API vs subscription: which is cheaper for coding agents?

Compare the monthly API bill for your workload with the flat plan price.

Data verified 2026-10-13 from official pricing pages.

Monthly API cost by workload

Light = 25% of default, Default = 3M input + 150K output tokens/day, 60% cache hits, 22 days; Heavy = 3x default. List prices, no discounts.

Model	Light	Default	Heavy
GPT-6 Luna	$1.17	$4.69	$14
Gemini 3.5 Flash-Lite	$4.34	$17	$52
Gemini 3.8 Flash	$8.79	$35	$105
Grok Build 0.1	$10	$41	$123
Claude Haiku 5.5	$10	$41	$124
Grok 4.3	$12	$49	$148
GPT-6.1 Sol	$22	$90	$269
Gemini 4 Argon	$22	$90	$269
Grok 4.7	$23	$92	$277
GPT-5.3 Codex	$25	$99	$298
Gemini 3.1 Pro Preview	$25	$100	$301
Claude Sonnet 4.6 (legacy)	$35	$141	$422
Claude Sonnet 5.5	$41	$165	$495
Claude Opus 5.5	$45	$180	$539
GPT-5.6 Sol	$47	$187	$562
Claude Opus 5 (legacy)	$59	$234	$703
GPT-6 Astra	$117	$469	$1,406
Claude Fable 5.1	$206	$825	$2,475

Rule of thumb

If the API figure for the model you actually use is below the plan price ($20 Pro tiers, $100 and $200 Max tiers), pay per token. If it is above, the plan is cheaper, provided its limits fit. Plans: all plans ranked.

Model prices: /models/. Your own numbers: calculator.

FAQ

When is the API cheaper than a subscription?

When your monthly token cost at list prices is below the plan price. The table shows light, default and heavy workloads for every model we track.

Do plans have hidden extra costs?

Often past the allowance: Claude offers usage credits at API rates, Cursor bills third-party models at API price, and Windsurf and Devin bill extra usage at API pricing.

Does caching change the answer?

Yes. Cache hits cut input cost sharply, so the default workload assumes 60% hits. Try other rates in the caching calculator.