Claude Opus 4.8 is Anthropic's top coding model, released May 2026, and inside Atlas it is the model most users pin for multi-file refactors where the agent has to hold an entire service in its head and still land a clean unified diff. It runs a 1M token context window and costs $5 / $25 per Mtok (input / output), with 128K max output tokens. The tradeoff is real: it is five times the input cost of Claude Haiku 4.5 for tasks that do not need frontier reasoning.
What is Claude Opus 4.8 best at inside Atlas?
Claude Opus 4.8, released May 2026, is the model most Atlas users pin for multi-file refactors. Anthropic's top coding model can hold an entire service in its 1M token context and still land a clean unified diff, which Atlas computes and surfaces for approval before writing anything.
The refactor case is where the pairing earns its price. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, and searches it with hybrid semantic and keyword retrieval fused by reciprocal rank fusion. That retrieval hands back whole declarations rather than truncated fragments, and Claude Opus 4.8 has the 1M token context window to actually take all of them at once. When the model then writes, Atlas computes a unified diff for every file edit and surfaces it for approval, so a frontier model that reasons carefully about cross-file invariants directly reduces the number of diffs you have to reject.
How much does Claude Opus 4.8 cost per million tokens?
Claude Opus 4.8 costs $5 / $25 per Mtok (input / output). That is a third of the old Opus 4.1 rate of $15 / $75, and Anthropic paired the cut with a strictly larger context window, so the 2026 model is both cheaper per token and able to take more of them.
Read the price against the alternatives before pinning it. At $5 per Mtok input, Claude Opus 4.8 is five times the input cost of Claude Haiku 4.5, which runs $1 / $5 per Mtok. If a task is mechanical, that multiple buys you nothing. Where it does pay is in wasted turns: Atlas permission-gates every tool call against allow, ask, and deny rules, and a model that proposes the right call the first time spends fewer cycles getting rejected. The output side matters too. Claude Opus 4.8 tops out at 128K max output tokens, enough to emit a large refactor in one turn without truncation stitching, which means you are not paying twice to re-send context for a continuation.
Does Atlas enable extended thinking for Claude Opus 4.8?
Yes. Atlas ships the anthropic-beta header interleaved-thinking-2025-05-14 by default, so Claude Opus 4.8 reasons between tool results instead of only before them. In an agent loop where every tool call is permission-gated, reasoning after a result lands is worth more than reasoning once at the top.
Interleaved thinking changes the shape of a long Atlas session. Because Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and connects to Model Context Protocol servers that return tool output mid-turn, Claude Opus 4.8 keeps getting new evidence after it has already started reasoning. The interleaved-thinking-2025-05-14 header lets it fold that evidence in. The cost is latency: extended thinking adds time to every turn, which is exactly why Claude Opus 4.8 is a poor fit for the small_model slot that drives titles and summaries.
When should you pick a different model than Claude Opus 4.8?
Pick something other than Claude Opus 4.8 whenever the task does not need frontier reasoning. At $5 per Mtok input, Opus 4.8 costs five times Claude Haiku 4.5's $1, and for session titles, commit summaries, and cheap subagent fan-out that premium buys nothing you can measure.
The clean split is by slot. Leave a frontier model in the main model slot and put a cheaper model in small_model, because Atlas lets you switch the active model and provider on the fly with favorites and recents, so nothing is locked in. Claude Opus 4.8 is specifically wrong for small_model: extended thinking adds latency, and a title generator that pauses to reason is a bad trade. Inside a session, run /models to drop down for the mechanical stretch and star Claude Opus 4.8 so it is one keystroke away when the hard part starts.
How do you make Claude Opus 4.8 the default model in Atlas?
Set "model": "anthropic/claude-opus-4-8" in atlas.json and every Atlas session starts on Claude Opus 4.8. For a single session instead, open /models in the TUI, choose Anthropic, pick Claude Opus 4.8, and star it so it pins to favorites for next time.
Atlas resolves models through a registry, so confirm the entry exists before you rely on it: run atlas models anthropic and look for the Opus 4.8 row. Credentials come from either the environment or the credential store. Export your key with export ANTHROPIC_API_KEY=sk-ant-... , or run atlas login and pick Anthropic to store it in the credential store instead, which keeps the key out of your shell history and dotfiles. Both paths resolve to the same provider entry, so the /models list looks identical either way.
Setup
- 01Export your key: `export ANTHROPIC_API_KEY=sk-ant-...`, or run `atlas login` and pick Anthropic to store it in the credential store instead.
- 02Confirm the model resolved from the registry: `atlas models anthropic`.
- 03In the TUI, run `/models`, choose Anthropic, then Claude Opus 4.8. Star it so it pins to favorites.
- 04To make it the default for every session, set `"model": "anthropic/claude-opus-4-8"` in atlas.json.
- 05Leave a cheaper model in the `small_model` slot, because extended thinking adds latency and Claude Opus 4.8 is a poor fit for titles and summaries.
Frequently asked questions
- how much does claude opus 4.8 cost per million tokens
- Claude Opus 4.8 is priced at $5 / $25 per Mtok (input / output). That is a third of the old Opus 4.1 rate of $15 / $75, and it comes with a strictly larger context window.
- what is the context window of claude opus 4.8
- Claude Opus 4.8 has a 1M token context window and a 128K max output token ceiling, which is enough to emit a large refactor in one turn without truncation stitching.
- how do i use claude opus 4.8 in atlas
- Export your key with `export ANTHROPIC_API_KEY=sk-ant-...` or run `atlas login` and pick Anthropic. Confirm it resolved with `atlas models anthropic`, then run `/models` in the TUI and choose Anthropic, then Claude Opus 4.8.
- is claude opus 4.8 worth it over claude haiku 4.5
- Claude Opus 4.8 is five times the input cost of Claude Haiku 4.5, so it is worth it only when the task needs frontier reasoning. For session titles, commit summaries, and cheap subagent fan-out, Haiku 4.5 does the same job.
- does atlas support extended thinking with claude opus 4.8
- Yes. Atlas ships the anthropic-beta header interleaved-thinking-2025-05-14 by default, so Claude Opus 4.8 reasons between tool results instead of only before them.
- how do i set claude opus 4.8 as the default model in atlas
- Set `"model": "anthropic/claude-opus-4-8"` in atlas.json to make Claude Opus 4.8 the default for every session. In the TUI you can also star it in `/models` so it pins to favorites.
- should i use claude opus 4.8 as the small_model in atlas
- No. Extended thinking adds latency to Claude Opus 4.8, which makes it a poor fit for the small_model slot that drives titles and summaries. Put a fast model there instead.
- can i switch off claude opus 4.8 mid session in atlas
- Yes. Atlas lets you switch the active model and provider on the fly with favorites and recents, so run `/models` and pick another model without restarting the session.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Fiber in 2026
Atlas is a terminal-native AI coding agent for Fiber in 2026. It knows fasthttp reuses buffers, tests handlers with app.Test(), and diffs every edit first.
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.
Atlas vs. Goose: Choosing Your AI Coding Agent in 2026
Compare Atlas and Goose for 2026. Atlas offers terminal-native TUI and code-specialized features. Goose provides shareable Recipes and 70+ MCP extensions for general agentic workflows.
Atlas vs CodeGPT in 2026: A Developer's Guide to Terminal and IDE AI Agents
Comparing Atlas and CodeGPT in 2026 for developers. Atlas offers terminal-native TUI and permission-gated tools, while CodeGPT provides IDE integration and a full repo Knowledge Graph.
Atlas for Actix Web in 2026
Atlas is a terminal-native AI coding agent for Actix Web in 2026. It reads extractors and app_data, then runs cargo test and cargo clippy behind a prompt.
Atlas for Deno: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Deno. Tighten --allow-all down to real permission scopes, add JSR deps, and run deno test and deno check in 2026.
Atlas for Groovy: A Terminal-Native AI Coding Agent for Gradle, Spock, and Jenkins in 2026
Atlas is a terminal-native AI coding agent for Groovy in 2026. It reads build.gradle closures and Jenkinsfiles, writes Spock specs, runs ./gradlew test, and applies Spotless.
Atlas for Quarkus in 2026
Atlas is a terminal-native AI coding agent for Quarkus in 2026. It reads CDI beans and JAX-RS resources, then runs ./mvnw test behind a permission prompt.