AI Model Cost Comparison: What the Same Build Cost on Claude, GPT-6, Gemini, Kimi and DeepSeek

This page compares what AI models cost per finished build: the full bill for the same real job on each model. Pat Simmons gave the same builds to three to five models at a time and logged each run's tokens, minutes and API-equivalent cost, meaning the real token counts priced at each company's published API rates.

Across 10 builds, DeepSeek V4.1 Flash cost $1.00 to $1.43 a build and Gemini 4 Argon $6.85 to $11.95, and both placed low in the blind rankings.

Opus 5.5 cost $11.66 to $160.67 a build, came first in all four builds of the Gemini 4 Argon test and won two of three in the Kimi K3.1 test. The same job can vary about 400-fold between models: the rebuild of Kimi's launch video cost $1.43 on DeepSeek V4.1 Flash and $592.37 on Fable 5.1.

Claude Pro and ChatGPT Plus are both $20 a month. The builds themselves ran on subscriptions, so the dollar figures show what the same tokens would cost on the API.

Last checked November 10, 2026: every API price and plan price on this page was re-checked that day on each company's own page, including Anthropic's API pricing, OpenAI's API pricing and Google's Gemini API pricing. The build costs come from Pat's published tests and don't change.

API-equivalent cost per build from Pat Simmons's three build-off rounds, run between September 23 and October 25, 2026. Ranks come from the two rounds that were ranked blind.

What did the same build cost on each model?

The cheapest single run was DeepSeek V4.1 Flash's 3D billiards game at $1.00, and the most expensive was Fable 5.1's rebuild of Kimi's launch video at $592.37. The chart puts all 40 runs side by side, grouped by build, and the tables under it add the tokens, minutes and blind rank for each one.

Each round had different builds and a different line-up of models, so compare models inside a round. Across rounds, the per-build ranges in the summary table above are the fair comparison.

Round 1: Opus 5.5, Fable 5.1 and GPT-6 Astra on three builds

The first round, run on September 23, 2026, gave three models a set of scenes from Moby Dick drawn entirely in code, a 20-second launch film with every sound made in code, and a 3D skateboarding game.

Totals for the three builds: Opus 5.5 $106.93 over 266 minutes, Fable 5.1 $90.99 over 143 minutes and GPT-6 Astra $60.20 over 102 minutes, from the Opus 5.5, Fable 5.1 and GPT-6 Astra build costs. That post published costs without a blind ranking, so these rows have no rank.

Round 2: Gemini 4 Argon against Opus 5.5, Fable 5.1 and GPT-6 Astra

The second round, from Pat's Gemini 4 Argon review, used four builds: a photoreal 3D cold brew bottle you can spin and pour, a product launch video, a Rocket League-style car soccer game and a pixel-for-pixel copy of the oryzo.ai website. Totals: Gemini 4 Argon $39.54, Opus 5.5 $189.04, Fable 5.1 $203.08 and GPT-6 Astra $409.96. Pat ranked every build blind, and Opus 5.5 came first and Argon fourth in all four.

Round 3: Kimi K3.1 Pro and DeepSeek V4.1 Flash against Claude and GPT-6 Astra

The third round, from the Kimi K3.1 Pro review, added two open-weight models (models anyone can download and run): a first-person 3D billiards game, a pixel-for-pixel copy of the Morpho water website, and a shot-for-shot rebuild of Kimi's own 94-second launch video.

Totals: Kimi K3.1 Pro $35.31, Opus 5.5 $276.48, Fable 5.1 $675.88, GPT-6 Astra $232.14 and DeepSeek V4.1 Flash $3.45. On the launch-video rebuild the review gives first, fourth and fifth place only, so Fable 5.1 and GPT-6 Astra show as "2nd or 3rd".

Which AI model is cheapest per build?

DeepSeek V4.1 Flash was the cheapest model in every build it ran, at $1.00 to $1.43, and Gemini 4 Argon was the cheapest in every build of its round, at $6.85 to $11.95. Neither won a build. DeepSeek placed third, fourth and fourth out of five, and Argon placed fourth out of four every time.

DeepSeek still beat Kimi K3.1 Pro in all three of its builds while costing a tenth as much: $3.45 for the three against Kimi's $35.31. DeepSeek took longer than Kimi on two of the three builds, though: 82 minutes against 63 on the billiards game, and 180 against 177 on the Morpho website copy.

Between ChatGPT's model and Claude's two, GPT-6 Astra was the cheapest in 7 of the 10 builds all three ran. Cost and placing pulled in opposite directions in these tests: the two cheapest models never placed above third, and Opus 5.5, which won six of the seven ranked builds, cost more than Astra on most of them.

Does Claude or ChatGPT cost more?

Per build, ChatGPT's top model usually cost less: GPT-6 Astra was cheaper than Opus 5.5 in 8 of the 10 builds and cheaper than Fable 5.1 in 7.

Both of the builds where Astra cost more than Opus were website copies. On the oryzo.ai copy Astra ran 499 minutes and $357.84, against $67.87 for Opus 5.5, and on the Morpho copy it cost $118.86 against $84.07. Fable 5.1 also came in under Astra on the 3D skateboarding game, $32.26 against $44.08.

Claude placed higher. Opus 5.5 finished above GPT-6 Astra in all seven builds that were ranked blind, which lines up with Pat's Claude vs ChatGPT vs Gemini on the same builds, where Claude won every round of the Argon test.

Round by round, Opus 5.5 cost more than GPT-6 Astra in the first round ($106.93 against $60.20) and the third ($276.48 against $232.14), and less in the second ($189.04 against $409.96), because of that one long website copy.

On subscriptions the entry prices are the same. Claude Pro is $20 a month, or $17 a month billed yearly ($200 up front), and Max and Ultra are $100 and $200, on Claude's plans page.

ChatGPT Go is $8, Plus is $20 and Pro starts at $100, on ChatGPT's pricing page. The full plan-by-plan breakdown, Grok included, is in our guide to every ChatGPT, Claude, Gemini and Grok plan.

Why per-token prices don't predict the bill

The bill is price per token times the number of tokens, and the number of tokens varied far more between models than the prices did. A token is a small chunk of text, about three or four characters of English, and AI companies charge per million of them; there's more on what a token is and what it costs.

Opus 5.5 lists at $4 per million input tokens and $20 per million output, less than half GPT-6 Astra's $10 and $50. Astra still cost less than Opus on 8 of 10 builds because it usually used far fewer tokens. On the Moby Dick scenes, Opus read 69.2 million input tokens and Astra read 4.2 million, and Astra finished for $8.16 against Opus's $37.22.

Price can win the other way too. On the same build, Fable 5.1 read 32.2 million input tokens, less than half of Opus 5.5's 69.2 million, and still cost more, $42.18 against $37.22, because Fable's per-token prices are two and a half times Opus's.

Most of those input tokens are cache reads: text the model already saw earlier in the session, billed at a small fraction of the full input price.

In the cost tables the models wrote for the Moby Dick build, Opus 5.5 logged 69,788,489 cache-read tokens that came to $13.96, while 1,840,724 tokens saved to the cache (cache writes) came to $14.73. A count of tokens alone says little about the price until you know which kind they were.

Time is the other multiplier. Both of the most expensive runs in the whole set went on for hours: Fable 5.1 spent 216 minutes and $592.37 rebuilding Kimi's launch video, and GPT-6 Astra spent 499 minutes and $357.84 copying the oryzo.ai site. A model that keeps working on its own keeps adding tokens.

AI model API prices per million tokens

On each company's own API price page on November 10, 2026, DeepSeek V4.1 Flash was the cheapest model in this data set and Fable 5.1 and GPT-6 Astra the most expensive. Developers pay these rates when they plug a model into their own app through its API (the connection other software uses).

Standard rates, checked November 10, 2026. OpenAI calls the cache-read price "cached input", Google calls it "context caching" and DeepSeek calls it a "cache hit". Sources: Anthropic, OpenAI, Google, Moonshot's Kimi API prices and DeepSeek's API prices.

Batch processing, for jobs that don't need an answer right away, costs less. Anthropic lists a 50 percent discount, which puts Fable 5.1 at $5 and $25 and Opus 5.5 at $2 and $10. OpenAI's batch price for GPT-6 Astra is $5 and $25, Google's for Argon is $1 and $5 through November 30, and Moonshot bills Kimi's batch jobs at 60 percent of the standard price.

DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays, and every other hour is off-peak at half the price. Google's page says Gemini 4 Argon isn't on its free tier and that every token is billed, including the model's thinking tokens.

Is a $20 subscription cheaper than paying for the API?

For builds this size, the API bill for a single run often passed a month of a $20 plan. One Opus 5.5 build cost $11.66 to $160.67 at API rates, and one Fable 5.1 build $16.55 to $592.37.

A subscription charges a flat monthly price and sets usage limits. Claude's plans page says every plan has usage limits and that how fast you reach one depends on the model, the length of your chats and the size of your files.

These builds hit those caps. Claude's usage limit ran out during the Moby Dick build, and the build waited about 30 minutes for a reset. When you reach a limit, Claude's page lists three options: wait for the reset, move to a higher plan, or, on paid plans, turn on usage credits to keep working at standard API rates.

Heavier plans stretch further. Pat's three GPT-6 Astra builds in Blender, the 3D modeling app, came to $1,771.61 in tokens over 54 hours on a $200-a-month ChatGPT plan, and cache reads made up 1.19 billion of the 1.22 billion input-side tokens. Pat also ran Gemini 4 Argon on Google's $99.99-a-month AI Ultra plan and hit its daily limit after about one build, the 3D bottle.

Plan prices from each company's own page, checked November 10, 2026: Claude and ChatGPT as linked above, Google's AI plans and Kimi's membership plans.

On Claude, the plan also decides which models you get. Pro pays for Fable with usage credits, and Max and Ultra include Fable for up to half of the weekly limit. Google's AI Ultra gives 5 times AI Pro's usage at $99.99 and 20 times at $199.99.

How the builds were run and priced

Every model in a round got the exact same prompt, with no steering, and ran at high effort until the build was finished. Opus 5.5 and Fable 5.1 ran in Claude Code and GPT-6 Astra in Codex. In the third round, Kimi K3.1 Pro ran in Kimi Code, Moonshot's coding tool, and DeepSeek V4.1 Flash ran in Claude Code through OpenRouter.

In the second and third rounds, Pat judged each set of builds blind and only then revealed which model made which. The costs are API-equivalent: each build's real token counts priced at the model's published API rates, with Kimi priced at Moonshot's rates. In the first round each model worked out its own cost from its session logs. No invoice exists for these amounts, because the builds ran on subscriptions.

The rounds ran between September 23 and October 25, 2026, so they reflect the prices and model versions of that period. Gemini 4 Argon's API price doubles on December 1.

Update November 10, 2026

This page launched with 40 runs from three rounds. Anthropic's API price page now also lists Claude Fable 5.2 at $8 per million input tokens and $40 per million output, below Fable 5.1's $10 and $50. Fable 5.2 has not yet been tested on these builds, so it has no rows here.

Sources

• Anthropic: Claude API pricing, checked November 10, 2026
• Anthropic: Claude plans, checked November 10, 2026
• OpenAI: API pricing, checked November 10, 2026
• OpenAI: ChatGPT plans, checked November 10, 2026
• Google: Gemini API pricing, checked November 10, 2026
• Google: Google AI plans, checked November 10, 2026
• Moonshot: Kimi API pricing, checked November 10, 2026
• Moonshot: Kimi K3.1 Pro membership plans, checked November 10, 2026
• DeepSeek: Models and pricing, checked November 10, 2026
• Pat Simmons: Gemini 4 Argon, No-Hype Full Review and Testing
• Pat Simmons: Kimi K3.1, No Hype Full Review and Testing