What a coding-agent model costs, next to how it ranks

A daily re-read of what the coding-agent models list for. The table, the field notes behind it, and the script that fetches it are on GitHub.

Every model that ranks in an agents category of the Design Arena, priced twice: what it lists for, and what a coding agent actually pays. Read from OpenRouter on 2026-08-23.

List price is not your bill, and here is by how much. A coding agent re-reads its whole context on every step, so about 95.6% of the tokens it sends are cache reads — billed at a fraction of the list input price that every other pricing table sorts by. Repriced at that mix, list input price overstates what an agent actually pays by a median 6.5× (range 3.4×–7.9× across the 58 rows where it can be computed).

The ranking barely moves; the bill does. 9 of the 10 cheapest by list price are still in the cheapest 10 repriced. So list price tells you which model is cheap and lies about what you will pay — and the 3.4×–7.9× spread in that multiple is what separates two models whose list prices look identical.

That 95.6% is one person's measurement of one coding agent (8.04B tokens, 2026-05-16), not an industry figure — it is simply the only public measurement we could find. 58 of these rows publish a cached-input price, so the weights ship in the JSON: recompute with your own mix. Cache-write prices are not in the catalog, so that 2.7% of tokens is folded into the cache-miss share, which understates cost by roughly 0.7%.

The batch rows are kept as their own entries rather than folded into the model above them. The prices belong to the catalog and the ranks belong to the arena; neither is ours, and both move.

All 60 models, cheapest to actually run

$0.0283/M at agent mix ($0.1875 in · $0.9375 out list) — Gemini 3.7 Flash · batch · #2 agenticgamedev · 1M ctx · cached in $0.0187
$0.0501/M at agent mix ($0.3969 in · $0.7938 out list) — DeepSeek V4 Pro 0423 · #27 godotgamedev · 1M ctx · cached in $0.0331
$0.0566/M at agent mix ($0.375 in · $1.875 out list) — Gemini 3.6 Flash · batch · #7 agenticgamedev · 1M ctx · cached in $0.0375
$0.0566/M at agent mix ($0.375 in · $1.875 out list) — Gemini 3.7 Flash · #2 agenticgamedev · 1M ctx · cached in $0.0375
$0.0731/M at agent mix ($0.30 in · $1.20 out list) — MiniMax M3 · #10 python-pptxslides · 1M ctx · cached in $0.06 · field note: 1.6 Billion Free Tokens Is a Compression Ratio, Not a Strategy
$0.0731/M at agent mix ($0.30 in · $1.20 out list) — MiniMax M3 · batch · #10 python-pptxslides · 524K ctx · cached in $0.06 · field note: 1.6 Billion Free Tokens Is a Compression Ratio, Not a Strategy
$0.0769/M at agent mix ($0.50 in · $3.00 out list) — Gemini 3 Flash Preview · #9 agenticslides · 1M ctx · cached in $0.05
$0.0918/M at agent mix ($0.45 in · $2.25 out list) — Kimi K2.5 · #9 godotgamedev · 262K ctx · cached in $0.07
$0.0979/M at agent mix ($0.40 in · $1.75 out list) — GLM 4.7 · #27 androidnative · 204K ctx · cached in $0.08
$0.1131/M at agent mix ($0.75 in · $3.75 out list) — Gemini 3.6 Flash · #7 agenticgamedev · 1M ctx · cached in $0.075
$0.1153/M at agent mix ($0.75 in · $4.50 out list) — Gemini 3.5 Flash · batch · #3 agenticslides(python-pptx) · 1M ctx · cached in $0.075
$0.1218/M at agent mix ($0.50 in · $2.00 out list) — GLM 4.6 · #13 godotgamedev · 204K ctx · cached in $0.10
$0.1396/M at agent mix ($0.875 in · $7.00 out list) — GPT-5.2 · batch · #18 godotgamedev · 400K ctx · cached in $0.0875
$0.1448/M at agent mix ($0.60 in · $1.92 out list) — GLM 5 · #16 godotgamedev · 204K ctx · cached in $0.12
$0.1508/M at agent mix ($1.00 in · $5.00 out list) — Claude Sonnet 5 · batch · #4 godotgamedev · 1M ctx · cached in $0.10
$0.1922/M at agent mix ($1.25 in · $7.50 out list) — GPT-5.4 · batch · #22 godotgamedev · 1.1M ctx · cached in $0.125
$0.1997/M at agent mix ($0.67 in · $3.40 out list) — Kimi K2.7 Code · #7 htmlslides · 262K ctx · cached in $0.17
$0.2033/M at agent mix ($0.95 in · $4.00 out list) — Kimi K2.6 · #2 agentichtmlslides · 262K ctx · cached in $0.16
$0.2042/M at agent mix ($1.25 in · $10.00 out list) — GPT-5.1-Codex · #24 mobileapps · 400K ctx · cached in $0.13
$0.2067/M at agent mix ($1.25 in · $4.25 out list) — Muse Spark 1.1 · #8 agenticgamedev · 1M ctx · cached in $0.15 · field note: 1.6 Billion Free Tokens Is a Compression Ratio, Not a Strategy
$0.2067/M at agent mix ($1.25 in · $4.25 out list) — Muse Spark 1.2 · #9 webapps · 1M ctx · cached in $0.15
$0.2197/M at agent mix ($0.966 in · $3.036 out list) — GLM 5.1 · #3 agenticslides · 204K ctx · cached in $0.1794
$0.2263/M at agent mix ($1.50 in · $7.50 out list) — Claude Sonnet 4.5 · batch · #26 mobileapps · 1M ctx · cached in $0.15
$0.2263/M at agent mix ($1.50 in · $7.50 out list) — Claude Sonnet 4.6 · batch · #4 mobileapps · 1M ctx · cached in $0.15
$0.2306/M at agent mix ($1.50 in · $9.00 out list) — Gemini 3.5 Flash · #3 agenticslides(python-pptx) · 1M ctx · cached in $0.15
$0.232/M at agent mix ($0.95 in · $4.00 out list) — Kimi K2.7 Code · batch · #7 htmlslides · 262K ctx · cached in $0.19
$0.2329/M at agent mix ($0.966 in · $3.036 out list) — GLM 5.2 · #10 agenticgamedev · 1M ctx · cached in $0.1932
$0.2494/M at agent mix ($1.25 in · $2.50 out list) — Grok 4.20 · #12 htmlslides · 2M ctx · cached in $0.20
$0.2494/M at agent mix ($1.25 in · $2.50 out list) — Grok 4.3 · #9 pptxslides · 1M ctx · cached in $0.20
$0.2792/M at agent mix ($1.75 in · $14.00 out list) — GPT-5.2 · #18 godotgamedev · 400K ctx · cached in $0.175
$0.2792/M at agent mix ($1.75 in · $14.00 out list) — GPT-5.2-Codex · #19 godotgamedev · 400K ctx · cached in $0.175
$0.2792/M at agent mix ($1.75 in · $14.00 out list) — GPT-5.3-Codex · #24 godotgamedev · 400K ctx · cached in $0.175
$0.29/M at agent mix ($1.20 in · $4.00 out list) — GLM 5V Turbo · #4 androidnative · 202K ctx · cached in $0.24
$0.3017/M at agent mix ($2.00 in · $10.00 out list) — Claude Sonnet 5 · #4 godotgamedev · 1M ctx · cached in $0.20
$0.3075/M at agent mix ($2.00 in · $12.00 out list) — Gemini 3.1 Pro Preview · #5 agentichtmlslides · 1M ctx · cached in $0.20
$0.3184/M at agent mix ($1.40 in · $4.40 out list) — GLM 5.2 · batch · #10 agenticgamedev · 1M ctx · cached in $0.26
$0.3379/M at agent mix ($2.00 in · $6.00 out list) — Qwen3.8 Max · #1 webapps · 1M ctx · cached in $0.25
$0.355/M at agent mix ($1.475 in · $4.425 out list) — Qwen3.7 Max · #8 godotgamedev · 1M ctx · cached in $0.295
$0.3771/M at agent mix ($2.50 in · $12.50 out list) — Claude Opus 4.5 · batch · #11 mobileapps · 200K ctx · cached in $0.25
$0.3771/M at agent mix ($2.50 in · $12.50 out list) — Claude Opus 4.6 · batch · #6 mobileapps · 1M ctx · cached in $0.25
$0.3771/M at agent mix ($2.50 in · $12.50 out list) — Claude Opus 4.7 · batch · #1 agenticslides · 1M ctx · cached in $0.25
$0.3771/M at agent mix ($2.50 in · $12.50 out list) — Claude Opus 4.8 · batch · #2 agenticslides · 1M ctx · cached in $0.25
$0.3771/M at agent mix ($2.50 in · $12.50 out list) — Claude Opus 5 · batch · #3 fullstack · 1M ctx · cached in $0.25
$0.3843/M at agent mix ($2.50 in · $15.00 out list) — GPT-5.4 · #22 godotgamedev · 1.1M ctx · cached in $0.25
$0.3843/M at agent mix ($2.50 in · $15.00 out list) — GPT-5.5 · batch · #7 agenticslides · 1.1M ctx · cached in $0.25
$0.3857/M at agent mix ($2.00 in · $6.00 out list) — Grok 4.5 · #3 godotgamedev · 500K ctx · cached in $0.30
$0.4525/M at agent mix ($3.00 in · $15.00 out list) — Claude Sonnet 4.5 · #26 mobileapps · 1M ctx · cached in $0.30
$0.4525/M at agent mix ($3.00 in · $15.00 out list) — Claude Sonnet 4.6 · #4 mobileapps · 1M ctx · cached in $0.30
$0.4525/M at agent mix ($3.00 in · $15.00 out list) — Kimi K3 · #1 mobileapps · 1M ctx · cached in $0.30 · field note: Beyond Token Pricing: How Indie Devs Should Really Evaluate AI Model Costs
$0.577/M at agent mix ($2.00 in · $6.00 out list) — Grok 4.6 · #1 androidnative · 500K ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Fable 5 · batch · #1 agenticgamedev · 1M ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Opus 4.5 · #11 mobileapps · 200K ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Opus 4.6 · #6 mobileapps · 1M ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Opus 4.7 · #1 agenticslides · 1M ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Opus 4.8 · #2 agenticslides · 1M ctx · cached in $0.50
$0.7542/M at agent mix ($5.00 in · $25.00 out list) — Claude Opus 5 · #3 fullstack · 1M ctx · cached in $0.50
$0.7687/M at agent mix ($5.00 in · $30.00 out list) — GPT-5.5 · #7 agenticslides · 1.1M ctx · cached in $0.50
$1.5084/M at agent mix ($10.00 in · $50.00 out list) — Claude Fable 5 · #1 agenticgamedev · 1M ctx · cached in $1.00
$0.25 in · $1.50 out (no cached price published) — Gemini 3 Flash Preview · batch · #9 agenticslides · 1M ctx
$1.00 in · $6.00 out (no cached price published) — Gemini 3.1 Pro Preview · batch · #5 agentichtmlslides · 1M ctx

The same rows, machine-readable: JSON · CSV · the same table on the web. CC BY 4.0.

Three things this table cannot tell you

It is a rate card. Your bill depends on your usage, and three of the biggest differences never appear in a price column. Put your own numbers in — one page, nothing to install, no account, and it reads the same JSON linked above, so it is never a day behind this table.

DeepSeek bills by the clock — off-peak is half price, and the weekend rule changed on 2026-08-23. The calculator resolves which phase you are in right now, what the same call costs in the other phase, and how long until it flips.
Long context is a cliff, not a slope. Cross a model's long-context threshold and every token in that request reprices, including the ones below it — one token over can double a bill. The table above cannot show you that; it depends on your prompt.
The mix above is the median agent, not yours. Slide the cache-read share to what your own logs say and the ordering of the cheapest ten changes.

The 49 field notes behind these numbers

Written while actually running these models — what broke, what the numbers turned out to mean, and what the price tag did not tell us.

Part of llm-api-pricing — field notes on AI coding agents. This one is also on the web, where it links out to the related write-ups. Every figure across the whole collection is also published as JSON and CSV, each row with the sentence it came from.

Did this save you an afternoon? A star on the repository is the whole ask — it is what puts these in front of the next person looking. The data is CC BY and does not require starring. Does your own number disagree? This same piece has a reply box — it ends on one question written from the figures above, and one line answers it. No template, no title, nothing to organise first. Want a figure that is not in here yet? Say which metric, which provider, which unit in one line — one required field, and the page you came from is already filled in. Requests get turned into rows. Got a better number? Open an issue — the form already knows which write-up you came from; corrections and counter-data are the point.

What a coding-agent model costs, next to how it ranks

All 60 models, cheapest to actually run

Three things this table cannot tell you

The 49 field notes behind these numbers

Report Page