Dropping "Read It in Small Chunks": Claude Code v2.1.296 Swallows Huge Files in One Gulp and Halves Sonnet 5.5's Cache Cost
Claude Code v2.1.296 adds allow_large to the Read tool so huge files can be read in a single pass, and loosens the limits on MCP descriptions and retry waits. Sonnet 5.5's cache reads are halved from $0.20 to $0.10. New knobs for subagent auto-compaction and model alignment were added as well.
On October 9, 2026, Anthropic released v2.1.296 of its coding agent, Claude Code. Where the previous day's v2.1.295 was a safety-leaning update that fixed the behavior of hooks (the checkpoints for automatic execution), this release loosens, in several places, the ceiling on how much you can hand the agent at once—and quietly cut the price of part of its flagship model along the way. Many of the changes bear directly on day-to-day usability and cost, and there are also tweaks aimed at unattended operation and subagent tuning.
Letting It Read a File Whole, in One Gulp
The most intuitive change this time is the allow_large option added to the file-reading tool, Read. Until now, Claude Code would break large text files at the usual size limit and read them in several passes. With allow_large, when you need the entire file and have room to spare in the context (the amount of working context it can hold), it can read even files that exceed the limit all at once in a single call.
When handling long logs, configuration files, or huge generated artifacts, split reading tended to drop context and multiply round trips. Being able to swallow it in one gulp makes it easier to reason about the whole.
Raising Several "Ceilings" at Once
Read isn't the only place the limit on how much can be passed at once got revisited. The constraints that forced things into small chunks were loosened across multiple spots.
- MCP descriptions: The ceiling on the tool descriptions and server instructions first sent to the model in external tool integration (MCP) was doubled from 2,048 to 4,096 characters. You can now convey how to use a tool in more detail.
- Persistence under overload: A new environment variable,
CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS, was added. You can set a longer retry wait than before for when the server rejects a request due to congestion (HTTP 529). For long-running unattended use, this makes it less likely to stall just because of congestion.
Sonnet 5.5's Cache Cost Cut in Half
On the cost side, the cache-read price for the flagship model Sonnet 5.5 was lowered. Cache reads are the cost incurred when reusing context you've already sent, and they add up in long conversations and repetitive work.
| Item | Before | After |
|---|---|---|
| Sonnet 5.5 cache read (per 1M tokens) | $0.20 | $0.10 |
This price is also reflected in the /cost display, the status line, the --max-budget-usd cap setting, and the SDK's cost calculation. Having the display follow suit is a welcome change in practice, too, as it reduces the gap between estimates and actual billing.
Fine-Tuning the Subagent "Swarm"
More tuning knobs were added for operations that run multiple subagents in parallel.
autoCompactWindow: Added to subagent definitions (front matter or--agents). It triggers auto-compaction (summarizing and compressing context) at an earlier stage than the main conversation, so worker agents can lighten their load before they run out of context.CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL: Lets you align just the agents within a workflow to a single model while letting other subagents keep running on their own models. This makes it easier to use different models for different roles.
Extending the Management Net to Claude Desktop
There was also attention to enterprise governance. A code key was added to Claude's gateway setting managed.policies[], so settings that previously applied to the CLI version can now be applied in the same way to the Code tab in Claude Desktop. Rules an administrator sets in one place now reach both the terminal and the desktop.
Alongside this, in the VSCode extension, the browser-control feature "Claude in Chrome" was changed to ask for confirmation on every action, including sessions connected via @browser (once you allow a site, subsequent confirmations stop). It brings the same caution found in the terminal to in-editor operations.
The Trade-Off Behind the Convenience
Loosening the limits so you can pass more at once can, if misused, lead to higher costs and wobblier judgment. If you use allow_large to read a huge file whole, you consume that many more tokens, which could offset the benefit of the halved cache-read price. The discipline of using it only when needed is still required. The setting to extend the retry wait under overload also calls for care in where you place the ceiling, so that in unattended operation you avoid the situation of "before you knew it, it had been waiting for hours." There's still room for the operations side to redesign the balance between convenience, cost, and control.
References: Claude Code changelog (official) / Releasebot: Claude Code updates / Gradually: Claude Code Changelog
