Aider 0.70 shipped this week with a meaningful update to architect mode. The core change: model routing per task type rather than the previous architect/editor split. Plus refinements to how aider handles cost ceilings within a session.

For BYOK users with strong feelings about model selection, this is the most useful aider release in some time.

What’s new

The previous architect mode used two models: an architect (for planning) and an editor (for diff generation). 0.70 expands this to four roles:

# .aider.conf.yml
model: claude-3-5-sonnet      # default for chat and main work
weak-model: claude-3-5-haiku  # used for commit messages, summaries
architect-model: o1-preview   # used for planning in architect mode
editor-model: claude-3-5-haiku # used for diff generation

You can configure each role independently. The model isn’t tied to provider — you can mix Anthropic and OpenAI freely (with both API keys configured).

Why the four-way split

The roles have different requirements:

• Default model: balanced reasoning + code generation; the workhorse
• Weak model: cheap, fast, doesn’t need to reason about code; commit messages, brief summaries
• Architect model: heavy reasoning; can be slow and expensive if the planning is high-leverage
• Editor model: fast diff generation; needs to follow plan precisely; doesn’t need creativity

A single model can’t be optimal for all four. Splitting lets you put the right model on each task. The actual cost of an aider session can drop substantially without quality loss.

Concrete cost impact

For a typical 30-minute aider session:

The all-Claude default is the baseline. Mixing models can drop cost 40-65% with comparable or better quality, depending on the task and the model selection.

The catch: you need to know what to put where. Anthropic-default users might not have set up an OpenAI account; they need to do that to use o1 for the architect role. The friction is real even though the savings are real.

What’s new beyond model routing

Other 0.70 improvements:

Per-session budget caps. You can set --budget 5.00 and aider stops the session if it would exceed the budget. Replaces the prior workaround of monitoring spend in the provider’s console.

Better failure recovery. When a model call fails (provider down, rate limited), aider’s retry logic is more resilient. Previously, a single transient failure could end the session; now it backs off and retries.

Improved repo-map precision. The repo map (aider’s automatic context loading) is more selective on large repos. The previous version sometimes loaded unrelated files; the new version is tighter.

Streaming output for the editor model. Diffs now stream as they generate. For users who like to read the diff as it’s being produced, this is more responsive.

What architect mode is for

For users who haven’t tried architect mode: it’s the workflow where one model produces a plan and another model produces the diffs. The plan describes what to change at a high level; the diffs implement the plan.

This is most useful for:

• Multi-file refactors where the plan is the hard part
• Tasks where you want to review the plan before code is touched
• Sessions where you want to use a strong reasoning model for thinking and a fast model for implementation

Single-file edits or quick bug fixes don’t benefit from architect mode. The overhead of two model calls outweighs the benefit when the planning is trivial.

Compatibility with existing config

0.70 is backwards-compatible. If you have an existing .aider.conf.yml, it keeps working. The new fields are optional. The old model and weak-model fields still apply as before.

To opt into the new four-way split, add architect-model and editor-model lines. The defaults if you don’t add them: architect-model = your model, editor-model = your weak-model. This matches the prior behavior.

What I’d configure

A few configurations that work well based on early experimentation:

For Anthropic-only users:

model: claude-3-5-sonnet
weak-model: claude-3-5-haiku
# architect-model defaults to claude-3-5-sonnet
# editor-model defaults to claude-3-5-haiku

For mixed-provider with focus on reasoning:

model: claude-3-5-sonnet
weak-model: claude-3-5-haiku
architect-model: o1-preview
editor-model: claude-3-5-haiku

For mixed-provider with focus on cost:

model: claude-3-5-sonnet
weak-model: claude-3-5-haiku
architect-model: claude-3-5-sonnet
editor-model: gemini-1.5-flash

The third option uses Gemini Flash for editor (very cheap, fast). Tradeoff: occasional diff format errors that aider has to retry. For sessions where cost matters more than time, it’s a reasonable choice.

What’s coming

The release notes hint at upcoming features:

• Per-task model selection (“use Claude for this prompt, Gemini for the next”)
• Better integration with local models (Ollama)
• Multi-language repo map improvements

Aider remains an open-source project with an active maintainer. The release pace is steady; new versions every 2-4 weeks. For a community-driven CLI tool, this level of polish and forward motion is unusual.

Worth upgrading?

Yes, especially if you’re using architect mode. The cost savings alone justify the upgrade for any regular user. The per-session budget cap is a useful safety net. The retry improvements reduce friction.

For users sticking with single-model configs, the upgrade is fine but less impactful. The new features are mostly orthogonal to the basic chat workflow.

Update via pip install --upgrade aider-chat (or your preferred install method). Existing configs continue to work without changes.