BackgrindFree tool · no signup
AI model pricing tracker
Every major coding model, one table: dollars per million tokens, cache rates, and what a real agent session costs on each. Sources linked on every row. Last verified July 23, 2026.
| Model | Maker | Tier | $ / M in | $ / M out | Session est. | Notes | Source |
|---|---|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | Frontier | $10 | $50 | $1.9$25$95 | Credits-only since Jul 20; subscription access paused | anthropic.com |
| GPT-5.5 | OpenAI | Frontier | $5 | $30 | $0.99$13$49 | benchlm.ai | |
| Claude Opus 4.8 | Anthropic | Flagship | $5 | $25 | $0.95$12.5$47.5 | Default heavy-lift model in Claude Code | claude.com |
| GPT-5.4 | OpenAI | Flagship | $2.5 | $15 | $0.49$6.5$24.5 | benchlm.ai | |
| Claude Sonnet 4.6 | Anthropic | Workhorse | $3 | $15 | $0.57$7.5$28.5 | claude.com | |
| Kimi K3 | Moonshot | Frontier (open-weight) | $3 ($0.3 cached) | $15 | $0.57$7.5$28.5 | Weights promised by Jul 27; benchmarks self-reported so far | benchlm.ai |
| Gemini 3.1 Pro* | Flagship (preview) | $2 | $12 | $0.4$5.2$19.6 | Public preview; third-party tracker figure, no official rate card yet | benchlm.ai | |
| gpt-5.3-codex | OpenAI | Coding-tuned | $1.75 | $14 | $0.37$4.9$18.2 | The BYO-key model behind Codex CLI | developers.openai.com |
| GLM-5.2 | Z.ai | Workhorse (open-weight) | $1.4 | $4.4 | $0.25$3.24$12.52 | Leads Chinese field on coding leaderboards | tech-insider.org |
| Claude Haiku 4.5 | Anthropic | Fast/cheap | $1 | $5 | $0.19$2.5$9.5 | claude.com | |
| Kimi K2.6 | Moonshot | Workhorse (open-weight) | $0.6 | $2.5 | $0.11$1.45$5.55 | Official-platform rate; strongest reputation for long agent runs | artificialanalysis.ai |
| DeepSeek V4-Pro | DeepSeek | Budget (open-weight) | $0.43 | $0.87 | $0.07$0.96$3.74 | Cache-miss rate; the cost floor of the frontier-adjacent tier | codersera.com |
| DeepSeek V4-Flash | DeepSeek | Budget (open-weight) | $0.14 | $0.28 | $0.02$0.31$1.2 | Lighter variant | codersera.com |
* Third-party tracker figure — no official rate card published yet. Standard list prices; caching, batch, and long-context surcharges vary by vendor. Subscription plans (Claude Pro/Max, ChatGPT, Copilot) meter differently — for those, use the coding agent cost calculator.
Subscriptions vs API: what a plan is "worth" in tokens
Plans don't sell tokens — they sell metered access. To compare them anyway, we price each plan's budget (sticker price, or the vendor's published credit grant where one exists) at API list rates for that vendor's workhorse model, using the agent mix of 20 input tokens per output token. "Feature hours" = sessions of 2M in / 100k out.
| Plan | $ / mo | Face value in tokens* | ≈ feature hours | What's actually included | Verdict vs raw API |
|---|---|---|---|---|---|
| Claude Pro | $20 | ≈ 5.6M @ Sonnet 4.6 | ≈ 2.7 | 5-hour window + weekly cap; meters time-windows, not tokens | Daily light-to-medium agent use beats API; heavy weeks hit the cap |
| Claude Max 5x | $100 | ≈ 28.0M @ Sonnet 4.6 | ≈ 13 | 5x Pro limits; can buy extra usage at API rates | Undercuts API for daily drivers — if the limits hold |
| Claude Max 20x | $200 | ≈ 56.0M @ Sonnet 4.6 | ≈ 27 | 20x Pro limits; extra usage at API rates | The heavy-agent plan; API-equivalent value typically exceeds sticker |
| ChatGPT Go (Codex) | $8 | ≈ 3.4M @ gpt-5.3-codex | ≈ 1.6 | Token credits drawn at API rates | ≈ face value in credits — a taster, not a workhorse plan |
| ChatGPT Plus (Codex) | $20 | ≈ 8.6M @ gpt-5.3-codex | ≈ 4.1 | Token credits at API rates; one long session can drain a month | ≈ face value; fine for occasional agent runs |
| ChatGPT Pro 20x (Codex) | $200 | ≈ 85.7M @ gpt-5.3-codex | ≈ 41 | Rate multiplier, not unlimited | For sustained Codex work; still metered |
| Cursor Pro | $20 | ≈ 5.6M @ Sonnet 4.6 | ≈ 2.7 | Included usage credited at ≈ API list prices, then metered | ≈ face value in model credits; agent-heavy weeks spill into billing |
| Copilot Pro | $10 ($15 credits) | ≈ 6.4M @ gpt-5.3-codex | ≈ 3.1 | $15 of model credits included | Explicit 1.5x sticker — the cheapest honest entry point |
| Copilot Pro+ | $39 ($70 credits) | ≈ 30.0M @ gpt-5.3-codex | ≈ 14 | $70 of model credits included | Explicit ~1.8x sticker |
| Copilot Max | $100 ($200 credits) | ≈ 85.7M @ gpt-5.3-codex | ≈ 41 | $200 of model credits included | Explicit 2x sticker — best published credits multiple |
* Face value = plan budget ÷ blended list price at the 20:1 agent mix — a floor, not the ceiling. Claude plans meter 5-hour windows and weekly caps rather than token grants, and daily drivers routinely extract more than face value before hitting them; Copilot publishes its credits outright (1.5–2x sticker); Codex and Cursor credit usage at roughly API rates. Plan prices verified 2026-07-03; the full subscription-vs-API breakeven for your own workload is in the cost calculator.
The honest summary: subscriptions are cheap prepaid tokens with a fair-use fuse. If you run agents daily, a $100–200 plan beats raw API until the exact moment the limits say otherwise — which is why the resilient setup is a subscription for the grind plus per-token credits (or a frontier model on credits) for the escalations.
How to read this table if you run coding agents
Two things matter more than the headline numbers. First, input dominates: an agent re-reads files, diffs, and its own history every turn, so real sessions burn 10–30x more input than output — that's why the session column swings harder than the price columns. Second, cache rates are the real price for long sessions: an agent loop that resends stable context hits the cache discount constantly. The full reasoning — with worked examples — is in our breakdown of Fable 5's pay-per-token math, and the honest way to pick between models at the same price point is a blind test on your own codebase.
AI model pricing FAQ
Is a Claude or ChatGPT subscription cheaper than paying per token?
For daily agent use, usually yes — that is the arbitrage subscriptions sell. At list prices a $20 plan buys only ~3 heavy agent sessions on Sonnet 4.6 if you priced it as raw tokens, but subscription limits are metered in time-windows and multipliers, and daily drivers routinely extract several times face value before hitting caps. The explicit case is Copilot, which publishes it: $10 buys $15 of credits, $100 buys $200. The catch is always the same — limits, not price.
How much do AI coding models cost per million tokens in 2026?
As of July 2026 the spread is roughly 70x: Claude Fable 5 tops the table at $10 per million input tokens and $50 per million output, flagships like Claude Opus 4.8 ($5/$25) and GPT-5.4 ($2.50/$15) sit in the middle, and open-weight models run far cheaper — Kimi K3 at $3/$15, GLM-5.2 at $1.40/$4.40, down to DeepSeek V4-Flash at $0.14/$0.28.
Why does input pricing matter more than output for coding agents?
A coding agent re-reads its context — files, diffs, tool results, its own history — on every turn, so input tokens outnumber output tokens by 10-30x in real sessions. A model with cheap output but expensive input can still be expensive to run as an agent, and cache-hit input discounts (like Kimi K3’s $0.30 rate) matter more than headline prices.
What does a typical coding agent session cost?
Rule of thumb at July 2026 prices: a scoped bug fix (150k tokens in, 8k out) runs from under a cent on DeepSeek V4-Flash to about $1.90 on Claude Fable 5. A one-hour feature session (2M in, 100k out) spans roughly $0.31 to $25 across the same table. The table on this page computes it live for every model.
How often is this page updated?
Every row links to its source and carries the date it was last verified. We re-verify after major releases and monthly at minimum; if a vendor changed prices since the stamp at the top of the table, trust their page over ours — and tell us.
Prices are half the decision — the other half is where your agent runs. Backgrind is an always-on-top overlay for whichever CLI and model you pick, and the Fusion council lets an expensive model be one seat at the table instead of the whole bill.