Best budget LLM for coding (2026)
The cheapest LLMs that are still genuinely usable for coding — ranked by input-token price, filtered to models with real benchmark scores so you are not just buying cheap-and-useless tokens.
🏆 Top pick: Qwen 3 (235B)
At free to self-host (no token cost) with a 53% SWE-bench score, Qwen 3 (235B) is the cheapest model that still holds up on real coding tasks.
The ranked list
| # | Model | Input price | Output price | SWE-bench | Context window |
|---|---|---|---|---|---|
| 1 | Qwen 3 (235B) | Free (self-hosted) | Free (self-hosted) | 53% | 128K |
| 2 | Qwen 3.6 | Free (self-hosted) | Free (self-hosted) | 57% | 256K |
| 3 | Qwen 3 Coder | Free (self-hosted) | Free (self-hosted) | 50% | 256K |
| 4 | Qwen 3 Coder Next | Free (self-hosted) | Free (self-hosted) | 60% | 256K |
| 5 | Llama 4 Maverick | Free (self-hosted) | Free (self-hosted) | 46% | 1M+ |
| 6 | GPT-OSS 120B | Free (self-hosted) | Free (self-hosted) | 62% | 128K |
| 7 | GPT-OSS 20B | Free (self-hosted) | Free (self-hosted) | 61% | 128K |
| 8 | GLM 5.1 | Free (self-hosted) | Free (self-hosted) | 58% | 128K |
Why each made the list
1 Qwen 3 (235B)
Powerful self-hosted reasoning, multilingual coding, agentic workflows with MCP
2 Qwen 3.6
Agentic coding with sustained multi-turn reasoning, frontend generation, local development
3 Qwen 3 Coder
Open-source coding specialist, long-context code generation, FIM completions on local hardware
4 Qwen 3 Coder Next
Next-gen open-source coding model optimized for agentic coding and local dev workflows
5 Llama 4 Maverick
Latest open-weights from Meta, large context, self-hosted coding with vision
6 GPT-OSS 120B
Best open-weight model for production coding, configurable reasoning for quality/speed trade-offs, fine-tunable on single H100, agentic workflows
7 GPT-OSS 20B
Local development, consumer hardware, fast reasoning loops, cost-effective agentic coding, laptop-friendly open-weight model
8 GLM 5.1
Top-tier agentic engineering, complex multi-step workflows, sustained long-horizon coding, self-improving agent loops
Found your pick? Build a full stack around it — Flowpicker shows compatibility warnings before you commit.
Open the stack planner →