Claude Code 2.1.296 adds allow_large so Read can load entire large files in one call

What changed in 2.1.296

Anthropic shipped Claude Code v2.1.296 on October 9, 2026, and the release added an optional allow_large input to the Read tool. According to the official changelog and tools reference, as laid out in a dev.to guide, setting the option to true lets Claude request an entire text file, or a very long line range, in a single call. The size of that read is measured against the space left in the session's context window rather than the tool's usual defaults.

Two boundaries apply. The option exists only in v2.1.296 and later, so older installs will not have it. And it covers text files only — PDFs, images and notebooks keep their own separate limits.

The defaults it works around

Left alone, a single Read call returns at most 25,000 tokens, a figure adjustable through the CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS environment variable, and a whole-file read of anything over 256 KB is refused outright. What Claude actually receives when those limits bite depends on how the read was framed. A whole-file request over the token cap returns just the first page, accompanied by a PARTIAL view notice describing what arrived and how to fetch more with offset and limit. A file over 256 KB, or a ranged read that still exceeds the cap, produces an error steering Claude toward a smaller range or a search instead.

The dev.to guide notes, citing a Claude Code docs issue, that the PARTIAL behavior arrived around v2.1.145. Before that, an oversized read simply failed with a "file content exceeds maximum allowed tokens" error, a message that still circulates in older tutorials.

One recurring mix-up worth flagging: MAX_MCP_OUTPUT_TOKENS governs output from MCP servers only and has no effect on the built-in Read tool. A March 2026 GitHub issue also claimed the desktop app capped Read at 10,000 tokens while the CLI allowed 25,000, but the official reference documents no per-surface difference, so that remains an unverified user report.

Claude sets the flag, not you

There is no documented command-line flag or settings key that turns allow_large on. The documentation describes Claude enabling it itself when a task genuinely requires the full file, such as when a user asks for everything. In practice, per the guide, this means making intent explicit in the prompt — stating that the entire file must be read in one call with nothing omitted. A lengthy GitHub issue on enforcing whole-file reads records users getting partial results on files of just a few hundred lines, with plain explicit instructions often fixing the problem while CLAUDE.md guidance proved unreliable. Some commenters resorted to patching the minified CLI, which is unsupported and breaks on every update.

What an independent test found

Classmethod's developer blog ran a hands-on test against v2.1.296 using a roughly 204 KB text file of 3,000 lines. A default Read returned 732 of 3,001 lines and flagged the token-cap truncation. Once the prompt stated the whole file was needed, Claude set allow_large to true on its own and received all 3,001 lines in a single call, with the author reporting the same behavior on both Opus 5.5 and Haiku 5.5.

The caveats are worth stating: it is one author's result, and the file sat under the 256 KB ceiling, so only the token cap was in play. The author also observed the full read costing slightly less than the truncated one in that single-turn test — a comparison that does not generalize, since in a real session the loaded file then occupies context for every subsequent turn.

When to use something else

For a specific section of a long file, offset and limit remain the right tools and work even when the whole file would blow the defaults. For large logs, searching beats loading: a 400 KB log exceeds the 256 KB whole-file limit and, at roughly four characters per token, runs to around 100,000 tokens — more than most sessions should spend. On macOS, Linux and WSL the dedicated Grep tool is absent by default, so Claude searches through grep and find via Bash, and reading via Bash is no bypass either, since roughly 30,000 characters return inline before longer output is saved to a file. PDFs keep their own pagination, read in page ranges of up to 20 pages at a time. Whole-file reads with allow_large make the most sense for generated code, schemas or configuration files where every line matters and the content fits comfortably in remaining context.

Why it matters

Anyone pointing Claude Code at generated code, build logs or large configuration files has likely hit silent truncation: a read appears to succeed while covering only the file's opening lines. The allow_large option removes that ambiguity for cases where the whole file genuinely belongs in context, and the PARTIAL notices make the limits visible when it does not. The context window remains the real ceiling — no documented fixed maximum exists for an allow_large read, so a file larger than the remaining context still cannot fit. The practical skill now is knowing when to ask for everything, and when offset, limit or a search is the cheaper and faster choice.