Cheapest LLM API (at public list prices)
“Cheapest” depends on the shape of your traffic. Chatbots that mostly answer in a few hundred tokens should optimize for low output $/1M. Retrieval-heavy apps that paste big contexts every turn may care more about input $/1M or effective caching. Free-tier rows with $0 list input and output can be real—but vendors still enforce quotas, rate limits, and fair-use rules that invoices do not show.
July 30, 2026: GPT-5.6 Luna redraws the cheap tier
OpenAI cut GPT-5.6 Luna API list prices by 80% to $0.20 input / $1.20 output per 1M tokens (cached input $0.02). Terra dropped 20% to $2 / $12; Sol stayed flat. At those numbers Luna undercuts most “nano / flash / mini” SKUs that still quote above $1/M output — while keeping a 1.05M context window and tool-calling support.
Confirm figures on OpenAI’s pricing page before you publish a forecast. Aggregator catalogs (including OpenRouter) sometimes lag a day or more after a cut — our AI model pricing tracker applies OpenAI’s official list for the GPT-5.6 family when the catalog is stale.
Model IDs: gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol. The bare alias gpt-5.6 routes to Sol (the expensive tier) — do not assume it means Luna.
Workflow that actually saves money
1) In the pricing table, set sort to Output price ↑ and modality to Text. Put Luna next to DeepSeek, Gemini Flash, and Claude Haiku on the same shortlist. 2) Run the same hundred production prompts through each—measure latency and failure rate, not vibes. 3) Multiply measured tokens by list $/1M from the table (or the cost calculator); if you batch or cache, re-check whether your vendor bills those lanes separately. 4) Sleep on “too cheap” models—sometimes the cheapest row is an old snapshot id that routers map differently, or a catalog that has not absorbed yesterday’s price cut.
Free tiers vs always-on APIs
List-price-zero SKUs are useful for experiments and low-volume prototypes, but production systems should plan for paid lanes with explicit rate limits. Treat zero-price rows as “start here,” not “scale forever,” unless your contract says otherwise. Luna is paid API — cheap enough that many teams will prefer it over $0 rows that throttle unpredictably.
When cheapest is wrong
Medical, financial, or safety-critical flows may need stronger models regardless of list price—saving cents per request while risking brand or liability is a bad trade. Use the cheap tier (Luna, Haiku, Flash, DeepSeek) for volume; keep Terra/Sol or Claude Opus on deck for escalation paths. Luna’s long-context recall also degrades earlier than Sol/Terra in independent testing — do not treat “1M context” as usable recall across the whole window.
Automation angle
If your goal is predictable monthly cost instead of optimizing every token, CloudyBot plans bundle AI work with hard caps—useful when you want scheduling and delivery without running your own token meter. The pricing table still helps you sanity-check what raw APIs would have cost for the same workload shape.
Related
GPT-5.6 pricing · Luna vs Haiku · Luna vs DeepSeek · OpenAI vs Anthropic · GPT-4o vs Claude Sonnet · AI cost calculator. All CloudyBot figures are aggregated catalog estimates (with OpenAI official overrides for GPT-5.6); confirm on each vendor console before you publish a forecast externally.