API leaderboards shift weekly—DeepSeek V3.2 still tops ultra-budget, Gemini Flash-Lite rides ecosystem defaults, but multi-turn Agents can turn "cheapest" into "most expensive." This post gives a reusable August 2026 ranking framework with blended token math and our monthly refresh policy.
How we rank "cheapest"
Most lists compare input price only or mix in OpenRouter volume (who gets called, not who costs least). Our main table uses three rules:
| Rule | Meaning |
|---|---|
| Blended cost | 60% input + 40% output per 1M tokens (USD) |
| Official list price | Vendor pages as of August 2026 |
| Sustainable | Free trials excluded; paid tiers only |
Bottom line: DeepSeek still leads ultra-budget; Gemini Flash-Lite ranks top 3 on blend; multi-turn Agents can multiply gaps 10–50×.
August 2026 main table: chat APIs
Blended = input×0.6 + output×0.4. Sources: OpenAI, Anthropic, Google AI, DeepSeek, Mistral pricing pages.
| Rank | Model | In $/1M | Out $/1M | Blend | Best for |
|---|---|---|---|---|---|
| 1 | DeepSeek V3.2 | 0.14 | 0.28 | ~0.20 | High volume, multilingual |
| 2 | Gemini 2.0 Flash-Lite | 0.075 | 0.30 | ~0.17 | Short replies, classification |
| 3 | GPT-4.1 Nano | 0.10 | 0.40 | ~0.22 | Tagging, extraction |
| 4 | Mistral Small 3.1 | 0.10 | 0.30 | ~0.18 | EU nodes, GDPR |
| 5 | Qwen3 32B | ~0.15 | ~0.60 | ~0.33 | Chinese-optimized |
| 6 | Claude Haiku 4.5 | 1.00 | 5.00 | ~2.60 | Low-latency support |
| 7 | GPT-4.1 Mini | 0.40 | 1.60 | ~0.88 | General chat |
| 8 | Gemini 2.5 Flash | 0.30 | 2.50 | ~1.18 | Multimodal |
| 9 | Claude Sonnet 4 | 3.00 | 15.00 | ~7.80 | Hard reasoning, code |
| 10 | GPT-5.4 / Opus class | 5.00+ | 25.00+ | ~13+ | Peak quality |
Tips: Ranks 1–4 are close—pick on latency and tool-call stability. Use three tiers: L1 ultra-cheap → L2 mid → L3 flagship escalation queue.
Specialty: Embedding & Batch
| Type | Leaders | $/1M |
|---|---|---|
| Embedding | Gemini / OpenAI text-embedding-3-small | 0.02–0.025 |
| Batch async | OpenAI / Google Batch | ~50% off list |
Embeddings are input-only; Batch is hours-latency, not user-facing.
Hidden costs
- Output tokens often 60–75% of chat bills.
- Agent chains ≈ 8–20× a single chat in tokens.
- Prompt cache can cut input cost 50–90% on hits.
- Aggregators (OpenRouter) may add markup; self-host trades GPU fixed cost for variable tokens.
Pick by scenario
| Product | Starter stack | Rough monthly @ DAU 100 |
|---|---|---|
| Chat / support | DeepSeek or Flash-Lite + Haiku escalation | $15–80 |
| RAG | Flash-Lite + embedding + Sonnet for hard Q | $80–350 |
| Coding Agent | Sonnet + Nano for lint | $150–600+ |
Formula: monthly tokens ≈ DAU × requests/day × tokens/request × 30; cost ≈ tokens × blend / 1e6.
Update policy
- First business day each month: recheck official prices.
- >15% price moves: footnote within 72 hours.
- 2026-08-04: DeepSeek V3.2 #1; Flash-Lite top 3 on blend; GPT-4.1 Nano cheapest Western nano tier.
Save money today
- Default small model; escalate to Sonnet
- Enable prompt caching
- Hard
max_tokenscaps - Separate embedding model
- Daily bill alerts at +50% DoD
- Dual-vendor fallback
Disclaimer: prices are public as of writing; your console bill is authoritative.
Tiered API routing—count the execution environment in TCO
Cheap tokens on an unstable VPS lose savings to retries and downtime.
Nuvcloud dedicated M4 Mac mini for Xcode, OpenClaw, and long-running Agents.