# Code execution tool (code_execution_20250825 · _20260120 · _20260521 · legacy _20250522)

Status: DOCUMENTED · LIVE_VERIFIED 2026-09-18 (e1 print(2+2) on haiku 4.5, e2 container reuse, e3 code_execution_20260521, g1/g2 Python cell via PTC on sonnet 4.6, k25 bad name; ≈$0.03 in tokens, container time inside the free tier). Sources: Code execution tool · Programmatic tool calling · Skills guide · Files API · Messages API. Last verified: 2026-09-18.

# Definition and versions

{"type": "code_execution_20250825", "name": "code_execution"}      // name fixed (k25: 400 "Input should be 'code_execution'")

Models: Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5, 4.8, 4.7, 4.6, 4.5, Sonnet 5, 4.6, 4.5, Haiku 4.5 (haiku: no PTC / REPL persistence — newer versions behave like 20250825). Platforms: Claude API, Claude Platform on AWS, Foundry Hosted-on-Anthropic; not Bedrock/Vertex. Not ZDR-eligible.

# Runtime

Python 3.11, Linux x86_64, 1 CPU, 5 GiB RAM, 5 GiB disk, no internet (only pre-installed libs: pandas, numpy, scipy, scikit-learn, statsmodels, matplotlib, seaborn, pyarrow, openpyxl, xlsxwriter, pillow, python-docx/pptx, pypdf, pdfplumber, reportlab, sympy, mpmath…; CLI: unzip, 7zip, bc, rg, fd, sqlite). Sandboxed per request's workspace. Whole-invocation timeout → execution_time_exceeded.

# Files in / out

• In: upload with the Files API then reference {"type":"container_upload","file_id":"file_…"} in the user content (CSV, xlsx, JSON, images, text…). Files preloaded bill container time even if the tool is not called.
• Out: files Claude leaves at the top level of $OUTPUT_DIR of a bash_code_execution call are returned as content: [{type: code_execution_output, file_id}] in the result; download via GET /v1/files/{id}/content. Media files carry C2PA Content Credentials.
• Skills: "container": {"skills": [{"type": "anthropic", "skill_id": "pptx", "version": "latest"}]} (max 20; Skills API CRUD documented by the Skills agent).

# Containers

Response top level: "container": {"id": "container_01YXRh6bhAJd4WeWsdo29Bgg", "expires_at": "2026-09-19T02:49:11.949149Z"} (live: expires_at ≈ 5 h ahead, rolling; the 30-day hard limit is not reported). Reuse by sending "container": "<id>" (or {"id": …, "skills": […]}) — e2 reused the same id and expires_at advanced. Checkpoint after ≈5 min idle, restore within 30 days; expired → error, resend without container. State independent of prompt cache. Required on continuation requests with a pending programmatic call. Sending an unknown id with a prompt that did not run code returned 200 without a container (k26) — behaviour for a real expired id remains UNVERIFIED.

# Response blocks (live e1 / g2)

{"type":"server_tool_use","id":"srvtoolu_017wg…","name":"bash_code_execution","input":{"command":"python3 -c \"print(2+2)\""}}
{"type":"bash_code_execution_tool_result","tool_use_id":"srvtoolu_017wg…","content":{"type":"bash_code_execution_result","stdout":"4\n","stderr":"","return_code":0,"content":[]}}
// text editor sub-tool: server_tool_use{name:"text_editor_code_execution", input:{command: view|create|str_replace, path, …}} ->
//   text_editor_code_execution_tool_result{content: text_editor_code_execution_view_result{file_type,content,num_lines,start_line,total_lines} | _create_result{is_file_update} | _str_replace_result{old_start,old_lines,new_start,new_lines,lines[]}}
// PTC python cell (20260120+): server_tool_use{name:"code_execution", input:{code}} ->
{"type":"code_execution_tool_result","tool_use_id":"srvtoolu_01XRk…","content":{"type":"code_execution_result","stdout":"[{'one': 1}]\n","stderr":"","return_code":0,"content":[],"abort_reason":null}}

Errors: {"type":"bash_code_execution_tool_result_error","error_code":"unavailable"}; codes unavailable, execution_time_exceeded, invalid_tool_input, too_many_requests, output_file_too_large (bash), file_not_found (text editor). abort_reason is undocumented (LIVE_DISCOVERED). pause_turn possible; mixed with client tools the code-execution result arrives after your tool_results. Streaming: sub-tool input as input_json_delta, result block whole.

Undocumented internal type ids seen in the per-model "Did you mean" error list (bash_code_execution_20250825, text_editor_code_execution_20250825, python_with_tools_code_execution_20250825/20260120/20260521) are not accepted as request tool types (k11/k12/k12b → 400).

# Usage / pricing

Docs show usage.server_tool_use.code_execution_requests; live responses (e1–e3, g1–g2) only contained web_search_requests/web_fetch_requests = 0 — no code-execution counter (discrepancy). Billing: free when web_search_20260209+ or web_fetch_20260209+ is in the request; otherwise container-hours with a 5-minute minimum, 1,550 free hours/org/month, then $0.05 per hour per container. Tokens as usual — e1 cost 4,611 input tokens on haiku (tool description is large). Container data retained 30 days; Files API outputs persist until deleted.

# Interplay

Web search/fetch dynamic filtering auto-provisions code execution (shared container); if you also declare code_execution it must be 20260120+. With a client bash tool present, tell Claude the two environments are separate. Skills load into the container. Programmatic tool calling: programmatic-tool-calling.md.

Examples: examples/anthropic/code-execution/code_execution.{sh,py,ts} (sh does the container reuse); test test_code_execution_and_container_reuse (expensive).