Claude Haiku 4.5 is Anthropic's small, fast model at $1 / $5 per Mtok (input / output) with a 200K token context window. In Atlas its natural home is the small_model slot, where it handles session titles, commit summaries, and cheap subagent fan-out. Haiku 4.5 supports reasoning and tool calls, so it is a real agent model rather than a chat-only cheap tier, but its 200K context is not the 1M window of Sonnet 5 and Opus 4.8.
What is the small_model slot in Atlas and why does Claude Haiku 4.5 belong there?
The small_model slot in Atlas is the cheap lane for background work, and Claude Haiku 4.5 at $1 / $5 per Mtok is built for it. Session titles, commit summaries, and cheap subagent fan-out all run there, and none of them need a frontier model's price.
Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and it reads git branches, status, and diffs and can stage and create commits on your behalf. Every one of those actions produces small model calls: name this session, summarize this diff, classify this file. Routed to a frontier model, that background chatter quietly becomes a meaningful share of the bill. Routed to Claude Haiku 4.5, it costs one fifth the input price of Opus 4.8, which is the difference between affordable and painful when subagents fan out in parallel.
Is Claude Haiku 4.5 a real agent model or just a cheap chat tier?
Claude Haiku 4.5 supports reasoning and tool calls, which makes it a real agent model rather than a chat-only cheap tier. That distinction is load-bearing in Atlas: a subagent that cannot call tools cannot search the index, read a file, or propose an edit, so it cannot do the job at any price.
Because Claude Haiku 4.5 can call tools, an Atlas subagent running on it can actually participate in the loop. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and Haiku 4.5 emits well-formed calls that clear the gate. It also carries 64K output tokens despite the small-model positioning, which is enough for a substantial summary or a focused patch. The reasoning support is what separates it from older cheap tiers that could only autocomplete.
What does the 200K context window cost you with Claude Haiku 4.5?
Claude Haiku 4.5 has a 200K token context window, not the 1M window of Sonnet 5 and Opus 4.8. In Atlas that means Haiku 4.5 cannot swallow a whole large repo index in one turn, so it is the wrong model for a task that needs the full codebase resident at once.
Context is the hard ceiling here. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion and indexes by AST declarations using tree-sitter, so retrieval hands back real declarations rather than arbitrary line windows. That retrieval is efficient, but a genuinely large service still exceeds 200K tokens. Claude Haiku 4.5 also degrades faster than Sonnet 5 on multi-step refactors that require holding several invariants at once. If your task is one of those, pay for the 1M window instead of fighting the 200K one.
How much cheaper is Claude Haiku 4.5 than Claude Opus 4.8?
Claude Haiku 4.5 costs $1 / $5 per Mtok, one fifth the input price of Claude Opus 4.8 at $5 / $25 per Mtok. Across a session where Atlas fans out subagents in parallel, that five-times gap on input is the difference between affordable and painful.
The gap compounds with fan-out, not with a single prompt. One Haiku 4.5 call saves cents. Ten parallel background subagents, each re-sending context every turn, is where $1 per Mtok instead of $5 per Mtok becomes visible on the invoice. The usual configuration keeps a frontier model in the model slot for the work you actually review and Claude Haiku 4.5 in small_model for everything else, so the expensive model only sees the turns that need it.
Can you run a whole Atlas session on Claude Haiku 4.5?
Yes, you can run an entire Atlas session on Claude Haiku 4.5 by switching the active model with /models, and at $1 / $5 per Mtok it is the cheapest way to work. The caveat is that Haiku 4.5 degrades faster than Sonnet 5 on multi-step refactors holding several invariants at once.
A full Haiku 4.5 session is a reasonable choice for narrow, well-scoped work: a single-file fix, a test you already know how to write, an exploration pass over an unfamiliar directory. Atlas lets you switch the active model and provider on the fly with favorites and recents, so you can start cheap on Claude Haiku 4.5 and escalate the moment the task turns into a cross-cutting refactor. Atlas snapshots file changes as git patches, so edits can be diffed and rolled back if the cheap model gets it wrong.
Setup
- 01Set `ANTHROPIC_API_KEY`, or authenticate with `atlas login`.
- 02Verify availability: `atlas models anthropic`.
- 03Wire it to the cheap slot in atlas.json: `"small_model": "anthropic/claude-haiku-4-5"` while leaving `model` on a frontier model.
- 04Switch the active model mid-session with `/models` when you want the whole session cheap.
- 05Escalate back to Sonnet 5 or Opus 4.8 before any multi-step refactor that has to hold several invariants at once.
Frequently asked questions
- how much does claude haiku 4.5 cost per million tokens
- Claude Haiku 4.5 costs $1 / $5 per Mtok (input / output), which is one fifth the input price of Claude Opus 4.8 at $5 per Mtok.
- what is claude haiku 4.5's context window
- Claude Haiku 4.5 has a 200K token context window and 64K output tokens. That is smaller than the 1M window of Sonnet 5 and Opus 4.8.
- how do i set claude haiku 4.5 as the small model in atlas
- Add `"small_model": "anthropic/claude-haiku-4-5"` to atlas.json and leave `model` on a frontier model. Set `ANTHROPIC_API_KEY` first, or authenticate with `atlas login`.
- can claude haiku 4.5 use tools and reasoning
- Yes. Claude Haiku 4.5 supports reasoning and tool calls, so it is a real agent model rather than a chat-only cheap tier, and it can drive Atlas subagents.
- is claude haiku 4.5 good for large refactors
- No. Claude Haiku 4.5 degrades faster than Sonnet 5 on multi-step refactors that require holding several invariants at once, and its 200K context cannot swallow a whole large repo index in one turn.
- what does atlas use the small_model slot for
- Atlas routes session titles, commit summaries, and cheap subagent fan-out through the small_model slot, which is why Claude Haiku 4.5 at $1 / $5 per Mtok is a good fit there.
- how do i check that claude haiku 4.5 is available in atlas
- Run `atlas models anthropic` to verify availability from the registry, then open `/models` in the TUI to select it for the current session.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Magic.dev: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent with permission-gated tools and diff review, against Magic.dev's research claims of 100 million token context models in 2026.
Atlas for Dart in 2026
Adopt Atlas, the terminal-native AI coding agent, for Dart development in 2026. Enhance productivity with intelligent code search, refactoring, and robust safety features across your Dart projects.
Atlas vs OpenHands: Terminal AI Coding Agents in 2026
Atlas and OpenHands comparison for 2026. Explore terminal-native TUI, self-hosting, code indexing, change review, and pricing models for AI coding agents.
Atlas vs Crush: Terminal AI Coding Agents in 2026
Comparing Atlas and Crush, two terminal AI coding agents for developers in 2026. Atlas offers robust planning and diffing, while Crush features LSP integration and mid-session model switching.
Atlas for TypeScript in 2026
In 2026, TypeScript developers leverage Atlas, the terminal-native AI coding agent, to enhance productivity. Atlas understands your types, ensures code quality, and offers robust safety features.
Atlas for Fortran: fpm.toml, Explicit Interfaces, and fprettify in 2026
Atlas is a terminal-native AI coding agent for Fortran in 2026. It reads modules, explicit interfaces, and intent declarations, runs fpm test behind a prompt, and runs fprettify.
Diagnose a Hanging or Long-Running Command with Atlas in 2026
How to diagnose a hanging command with Atlas in 2026: the bash tool races every command against a timeout and tells you whether it is slow or blocked on input.
Atlas vs Cursor: terminal AI coding agents compared (2026)
A grounded 2026 comparison of Atlas and Cursor across workflow, change review, extensibility, and pricing for developers choosing an AI coding agent.