Most developers leave every request on the most capable model and pay frontier prices to rename a file. The cheapest capable coding model is a fraction of that, and for a lot of work it is just as good. Here is the real 2026 pricing, the cost per solved task, and exactly which model to reach for on which job.
Quick answer
Claude Haiku 4.5 is the cheapest capable coding model, at roughly $1 per million input tokens and $5 per million output, and it is more than twice as fast as larger models. It is also among the best on cost per solved task. Use Haiku for edits, file navigation, and subagents; Sonnet 4.6 (about $3 and $15) for writing full features; and reserve Opus 4.8 (about $5 and $25) or GPT-5.5 for planning, architecture, and hard debugging. Cursor's in-house Composer is cheaper still, at about $0.50 and $2.50, if you work inside Cursor. Tokens 4 Breakfast shows your real spend per model so you can see whether an expensive default is inflating your bill.
Each step stands on its own — skip to the one that matches where you are.
Cursor's Composer runs about $0.50 per million input tokens and $2.50 output, but only inside Cursor. Claude Haiku 4.5 is roughly $1 and $5. Claude Sonnet 4.6 is about $3 and $15. Claude Opus 4.8 is about $5 and $25, and OpenAI's GPT-5.5 is about $5 and $30. Note that output tokens cost three to five times input, and coding produces a lot of output, so the output price matters more than the input price for writing code.
The number that matters is cost per solved task, not cost per token. A cheap model that needs three tries costs more than a mid model that gets it right once. On that measure Haiku 4.5 is the standout value, at roughly a dollar of output per benchmark point, because it is both cheap and genuinely capable. A pricier model only wins when it clears work the cheap one cannot, which is a smaller slice of day-to-day coding than most people assume.
Haiku 4.5 handles the bulk of everyday coding: small edits, renames, file navigation, running tests, and fanning out subagents to explore a codebase. It gives roughly Sonnet-level coding on these tasks at a fraction of the cost and more than twice the speed. If you spawn subagents, put them on Haiku by default; four cheap subagents cost less than one expensive one and do the exploration just as well.
When you are writing a whole feature from a clear spec, step up to Sonnet 4.6, or to Cursor's Composer if you live in Cursor. Both are strong implementation models at a small fraction of frontier cost, and Composer in particular beats far pricier models on real agent benchmarks. This is the tier that does the actual code writing once the plan exists.
Keep the frontier models, Opus 4.8 and GPT-5.5, for the work that genuinely needs them: understanding an unfamiliar codebase, choosing an architecture, writing the spec, and untangling a hard bug. This is where their extra capability earns the price. The mistake is using them for the routine execution that a cheaper model finishes just as well.
Picking the cheapest capable model for each task is the core of model routing, and it is worth 40 to 70 percent on a coding bill. You can do it manually by switching models, or automatically with Cursor's Auto mode or an API router. The companion guide on model routing walks through every method and the full math.
You cannot pick the cheapest model well if you cannot see where the money goes. Tokens 4 Breakfast breaks your spend down by model and by project, live in the Mac menu bar, so an Opus-heavy habit shows up as a real number. Switch your routine work to a cheaper model, then watch the per-model spend fall.
Pro tips
Short, direct answers to the things people ask most about this.
Claude Haiku 4.5, at roughly $1 per million input tokens and $5 per million output, is the cheapest capable coding model and more than twice as fast as larger models. Inside Cursor, the in-house Composer model is cheaper still at about $0.50 and $2.50, but it has no public API.
Haiku 4.5 handles most everyday coding well: edits, file navigation, tests, and subagents. Step up to Sonnet 4.6 for writing full features from a spec, and reserve Opus 4.8 or GPT-5.5 for planning, architecture, and hard debugging. Match the model to the difficulty of the task, not its importance.
Haiku 4.5 is roughly five times cheaper than Opus 4.8 per token (about $1 and $5 versus $5 and $25), and faster. For routine coding where both produce correct output, using Opus means paying about five times over for quality you did not need.
Use the cheapest capable model, usually Haiku 4.5, for subagents. Subagent work is mostly exploration and file reading, which does not need a frontier model. Keeping subagents on a cheap model, and to about four at a time, is one of the easiest ways to cut cost.
Track your spend by model. Tokens 4 Breakfast shows your real cost per model and per project live in the Mac menu bar, so you can see whether an expensive default is quietly inflating your bill and switch routine work to a cheaper model.
Privacy-first analytics
We use optional analytics to understand aggregate website usage and make the product easier to discover. Google Analytics only loads after you accept. No account data, app data, or personal AI usage is sent, and you can change your choice anytime.
Read the privacy policy