Claude Sonnet 4.5 is the September 2025 Anthropic Sonnet, and the first Anthropic model in the registry to carry a 1M tokens (1,000,000) context. Max output stayed at 64K, and pricing is $3 per Mtok input, $15 per Mtok output. Inside Atlas, the terminal-native AI coding agent, Claude Sonnet 4.5 is the cheapest way to hold a very large codebase in context on a reasoning class Claude. It was long the default coding workhorse, strong at multi file edits driven by the Atlas unified diff review loop, and it remains a common pin when a team wants a frozen, well understood model snapshot.
Why was Claude Sonnet 4.5 the first Claude with a 1M token window?
Claude Sonnet 4.5 shipped its 1M tokens (1,000,000) window in September 2025, a full quarter before any Opus had one, since Opus was still at 200K at the time. That inversion made the mid tier Sonnet, not the frontier Opus, the long context option on Anthropic for months.
The unusual thing about Claude Sonnet 4.5 is that the big window landed on the cheaper tier first. At $3 per Mtok input and $15 per Mtok output, Claude Sonnet 4.5 is the cheapest way to hold a very large codebase in context on a reasoning class Claude. Inside Atlas that combination is practical rather than theoretical: Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so what fills the 1M token window of Claude Sonnet 4.5 is coherent declarations. A wide read across many services on Claude Sonnet 4.5 costs a third of what the same read would cost on a frontier Opus tier.
What is the 64K max output limit of Claude Sonnet 4.5 in practice?
Claude Sonnet 4.5 caps max output at 64K tokens, half the 128K ceiling of Sonnet 4.6 and Sonnet 5. The read side is huge at 1M tokens (1,000,000), but the write side is the older, smaller number, and that asymmetry shapes how you should scope Atlas tasks.
The practical consequence of the 64K output ceiling on Claude Sonnet 4.5 is that very wide patches want splitting. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a task that needs to rewrite twenty files will produce several diffs rather than one, and each one gets its own approval. That is not a failure mode, it is just a rhythm: plan wide, write narrow. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which is the natural place to break the work into 64K sized chunks. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, so a partial multi patch run stays recoverable.
How do you keep subagent fan out cheap on Claude Sonnet 4.5?
Set "small_model" to a Haiku in atlas.json so subagent fan out stays cheap while Claude Sonnet 4.5 handles the main loop at $3 per Mtok input, $15 per Mtok output. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and each one bills separately.
Subagent economics are where a Claude Sonnet 4.5 configuration goes right or wrong. The main loop model deserves the reasoning class, but titling, summarizing, and small mechanical lookups do not, and every parallel background session on Claude Sonnet 4.5 pays the $15 per Mtok output rate for work a Haiku could do. Pointing the small_model slot at a Haiku keeps the fan out cheap while leaving the multi file edit work on Claude Sonnet 4.5, where it belongs. Atlas lets you switch the active model and provider on the fly with favorites and recents, so verify the active model in the TUI status line after switching with /models, and confirm the main slot really is Claude Sonnet 4.5 before starting a long run.
When should you pick Sonnet 5 instead of Claude Sonnet 4.5?
Pick Sonnet 5 when you want current defaults, because Sonnet 5 is both newer and cheaper at $2 per Mtok input and $10 per Mtok output, against the $3 / $15 of Claude Sonnet 4.5. In 2026 Claude Sonnet 4.5 is mostly a pin for reproducibility rather than a fresh choice.
The case for staying on Claude Sonnet 4.5 in 2026 is a narrow but real one: a frozen snapshot. Atlas surfaces both the plain claude-sonnet-4-5 id and the dated claude-sonnet-4-5-20250929 through `atlas models anthropic`, and pinning the dated id in atlas.json freezes the exact model behind a build. Teams with regression suites, evaluation baselines, or compliance reasons to keep model behavior fixed get value from that. Everyone else should note that Sonnet 5 undercuts Claude Sonnet 4.5 on price and beats it on max output, 128K against 64K. Use /models in the Atlas TUI to run both on the same repo task and compare the cost readout before you commit a default to atlas.json.
Setup
- 01Export ANTHROPIC_API_KEY in your shell, or store it with `atlas login`.
- 02Run `atlas models anthropic` and pick between claude-sonnet-4-5 and the dated claude-sonnet-4-5-20250929.
- 03Set "model": "anthropic/claude-sonnet-4-5-20250929" in atlas.json to freeze the exact snapshot.
- 04Set "small_model" to a Haiku so subagent fan out stays cheap.
- 05Verify the active model in the TUI status line after switching with /models.
Frequently asked questions
- what is the context window of claude sonnet 4.5
- Claude Sonnet 4.5 has a 1M token context window (1,000,000 tokens). Released in September 2025, it was the first Anthropic model in the registry to carry a 1M window, while Opus was still at 200K.
- how much does claude sonnet 4.5 cost per million tokens
- Claude Sonnet 4.5 is priced at $3 per Mtok input and $15 per Mtok output. Sonnet 5 later undercut it at $2/$10, so Claude Sonnet 4.5 is now the more expensive of the two.
- what is the max output of claude sonnet 4.5
- Claude Sonnet 4.5 caps max output at 64K tokens, half the 128K ceiling of Sonnet 4.6 and Sonnet 5. Wide multi file patches in Atlas therefore arrive as several diffs rather than one.
- how do I pin claude sonnet 4.5 to a dated snapshot in atlas
- Run `atlas models anthropic`, then set "model": "anthropic/claude-sonnet-4-5-20250929" in atlas.json. The dated id freezes the exact snapshot instead of tracking the rolling claude-sonnet-4-5 alias.
- how do I make atlas subagents cheaper on claude sonnet 4.5
- Set "small_model" to a Haiku in atlas.json. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and routing that fan out away from the $15 per Mtok output of Claude Sonnet 4.5 keeps it cheap.
- is claude sonnet 4.5 still good for coding in 2026
- Claude Sonnet 4.5 remains strong at multi file edits driven by the Atlas unified diff review loop, but Sonnet 5 is newer and cheaper at $2/$10. Use Claude Sonnet 4.5 when you need a fixed, reproducible snapshot.
- how do I confirm atlas is actually running claude sonnet 4.5
- Verify the active model in the Atlas TUI status line after switching with /models. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the status line is the source of truth mid session.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas for Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
Atlas vs Amp: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Amp, Sourcegraph's agent featuring Oracle and Orbs, for developers in 2026.
Atlas for Flutter in 2026
Discover Atlas for Flutter in 2026. This terminal-native AI coding agent helps Flutter developers build apps faster and safer, integrating with widgets, state, and the Dart toolchain.
Atlas vs Roo Code: Terminal AI Coding Agents in 2026
Comparing Atlas and Roo Code in 2026. Atlas offers terminal-native TUI, permission-gated tools, and diff review. Roo Code, a VS Code extension, shut down May 15, 2026.
Locate Where a Behavior Is Implemented with Atlas in 2026
How to locate where a behavior is implemented with Atlas in 2026: codebase_search for meaning, grep for exact text, and the lsp tool for the symbol graph.
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 for Blazor: Terminal-Native AI Coding for .razor Components in 2026
Atlas is a terminal-native AI coding agent for Blazor developers in 2026. Work across .razor components, render modes, and the C# and JS interop boundary safely.
Atlas for SolidJS: A Terminal-Native AI Coding Agent Built on Solid, for Solid, in 2026
Atlas is a terminal-native AI coding agent for SolidJS in 2026. It reads your createSignal and createMemo graph, fixes destructured props that break reactivity, and runs vitest.