Arc: v0.6.0 Release

Arc version 0.6.0, first released on August 5, 2026.

Highlights

Auto Mode (Beta)

v0.6.0 brings the new Auto mode to Arc, which automatically routes prompts to the cheapest model that can handle the task. Relying on an in-house fine-tuned model, Auto mode makes sub-10ms decisions¹ at 0.757 AUC² based on internal testing. Crucially, Auto mode isn't limited to a specific set of models, and instead adapts to your configuration. Learn more.

Prompt Polishing

When enabled, Arc automatically polishes your prompts before sending them to the model, and can be configured to provide either grammar/spelling-only corrections or full prompt rewriting.

Chat History Encryption

Chat history is now stored in our own encrypted, compact ARCX binary format by default, providing better security and storage efficiency than JSON or SQLite.

Changelog

Chat

Auto mode is available from the model picker. During beta, the router shows you its routing option with accept/reject controls before sending the prompt, giving you the final decision.
Added prompt polishing. Polished prompts appear in the composer for approval before being sent.
You can now see the model used for subagents.
The stop button now stays visible whenever a task is running so you can cancel at any time.
The queue/steer send button is now a single accent-coloured split control, matching the main send button.

Sidebar Chat

Added a collapsible plan bar above the composer, replacing the old to-do list UI.

Fullscreen Chat

None.

Settings

Added prompt polish settings, with off, basic, and polish modes.
Added Auto quality bias, letting you route toward cheaper or stronger models.
Added notification sound settings.

Backend

Pressing stop now always ends the current task. The busy state now reliably clears and you can send again, even if the agent got stuck mid-turn.
Tasks that get repeated provider errors now end with a clear error instead of retrying forever.
History compaction no longer splits tool-call chains across the compaction boundary.
Larger sessions no longer lose older tool steps from the transcript after a restart.
Post-edit LSP diagnostics and verification results are now part of the edit's own tool result instead of separate duplicate responses, fixing some provider errors and improving efficiency.
Reverting a checkpoint now reliably removes the later snapshots, even when they were created moments apart.
File tools now reject missing arguments with a clear error instead of writing the literal text undefined into files or searching for it.
Editing files with Windows line endings no longer corrupts nearby lines when the search text has to be matched loosely.
Added configurable per-session usage caps for web.search, subagent.spawn, and mcp.call tools.
Writes to config files (.arc/, .vscode/, .cursor/, .claude/, AGENTS.md, CLAUDE.md, .arcrules*) now always require approval, even with auto-approve on.
The agent now pauses and asks after repeated failing or identical tool calls instead of getting stuck.
Messages marked noCompact (like plan instructions) now survive history compaction.
Chat history automatically migrates to the encrypted ARCX format, cleaning up the old JSON file after the first save.
Added pre.compact hook event with a block decision, plus an instructions.loaded hook event.
Workspace conventions are now also picked up from .cursorrules, .windsurfrules, .clinerules, and copilot-instructions.md.
Added ${env:VAR} interpolation support in MCP server configs.

Performance

Measured on an Intel® Core™ Ultra 9 Processor 185H, using a 5,000-prompt pool (p50 500-char / p90 1,300-char). Includes estimated data for an Intel® Core™ i5-12400F.

Metric Ultra 9 185H i5-12400F
mean 1.96 ms 2.33 ms
p50 1.57 ms 1.87 ms
p90 3.33 ms 3.96 ms
p99 7.07 ms 8.41 ms
throughput ~510 prompts/s ~429 prompts/s

Metric	Ultra 9 185H	i5-12400F
mean	1.96 ms	2.33 ms
p50	1.57 ms	1.87 ms
p90	3.33 ms	3.96 ms
p99	7.07 ms	8.41 ms
throughput	~510 prompts/s	~429 prompts/s

Accuracy

We evaluated Auto mode on 1,190 prompts, selected with seed 42 from the canonical pool of 5,947 labeled prompts.

Category	Source	AUC	n
Agentic	swechat	0.711	757
	gdpval	0.604	37
	tbench	n/a¹	19
	agenttrace	n/a¹	2
Knowledge QA	hle	0.443	88
	gpqa	0.603	140
	scicode	0.205	19
	critpt	n/a¹	8
Environment-graded	τ²-bench (trajectories)	0.993	120
Overall	all sources	0.757	1,190
Category aggregates	agentic (swechat+gdpval+tbench+agenttrace)	0.714	815
	knowledge (hle+gpqa+critpt+scicode)	0.604	255
	env-graded (τ²)	0.993	120

¹ The tbench, agenttrace, and critpt results are not reported because those groups are too small or contain only one outcome type, while across all 1,190 evaluated prompts Auto mode achieved a balanced accuracy of 0.611 and a Brier score of 0.187.