What is the best-value coding model available through an API?

As of 19 July 2026, **GPT-5.6 Terra is the best general default for coding capability per unit cost**. It costs **$2.50 per million input tokens and $15 per million output tokens**, retains the full 1.05-million-token context window of the GPT-5.6 family, and OpenAI reports that it performs just above Claude Fable 5 on the Artificial Analysis Coding Agent Index.\n\nFor more cost-sensitive workloads:\n\n- **GPT-5.6 Luna** is the better high-volume coding-agent option at $1 input and $6 output per million tokens;\n- **Gemini 3.5 Flash** is a strong low-cost multimodal and agentic alternative at $1.50 input and $9 output per million tokens;\n- **Claude Sonnet 5** is especially attractive during its introductory $2 input and $10 output pricing, which lasts until 31 August 2026.\n\nPrice per token is not the same as cost per completed task, so the final decision should be based on successful-task cost, retries, latency, and human correction time.

A cheap model can become expensive if it needs repeated prompts, produces large reasoning traces, fails late in a workflow, or creates defects that require human repair. A more expensive model can be cheaper overall if it completes difficult tasks correctly on the first attempt.\n\nGPT-5.6 Terra is the best balanced default because it combines near-frontier coding-agent results with mid-tier pricing. OpenAI reports an Artificial Analysis Coding Agent Index score of **77.4**, compared with **77.2** for Claude Fable 5, and a Terminal-Bench 2.1 score of **87.4%**. These are vendor-presented results and must be verified in the intended harness, but they place Terra on a strong capability-cost frontier.\n\nGPT-5.6 Luna is better suited to routine code generation, classification, test creation, repository summarisation, and parallel sub-agent work where a large number of calls matters more than obtaining the highest possible success rate on every difficult task.\n\nGemini 3.5 Flash is useful when the workflow includes images, audio, video, PDFs, search grounding, code execution, or rapid agent loops. It provides a one-million-token context window and broad tool support at a lower list price than Terra.\n\nClaude Sonnet 5 is a credible alternative for teams already using Claude Code or Anthropic's tooling. Anthropic describes it as close to Opus 4.8 on some agentic workloads, with adjustable effort. Its introductory price makes it particularly competitive before 1 September 2026; the published standard price after that date is $3 input and $15 output per million tokens.\n\nA proper cost evaluation should record:\n\n1. total input, cached-input, reasoning, and output tokens;\n2. tool-call and search charges;\n3. success on the first run and after retries;\n4. wall-clock time;\n5. the amount of developer review and correction;\n6. defect escape rate;\n7. cache hit rate;\n8. the percentage of tasks escalated to a more capable model.\n\nThe strongest production pattern is usually model routing: a lower-cost model handles routine work, while difficult or high-risk tasks are escalated to a frontier model.

- [OpenAI — GPT-5.6 model catalogue](https://developers.openai.com/api/docs/models) — establishes current Sol, Terra, and Luna pricing, context windows, and intended capability tiers. - [OpenAI — GPT-5.6](https://openai.com/index/gpt-5-6/) — reports Terra and Luna coding-agent benchmark results and current family pricing. - [Google — Gemini 3.5 Flash model documentation](https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash) — establishes the model's token limits, modalities, agentic positioning, and supported tools. - [Google — Gemini Developer API pricing](https://ai.google.dev/gemini-api/docs/pricing) — establishes Gemini 3.5 Flash's current API prices. - [Anthropic — Claude Sonnet 5](https://www.anthropic.com/news/claude-sonnet-5) — establishes Sonnet 5's agentic positioning and introductory and standard API pricing.