Which LLM has the lowest input price in 2026?

Gemini 2.5 Flash-Lite at $0.075 per million input tokens is the cheapest credibly-rated model from a major provider. Some hosted Llama 3.2 1B deployments go even lower (~$0.05) but quality drops to ~15–20 — usable for embeddings-adjacent tasks but not for general inference.

Is DeepSeek R1 really production-ready at $0.55 per million tokens?

Yes. R1 scores 73 on the blended TokenRate Quality index, putting it in the balanced tier — comparable to Claude Sonnet 4 at $3 input. The catch is higher latency from chain-of-thought tokens and a more limited ecosystem (fewer official SDKs, smaller community).

How do I filter the TokenRate calculator to under-$1 models?

Click 'Filters' on the calculator, then pick the 'Under $1' chip under 'Input cost / 1M tokens'. Stack with 'Good (50+)' under Quality score to filter further to the production-viable subset.

What's the catch with under-$1 LLMs?

Three things: (1) Output cost is sometimes high relative to input (DeepSeek R1 is $2.19/1M output); (2) Quality on hard problems drops below flagships — multi-step reasoning, novel coding, complex math; (3) Ecosystems are less mature for some — open-source models often lack official SDKs for every language. Test before shipping.

Best LLMs Under $1 Per Million Tokens in 2026

The Under-$1 Tier Has Become Production-Viable

Two years ago 'cheap' LLMs meant 'demo only' — quality was too low for anything user-facing. In 2026 that's no longer true: a dozen models priced under $1 per million input tokens now score 50–70 on the blended Arena AI + Artificial Analysis Quality index, putting them comfortably in production-grade territory for routine workloads. That's the whole reason TokenRate added a one-click 'Under $1' cost preset to the Filter panel — it's now a real production tier, not a budget afterthought. This post ranks the best of that tier by Quality × Value (the quality per dollar metric) so you can see which sub-$1 models are actually worth shipping with.

The Top Picks at Under $1 Input

As of May 2026, ranked by quality score (descending) with input price for context: DeepSeek R1 — quality 73, $0.55 input (the bargain reasoning king); GPT-5 mini — quality 68, $0.30 input (OpenAI's value flagship); Gemini 2.5 Flash — quality 66, $0.15 input (Google's volume play); Claude Haiku 4.5 — quality 65, $0.25 input (Anthropic's fast tier); DeepSeek V3 — quality 65, $0.30 input (non-reasoning sibling of R1); Llama 4 Maverick — quality 62, $0.20 input; Llama 4 Scout — quality 55, $0.10 input; Gemini 2.5 Flash-Lite — quality 55, $0.075 input (cheapest credible quality); Mistral Small — quality 42, $0.20 input (good for European data-residency requirements); Qwen 2.5 32B — quality 48, $0.30 input. To see these ranked live and sortable, open the calculator, click Filters → 'Under $1', then sort by 'highest quality' or 'best value'.

Picking Between the Top Three

If your workload includes reasoning (multi-step planning, code synthesis, math), DeepSeek R1 is the clear winner — quality 73 at $0.55 input puts it within 10 points of $10+ flagships at 5% of the cost. If your workload is general chat, summarization, or RAG with short context, Gemini 2.5 Flash and Claude Haiku 4.5 split the win: Flash is cheaper ($0.15 vs $0.25 input) and has 1M context vs Haiku's 200K, but Haiku has slightly better instruction-following and is part of Anthropic's tooling ecosystem. If your workload is very high volume but quality requirements are modest (classification, intent detection, simple extraction), Gemini 2.5 Flash-Lite at $0.075 input is unbeatable. For the side-by-side comparison, use /tools/compare-prices to pick all three and see input/output/context in one grid. See also Gemini Flash vs GPT-4o-mini budget model.

Output Cost Matters at This Tier Too

Under-$1 input pricing tells half the story. Output costs at this tier range from $0.30/1M (Gemini Flash-Lite) to $4.40/1M (o3-mini, which slips just above $1 input depending on your routing). For generation-heavy workloads — content creation, code generation, long-form writing — output cost dominates and rankings shift. DeepSeek R1's $2.19 output is high relative to its input, so its true value depends on your input:output ratio. Claude Haiku 4.5's $1.25 output keeps it efficient on chat. Gemini 2.5 Flash-Lite's $0.30 output is the absolute floor. The Compare Prices view shows input and output columns side by side so you can apply your specific token ratio. For background on the input/output split, see input vs output token ratio optimization and output token pricing explained.

Combining Under-$1 Models With Caching and Batching

Under-$1 input pricing already cuts your bill by 70–95% vs flagships. Add prompt caching (90% discount on repeated context) and batch APIs (50% discount on async traffic) and the effective rate drops to single-digit cents per million tokens. A workload that costs $5,000/month on Claude Opus 4 can drop to $50/month on cached, batched Gemini Flash-Lite — a 100x reduction — without dropping below acceptable production quality for routine tasks. The right architecture is tiered routing: cheap model as default, escalate to a balanced or flagship only when the cheap model fails a quality check. For implementation, see multi-model routing with quality scores and building cost-aware AI agents.