The cheapest LLMs for coding in 2026: real list prices

Coding-capable models now run from $0.15 to $3 per million input tokens, a twentyfold spread. Most of the saving comes from sending routine turns to the bottom of that range and keeping the top for the turns that need it.

What changed in this update. The first version of this post, in June, had a pass-rate benchmark and prices labelled "kRouter-tested". We cannot back those numbers, so they are gone. The model line-up has moved on too (GLM-5.3, Kimi K3, DeepSeek V4, MiniMax M3), and two free providers it recommended, Qwen Code and iFlow, are no longer offered in kRouter. Every price below comes from the provider's own pricing page, checked in October 2026. Prices change often, so check again before you commit.

The price list

Per million tokens, standard tier, uncached input:

DeepSeek's peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Every other hour, weekends and Chinese public holidays included, is off-peak at half price.

Two caveats before reading anything into the gaps:

• Tokens are not the same size everywhere. Each model family has its own tokenizer, so the same prompt is a different number of tokens on each. Anthropic says the tokenizer it introduced with Claude Opus 4.7 produces about 30% more tokens for the same text than the one before.
• Price says nothing about whether the model finishes your task. A cheap model that needs three attempts costs more than an expensive one that needs one. Try the candidates on your own repository before moving everything over.

What the gap looks like

Take a million input tokens and 200,000 output tokens as a unit of work. At list prices, before any caching discount, that costs:

Agents resend the same context on every turn, so much of their input can be served from cache, and every provider here bills cached input at a fraction of the uncached price. Your real bill depends heavily on how well the cache holds.

Flat-rate plans

Per-token prices are half the picture. Several providers sell coding subscriptions that cap usage by time window instead:

• OpenCode Go, $10 a month (Go Plus is $40). One subscription covers Kimi K3, GLM-5.3, DeepSeek V4 Pro and Flash, MiniMax, Qwen, MiMo and more. Each model has a monthly allowance, and any 5-hour window may use 20% of it, any week 50%. In kRouter it is the ocg/ provider, and since v0.5.163 its Responses-only models (Grok, GPT Luna, Muse Spark) work as well -- see OpenCode Go's three protocols.
• GLM Coding Plan from Z.ai, from $18 a month, metered in credits per 5 hours and per week. Requests for GLM-5.1 and GLM-5.2 are served by GLM-5.3. kRouter's glm provider takes a Z.ai key.
• Kiro: a free plan with 50 credits a month that includes Claude Sonnet 4.5 and some open-weight models, rate limited, and Pro at $20 for 1,000 credits. kRouter's kr/ provider connects to it.

Which is cheaper depends on how much you use. A flat plan buys a predictable bill; pay-as-you-go on a Flash-class model wins when usage is light.

What about the free tiers?

Free still beats every price above. These are still available:

• OpenCode Free (oc/), no account needed. kRouter fetches the model list live from OpenCode, and the list changes as OpenCode adds and removes free models.
• MiMo Code Free (mmf/), no account needed. It serves one model, mimo-auto.
• OpenRouter :free models: 20 requests a minute and 50 a day, or 1,000 a day once you have bought at least $10 of credits.
• Kiro Free, the 50 monthly credits above.

Some are gone. Alibaba discontinued Qwen Code's free OAuth tier on April 15, 2026, and kRouter hides that provider. kRouter no longer offers iFlow. Gemini CLI's free tier for personal accounts ended on June 18, 2026 -- what still works for Gemini.

The catch with free is quota and churn, so the strategy is free first, with a cheap paid tier as overflow:

1. oc/<a current free model>   # OpenCode Free, no account
2. mmf/mimo-auto               # MiMo Free, no account
3. ocg/glm-5.3                 # OpenCode Go, flat $10/month
4. ds/deepseek-v4-flash        # DeepSeek, pay per token

In a fallback combo kRouter tries the entries in order and moves to the next one when an entry is rate limited or out of quota; entries that report their remaining quota are reordered so the fullest goes first. Combos & fallback lists what each kind of error does.

Two cautions. kRouter shows a risk notice before you connect a Kiro account, because subscription sessions like it are not licensed for proxy use and an account can be restricted. And read each free provider's data terms before sending it code you cannot share.

A pure-paid combo (no free tier)

If you cannot or do not want to use free tiers (commercial workload, compliance constraints, or you just do not trust them):

1. ds/deepseek-v4-flash          # cheapest per token, off-peak especially
2. minimax/MiniMax-M3            # flat price up to 512k input
3. glm/glm-5.3                   # stronger, still well under Claude prices
4. anthropic/claude-sonnet-5-5   # safety net

Model ids after the slash are passed to the provider as written, so the last line can name any current Anthropic model id.

Choosing

• Lowest bill for routine edits: a Flash-class model on pay-as-you-go -- GLM-5.3-Flash or DeepSeek V4.1 Flash, off-peak if you can.
• Several hours a day, predictable bill: OpenCode Go at $10 a month.
• Very long prompts: MiniMax M3 keeps its base price up to 512k input tokens, and the DeepSeek models take a 1M-token context.
• Hard problems: keep one premium model at the end of the combo for the turns that need it.

Install

npm install -g @sifxprime/krouter
krouter -t

The provider catalog is at /providers, and combo strategies are in /docs/combos.

Common questions

What is the cheapest LLM for coding right now?

By list price per token, GLM-5.3-Flash and DeepSeek V4.1 Flash, both at $0.15 per million input tokens (DeepSeek's off-peak rate), as of October 2026. Free tiers such as OpenCode Free and MiMo Code Free cost nothing, but their model lists and limits change without notice.

Is a flat-rate plan cheaper than paying per token?

It depends on volume. OpenCode Go costs $10 a month whether you use it or not, with per-model limits per 5 hours and per week. At the list prices above, $10 buys about 37 units of a million input and 200,000 output tokens on DeepSeek V4.1 Flash off-peak, so light users usually come out ahead paying per token, while heavy daily use is easier to budget on the flat plan.

Are the free tiers safe to rely on?

Not as your only backend. Qwen Code's free tier and Gemini CLI's free tier both ended in 2026. Put free providers first in a combo and keep a cheap paid provider behind them, so a free tier closing costs you a little money instead of a stopped task.

Why does the same task cost different amounts on models with similar prices?

Tokenizers differ, so the same text is a different number of tokens on each model. Cache hit rates differ. And a model that needs more attempts or longer answers spends more tokens on the same task.

Related

• Free AI coding assistant: what is actually free
• OpenCode Go in Claude Code and Codex
• OpenRouter free tier 429 and the 50-a-day cap
• Claude Code on Kimi K3, GLM-5.3 and DeepSeek V4