nearly half of AI tokens. Xiaomi’s MiMo-V2.6-Pro scored 46, near GPT-6 Sol, for less than $0.15 per task.

A worked bill for a support agent

Draft a reply to a failed-renewal ticket using the last invoice, the dunning state, and the public help article. Cap the agent at four tool calls. If the prompt and records are 8,000 input tokens and the model writes 780 output tokens, the call is about $0.022 at $2 and $10. Forty a day is under $27 a month.

Let it wander and the shape changes. Four retries, a 20,000-token scratchpad, and a full thread replay can push one ticket past 150,000 tokens. Forty of those a day can exceed a $49 plan. Add the human term PitchBook analyst Harrison Rolfes used in a July note: about $17 for a failed attempt, roughly ten minutes of time. On a solo product that time is yours. A model $0.50 cheaper per call that fails twice as often is the expensive model.

Subscriptions do not erase the gap. Huryn found usage worth about 10 times a $100 ChatGPT plan and about 45 times Claude Max 20x at $200, measured at API rates. Rotor estimated a maximum-effort Opus 5.5 coding run at about $1.30 of Claude Max, against under $0.60 for Sol on ChatGPT. Appetite offset the larger allowance. If your product calls the API for customers, you pay the loop, not the consumer multiplier.

How to budget the week’s models

1. Split jobs before you split vendors. Use three buckets: extract and classify, draft with a human send, and act in production. Extraction can sit on a cheap model. Drafts use a mid-tier model with a short output cap. Card charges and customer email stay on the smallest capable model plus a hard approval.

2. Price one finished job, not one million tokens. Pick five tasks from last week: a bug repro, a help reply, a SQL draft, a changelog, a churn note. Run each three times on the current model and one alternative, at the effort you will ship. Log tokens, cache reads, steps, wall time, and whether you accepted the result. Divide dollars by accepted results.

3. Cap steps before you cap the monthly bill. For support drafts, four tool calls and 1,500 output tokens is a sane start. If a coding agent has not produced a readable diff in 15 minutes, count the run as failed. A five-hour research result is not a workflow.

4. Treat cache discounts as temporary. Huryn’s Sol saving leaned on a cache rate he flagged as possibly temporary. Anthropic’s $0.20 cache price is identical on Sonnet 5.5 and Opus 5.5, so long sessions do not reward the cheaper sticker. Keep a 30 percent contingency where cached input dominates.

5. Leave Argon out of this month’s margin. It is not generally available, and the $2 and $10 rate is introductory. Retest when paid access exists and you can rerun the same five tasks.

6. Put repair time on the spreadsheet. Log $17, or your real hourly rate, for each ten minutes spent fixing a draft. After 30 tasks, drop any model whose accepted-result cost is not at least 20 percent better than the incumbent.

7. Publish the cap if you resell agent work. A $0.02 support reply and a $7 coding run are different features. Include a draft allowance, then meter the rest. That packaging survives the next rate cut.

Do not rewrite the stack because three labs printed the same numbers. Sol’s lower task cost is a reason to test bounded jobs, not a reason to rip out a workflow that already accepts its output. The middle of the market now has a shared sticker and a split bill. Founders who measure finished jobs will feel the split. Founders who compare $2 with $2 will meet it on the invoice.

FAQ

Does the shared $2 and $10 price mean these models cost the same?
No. Those are list prices per million tokens. Task costs cited on 2 October were about $0.72, $1.99 and $7.67. Argon is introductory and not on general sale.

Should a solo SaaS switch coding agents to GPT-6.1 Sol this week?
Only after a capped trial on your repository. Huryn’s hunt showed a large gain over GPT-6 Sol and a much smaller bill than GPT-5.6 Sol, but Sol was slow, and one benchmark is not your codebase.

Is maximum effort the right default?
No. Artificial Analysis recommends High effort on Sonnet 5.5 for value. Keep maximum effort for a weekly eval, not for every ticket.

What if the bill is a ChatGPT or Claude subscription?
Consumer plans subsidise heavy use relative to API rates. Token appetite can erase that edge. If your product calls the API for users, ignore the consumer multiplier.

When should Argon enter the comparison?
When paid API access exists and the same five tasks can be rerun. Until then it is a benchmark story, not a vendor.