Checked 2026-10-15 · official API list prices
Real value: what your plan would cost at API prices
I spend most days inside these tools, and I kept wondering whether my $20 plan was a steal or a trap. So I priced it out: take what each plan says you can use, multiply by real agentic-coding token counts, and charge it at the vendors' own API list prices. The answer surprised me. Where a vendor publishes the number, we use it. Where it doesn't, we estimate, show the method, and paint it orange so you never mistake a guess for a fact.
Your scenario
- Conservative (low end of the published range)
- Up to (high end)
- Fixed $ credits
- Estimate (method in each row)
| Plan | $/mo | Allowance used | Model priced | API $ per turn | API-equiv / month | Multiplier |
|---|
Free plans and unquantified quotas
- Claude Free: Chat only: Claude Code isn't on Free. Worth ≈ $0.00/mo at API prices. source
- GitHub Copilot Free: 50 chat requests a month (GitHub). Est.: priced as one Typical Sonnet 5.5 turn each; GitHub doesn't say which models Free requests reach. Worth ≈ $34/mo at API prices. source
- Kiro Free: 50 credits × Kiro's $0.04 add-on credit price. A credit is Kiro's unit of work, not an API dollar. Worth ≈ $2.00/mo at API prices. source
- JetBrains AI AI Free: 3 AI credits per 30 days; JetBrains: 1 credit = $1. Worth ≈ $3.00/mo at API prices. source
- Amp Hobby: No included agent usage; pay-as-you-go at provider prices (no markup). Worth ≈ $0.00/mo at API prices. source
- Cursor Free: Limited Grok-only quota, not quantified by Cursor. source
- Windsurf Free: Light daily/weekly quota, not quantified; free models (e.g. SWE-1.5) don't count against it. source
- OpenAI Codex Free: GPT-6 Luna in the desktop app; OpenAI publishes no Free allowance (assume ≤ Go). source
- Google Antigravity Free: 'A meaningful quota, refreshed weekly'. Not quantified. source
How the math works
API-equivalent value / month = allowance × tokens per agent turn × API list price. Everything below is checkable.
1. What a "turn" costs in tokens
One prompt you type into an agent isn't one API call. The agent reads files, runs tools and loops, so most of the tokens are context being re-read from the prompt cache. Our per-call baseline is real logged data: BSWEN: 1,289 Claude Code requests = 100.3M input (84.2M cache reads) + 616K output → per request ≈ 65.3K cached, 12.5K fresh, 0.48K output. (BSWEN, 100M tokens tracked). An independent log of 3,007 Claude Code sessions found the same pattern: cache reads outnumbered fresh input more than 1,000:1 (Hetland).
- Light: 5 model calls per turn → 327K cached + 63K fresh input + 2.4K output.
- Typical: 15 model calls per turn → 980K cached + 188K fresh input + 21.6K output (output ×3; see below).
- Heavy: 40 model calls per turn → 2612K cached + 500K fresh input + 192.0K output (output ×10; see below).
Assumptions, stated plainly: calls per turn (5 / 15 / 40) is our modeling choice. It isn't a published figure, and you can change the profile above. The output multiplier corrects a known logging gap: Claude Code's JSONL logs leave thinking tokens out of output, which undercounts output 10–17× on Opus (ccusage #866). Fresh input is priced at the base input rate. Real cache writes cost 1.25× on Anthropic, so our numbers lean slightly low.
2. Allowances (only published numbers)
- OpenAI Codex Plus ($20): OpenAI's own estimate of local messages per 5 hours: 5–45 GPT-6 Astra, 15–160 GPT-6.1 Sol, 350–3,000 GPT-6 Luna. We price whichever model gives the most value. Weekly limits "may also apply" and aren't quantified. source
- Claude Pro / Max: Anthropic no longer publishes counts. We use its last official estimate (2025 help page, archived): Claude Code prompts per 5 hours of 10–40 (Pro), 50–200 (Max 5x) and 200–800 (Max 20x). Anthropic has raised 5-hour limits twice since then (May and Sept 2026, per MorphLLM's timeline), so treat these as floors. Labeled est. on the site.
- GitHub Copilot: the dollar AI-credit allowance GitHub publishes ($15 / $70 / $200). We count it at face value. GitHub says credit use varies by model and doesn't say credits are charged at exactly API list price. source
- Every other plan uses its own row's method, shown under the plan name in the table: $ credits (Warp, JetBrains, Zed, Augment after its 40% fee, Kiro at its $0.04 add-on price), credit formulas (Z.ai's published credit multipliers × GLM-5.3 list price), call quotas (Alibaba 90K/month, MiniMax per 5h, Kimi 300–1,200 per 5h, Gemini 1,500/day; one request = one model call), per-day message tables (Windsurf), vendor multipliers applied to a base plan (Cursor 4/15/50x Plus, Codex Pro 5x/10x Plus, Antigravity Ultra 5x/20x, Factory 5x/10x) and community measurements (Antigravity, Cursor). Anything not printed by the vendor is orange / est.
- Windows per month is your input (default 20, roughly one full window per workday). Weekly caps may stop you before you reach it. Nobody publishes those caps, so the real multiplier can be lower than shown.
3. API list prices (USD per 1M tokens, standard tier)
| Model | Input | Cached input | Output | Source |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $0.2 | $20 | ↗ |
| Claude Sonnet 5.5 | $2 | $0.1 | $10 | ↗ |
| Claude Fable 5.1 | $10 | $0.25 | $50 | ↗ |
| Claude Haiku 5.5prompts ≤100K tokens; $0.50/$0.05/$2.50 above | $0.1 | $0.01 | $0.5 | ↗ |
| GPT-6 Astra≤272K context | $10 | $1 | $50 | ↗ |
| GPT-6.1 Sol | $2 | $0.1 | $10 | ↗ |
| GPT-6 Luna | $0.1 | $0.01 | $0.5 | ↗ |
| GPT-5.6 Luna | $0.2 | $0.02 | $1.2 | ↗ |
| Grok 4.7Cursor's published rate beyond the pool | $2 | $0.5 | $6 | ↗ |
| Grok 4.6 | $2 | $0.5 | $6 | ↗ |
| Grok 4.5 | $2 | $0.5 | $6 | ↗ |
| Composer 2.5 | $0.5 | $0.2 | $2.5 | ↗ |
| Gemini 4 Argonas listed by Cursor at API list price | $2 | $0.1 | $10 | ↗ |
| Claude Opus 5provider price as listed by Zed | $5 | $0.5 | $25 | ↗ |
| Claude Opus 4.6provider price as listed by Zed | $5 | $0.5 | $25 | ↗ |
| Claude Sonnet 4.5provider price as listed by Zed | $3 | $0.3 | $15 | ↗ |
| GLM-5.3in/out from Vals AI; cached est. = input × Z.ai's cached/input credit ratio (1.7/6.9) | $1.4 | $0.345 | $4.4 | ↗ |
| Kimi K2.7 Code | $0.95 | $0.19 | $4 | ↗ |
| MiniMax M2.7 | $0.3 | $0.06 | $1.2 | ↗ |
| Qwen3.7-Plus≤256K tokens; cached est. at 10% of input (not on the price page we read) | $0.4 | $0.04 | $1.6 | ↗ |
| Gemini 3.5 Flashprovider price as listed by Zed; cached est. at 10% of input | $1.5 | $0.15 | $9 | ↗ |
| SWE-2Cognition list price (75% off through Dec 31, 2026) | $3 | $0.3 | $15 | ↗ |
4. Winners
The "Real value" badge in each tier on the home page goes to the highest conservative multiplier (low end of the range, Typical turns, 20 windows a month). A badge needs at least two plans with numbers. Benchmarks tell you how smart a model is; this tells you how much of it you get for your money.
A solid ★ goes to the best verified value. If an estimate beats it, the estimate also gets a dashed ★ est., so you can see both the safest pick and the likely-best pick.
Found a better data source or measured your own cap? Send it; we replace estimates with data.