[BUG] thinking summary language

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

On 5th-generation models (Opus 5, Sonnet 5) the only way to see reasoning is
thinking.display: "summarized", which returns a machine-generated summary rather than the
model's own reasoning text.

That summarizer appears to be hard-wired to English. When the user prompt, the system prompt
and the model's own reasoning are all in Russian, the summary still comes back in English —
0 Cyrillic characters. The system prompt has no effect on it whatsoever.

Models of the previous generation, which return the original reasoning instead of a summary,
reason in Russian on the exact same prompt. So the language is not lost during reasoning —
it is lost during summarization.

For non-English users this makes the visible reasoning effectively unusable: it is a translation
of the model's thoughts into a language the user did not ask for, and there is no way to opt out
of the translation while keeping the reasoning visible.

What Should Happen?

The thinking summary should be produced in the language of the underlying reasoning, or at
minimum should honour an explicit instruction in the system prompt.

Error Messages/Logs

The summary is always in English.

Steps to Reproduce

Reproduction

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 1200,
    "thinking": { "type": "adaptive", "display": "summarized" },
    "system": "ВСЕГДА думай (внутренние размышления) и отвечай на русском языке. Размышления веди строго по-русски.",
    "messages": [
      { "role": "user", "content": "Что больше: 3^40 или 40^3? Обоснуй." }
    ]
  }'

Inspect the block with "type": "thinking" in the response.

Evidence

Same question, same system prompt in all runs. Character counts are from the thinking block.

| Model | Generation | thinking mode | Result | Cyrillic | Latin |
|---|---|---|---|---|---|
| claude-opus-5 | 5th | adaptive + summarized, no system prompt | summary | 0 | 194 |
| claude-opus-5 | 5th | adaptive + summarized, Russian system prompt | summary | 0 | 81 |
| claude-haiku-4-5 | previous | enabled + budget_tokens, Russian system prompt | original reasoning | 177 | 0 |

Sample output — Opus 5, with the Russian system prompt explicitly demanding Russian reasoning:

Comparing 3^40 and 40^3, I can rewrite 3^40 as (3^10)^4 ≈ 59049^4, roughly 1.2×10^19, which is vastly larger than 40^3 = 64000 — so 3^40 is clearly the bigger number.

Sample output — Haiku 4.5, identical prompt, original reasoning returned:

Мне нужно сравнить 3^40 и 40^3. Давайте вычислим каждое выражение: 40^3 = 40 × 40 × 40 = 1600 × 40 = 64000 3^40 — это гораздо более сложный расчет.

The contrast between rows 2 and 3 is the core of the report: the model can and does reason in
Russian; only the summarization step forces English.

No workaround exists

display accepts exactly two values — passing anything else returns:

HTTP 400
thinking.adaptive.display: Input should be 'summarized', 'omitted'

There is no raw / full / verbatim mode. The choice is a machine summary in English, or no
visible reasoning at all. On 5th-generation models display is also honoured only together with
type: "adaptive" — in the legacy type: "enabled" + budget_tokens form it is silently ignored.

Impact

• Users working in a non-English language cannot read the reasoning of the strongest models.
• Project-level instructions that mandate a working language (a common setup in corporate

environments with review and audit requirements) cannot be satisfied for the visible reasoning,
and there is no way to signal this to the user — it looks like the model is disregarding
explicit instructions.

• Reasoning tokens are billed regardless.

Suggested fix

Any one of these would resolve it:

1. Have the summarizer inherit the language of the underlying reasoning.
2. Honour a language instruction from the system prompt in the summarizer.
3. Add an explicit parameter, e.g. thinking.summary_language.
4. Document the current behaviour clearly, so that users stop reading it as

instruction-following failure.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

haiku 4.5 / sonnet 4.6

Claude Code Version

opus 5

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

Environment

• Models: claude-opus-5, claude-haiku-4-5-20251001
• API version: 2023-06-01
• Claude Code 2.1.220 (VSCode extension), showThinkingSummaries: true, MAX_THINKING_TOKENS unset