Quick answer: DeepSeek V4 Flash has the lowest standard API list price here. OpenAI GPT-5.6 Luna is the lowest-priced OpenAI option, while Claude Haiku 4.5 is Anthropic’s budget model.
That does not make any one model the cheapest for every coding task. A low token rate can lose its advantage after failed patches, long reasoning output, tool calls, and retries.
Muse here means Meta Muse Spark 1.2 and the Muse Code coding agent.
Cheap AI coding models compared
Prices are vendor list prices per one million tokens unless marked as a monthly plan. Currencies are not converted. Promotions, caches, tools, routing, and taxes can change the final bill.
| Model or service | Type | Context | Input / output price | Best fit to test |
|---|---|---|---|---|
| DeepSeek V4 Flash | API model | 1M | $0.14 / $0.28 | High-volume, price-sensitive coding |
| DeepSeek V4 Pro | API model | 1M | $0.435 / $0.87 | Harder agentic coding tasks |
| OpenAI GPT-5.6 Luna | API and Codex model | 1.05M | $0.20 / $1.20 | Lower-cost Codex and subagent work |
| OpenAI GPT-5.3-Codex | Coding model | 400K | $1.75 / $14 | Agentic work in Codex-like harnesses |
| Claude Haiku 4.5 | API model | 200K | $1 / $5 | Fast, bounded coding tasks |
| Claude Sonnet 5 | API model | 1M | $2 / $10 promo | Longer coding and agent workflows |
| GLM-5.2 | Model and Coding Plan | 1M | From ¥49/month | Project-level coding in supported tools |
| Kimi K2.7 Code | Coding API model | 256K | ¥6.50 / ¥27 | Thinking-mode coding with multimodal input |
| Kimi K3 | API model | 1M | ¥20 / ¥100 | Long-horizon programming |
| Meta Muse Spark 1.2 | Coding model | 1.05M | Check live Meta pricing | Muse Code and multi-agent workflows |
Claude Sonnet 5’s $2/$10 price is promotional through 31 August 2026. Anthropic lists $3/$15 from 1 September. GLM’s stable public API token price was unavailable, so this table does not invent one.
What the price table leaves out
API prices and coding-product subscriptions do not buy the same thing. Codex, Claude Code, GLM Coding Plan, Kimi Code, and Muse Code can bundle quotas, tools, credits, or agent-generated tokens.
Cached input can be much cheaper than fresh input. Reasoning output, tool definitions, tool results, delegated workers, web search, and failed fallback attempts can raise the actual cost.
Cost per successful task = total measured spend across every attempt ÷ verified successful tasks
Count failed attempts and retries. A completion is successful only when the task’s tests, verifier, or acceptance check passes.
DeepSeek: the lowest list price in this comparison
DeepSeek V4 Flash and Pro both expose 1M context windows. Flash lists $0.14 input and $0.28 output; Pro lists $0.435 and $0.87.
DeepSeek positions Flash for speed and economy and Pro for stronger agentic coding. Treat those as vendor claims until both models run through the same held-out tasks in the same harness.
OpenAI Codex: separate the product from the model
Codex is a coding-agent product that can use several models. Its credit rate card meters supported work by actual token usage.
GPT-5.6 Luna costs $0.20 input and $1.20 output and targets cost-sensitive work. GPT-5.3-Codex costs $1.75 and $14 and is optimized for agentic coding.
Measure delegated-worker and reasoning usage instead of assuming the API rate equals the price of one completed Codex task.
Claude: Haiku for price, Sonnet for harder work
Anthropic’s own guide says Haiku falls further behind on long coding tasks than stronger models.
Haiku 4.5 lists $1 input and $5 output. Sonnet 5 targets coding and agent workflows, with a 1M context window and promotional $2/$10 pricing through 31 August 2026.
A cheap-model test should include both short fixes and long repository tasks. That reveals whether Haiku’s lower rate survives additional retries.
GLM: a coding subscription with quota rules
GLM-5.2 supports a 1M context window, tool calling, MCP, caching, structured output, and thinking modes.
The official Coding Plan starts at ¥49 per month, with Pro at ¥149 and Max at ¥469. Prompt figures are estimates because one prompt can trigger many model calls.
Use the plan’s live quota and supported-tool terms. Do not convert estimated prompts into a guaranteed per-request price.
Kimi: compare API models and Kimi Code separately
Kimi K2.7 Code is a 256K thinking-only coding model at ¥6.50 input and ¥27 output. The faster HighSpeed service doubles those rates.
Kimi K3 targets long-horizon programming with 1M context at ¥20 input and ¥100 output. Kimi Code membership is a separate quota product with its own endpoint and tier limits.
Meta Muse: coding model and agent
Muse Spark 1.2 is Meta’s coding model. Muse Code is the terminal agent built to run it across large repositories and multi-agent workflows.
Meta documents a roughly 1.05M context window but does not expose a stable public price table on the cited product page. Check the live Meta Model API price before purchasing.
How to choose the cheapest useful coding model
- Choose real tasks. Include the repository repairs, terminal work, and browser checks you expect the agent to perform.
- Pin the setup. Keep the model version, harness, tools, context, environment, timeout, provider, and retry policy comparable.
- Capture every attempt. Retain usage, spend, duration, tool calls, retries, verifier output, and the final trace.
- Compare within task families. Do not average autocomplete, repository repair, and long autonomous work into one score.
- Rerun close results. Recheck after model, price, prompt, skill, or harness changes.
| Work surface | Evidence to retain | ClawBench benchmark |
|---|---|---|
| Repository repair | Accepted patch and verifier result | SWE-Bench Verified |
| Terminal work | Commands, files, and final state | Terminal Bench |
| Browser workflow | Navigation, recovery, and outcome | Web Tasks |
| Fast baseline | Repeatable mixed-domain result | ClawBench Entry Test |
Cheap AI coding model FAQ
Which coding model is cheapest after retries?
DeepSeek V4 Flash has the lowest standard list price in this dated comparison. Your cheapest option is the one with the lowest measured cost per verified success on your tasks.
Does a larger context window make a coding model better?
No. Context size is a limit, not a quality score. Sending more repository context can increase cost and reduce focus.
Does the agent harness change model performance?
Yes. Tools, context assembly, timeouts, retries, and verification can change both success and cost. Compare models inside the same pinned harness before drawing a conclusion.
Should I choose API pricing or a coding subscription?
Choose the unit that matches your workflow. Compare API usage for embedded agents and product quotas for interactive coding, then calculate completed-task cost for each.
Compare models with trace-backed evidence
Use the same task family and harness, then inspect the run behind every score. The lowest token price is only a starting hypothesis.