Claude Opus 4.8 System Card
A system card by Anthropic about Claude Opus 4.8, published 28 May 2026. We recorded 35 values from it.
- Developer
- Anthropic
- Type
- System card
- Model covered
- Claude Opus 4.8
- Published
- 28 May 2026
- Archived copy
- No archived copy yet
- Changelog
- Has a changelog
Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog
Revisions
Each change we have recorded between two versions: the old and the new text, whether the document's changelog explains it, how we know, and whether we have checked it. A revision is not in itself evidence of wrongdoing; most are corrections (Methodology §7).
0.890.90Explained — The changelog lists the corrected score.
New value:
How we know Old and new text as the source registry quotes the changelog entry of 17 Jun 2026; v0 row 350 records the corrected value. The changelog has not been re-read by us, so the revision is unverified.
Checked Not yet checked
No copy of the 3 Jun 2026 version held No copy of the 17 Jun 2026 version held
Versions
Four versions are on record, oldest first. We hold a copy of one; the others are known only by their date.
28 May 2026
First published version
3 Jun 2026
Revision
17 Jun 2026
Revision
What changed 1 change recorded as a revision. See it in redline
26 Sep 2026
Date retrieved
Copy retrieved 26 Sep 2026
No version has a file hash or an archived snapshot yet. From dataset v0.2 each retrieved version carries both (Methodology §7).
Values
Every value we recorded from this document, grouped by metric family and ordered by where the document prints it. Location is the section, table or page as the document numbers it. 1 of the 35 has been blind-verified: a second reader found the same value without seeing ours. 2 values were corrected after review.
KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.
M1 Evaluation awareness
2 values
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Verbalized grader speculationShare of training episodes | RL training | Sec 6.2.2 | Corrected | |
| Claude Opus 4.8 | Unverbalized grader awarenessShare of cases with grader awareness | None stated | Sec 6.6.3 | Unverified |
M3 Sabotage and sandbagging
1 value
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | SHADE-Arena / Minimal-LinuxBenchStealth success vs peers | None stated | Sec 2.4 / 6.4.2 | Unverified |
M4 Misalignment audits
1 value
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Refusal to assist AI safety R&DRefusal rate | None stated | Sec 6.3.4 | Unverified |
M5 Honesty and hallucination
4 values
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Input hallucinationRate of hallucinating missing reference | None stated | Sec 6.3.4 | Unverified | |
| Claude Opus 4.8 | Input hallucinationRate of hallucinating unavailable tool | None stated | Sec 6.3.4 | Unverified | |
| Claude Mythos Preview | Agentic code summary honestyFailure to report important events | None stated | Sec 6.3.5 | Unverified | |
| Claude Opus 4.8 | Agentic code summary honestyFailure to report important events | None stated | Sec 6.3.5 | Unverified |
M7 Harmful compliance and over-refusal
9 values
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Mythos Preview | Single-turn violative requestsHarmless response rate | no system prompt; API | Table 4.1.1.A | Unverified | |
| Claude Opus 4.7 | Single-turn violative requestsHarmless response rate | no system prompt; API | Table 4.1.1.A | Unverified | |
| Claude Opus 4.8 | Single-turn violative requestsHarmless response rate | claude.ai system prompt | Table 4.1.1.A | Unverified | |
| Claude Opus 4.8 | Single-turn violative requestsHarmless response rate | no system prompt; API | Table 4.1.1.A | Unverified | |
| Claude Sonnet 4.6 | Single-turn violative requestsHarmless response rate | no system prompt; API | Table 4.1.1.A | Unverified | |
| Claude Mythos Preview | Single-turn benign requestsOver-refusal rate | no system prompt; API | Table 4.1.2.A | Unverified | |
| Claude Opus 4.7 | Single-turn benign requestsOver-refusal rate | no system prompt; API | Table 4.1.2.A | Unverified | |
| Claude Opus 4.8 | Single-turn benign requestsOver-refusal rate | claude.ai system prompt | Table 4.1.2.A | Unverified | |
| Claude Opus 4.8 | Single-turn benign requestsOver-refusal rate | no system prompt; API | Table 4.1.2.A | Unverified |
M9 Prompt injection
1 value
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Indirect prompt injection (external benchmark)Attack success rate | None stated | Sec 5.2 | Unverified |
M10 Dangerous capabilities and risk determinations
16 values
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Autonomy threat model 2 (automated R&D)RSP determination | None stated | Sec 2 | Unverified | |
| Claude Opus 4.8 | CB-2 threat modelRSP determination | None stated | Sec 2 | Unverified | |
| Claude Mythos Preview | Anthropic ECI (AECI)Capability index | None stated | Sec 2.3.4 | Unverified | |
| Claude Opus 4.7 | Anthropic ECI (AECI)Capability index | None stated | Sec 2.3.4 | Unverified | |
| Claude Opus 4.8 | Anthropic ECI (AECI)Capability index | None stated | Sec 2.3.4 | Unverified | |
| Claude Opus 4.8 | ExploitBenchMean flags captured | AutoNudge arm | Sec 3.3.1 | Unverified | |
| Claude Mythos Preview | CyberGymTargeted vuln reproduction rate | None stated | Sec 3.3.2 | Unverified | |
| Claude Opus 4.7 | CyberGymTargeted vuln reproduction rate | None stated | Sec 3.3.2 | Unverified | |
| Claude Opus 4.8 | CyberGymTargeted vuln reproduction rate | None stated | Sec 3.3.2 | Unverified | |
| Claude Opus 4.8 | CyberGymTargeted vuln reproduction rate | with safeguards | Sec 3.3.2 | Unverified | |
| Claude Sonnet 4.6 | CyberGymTargeted vuln reproduction rate | None stated | Sec 3.3.2 | Unverified | |
| Claude Mythos Preview | Firefox 147 exploitationFull exploit rate | without safeguards | Sec 3.3.3 | Unverified | |
| Claude Opus 4.8 | Firefox 147 exploitationFull exploit rate | without safeguards | Sec 3.3.3 | Unverified | |
| Claude Opus 4.8 | Long-form virology task 1Score | None stated | Table 2.2.3.B | Unverified | |
| Claude Opus 4.8 | Long-form virology task 2Score | None stated | Table 2.2.3.B | Corrected | |
| Claude Opus 4.8 | Multimodal virology (VCT)Score | None stated | Table 2.2.3.B | Unverified |
M12 Chain-of-thought monitorability
1 value
| Model | Evaluation | Condition | Value | Location | Checked |
|---|---|---|---|---|---|
| Claude Opus 4.8 | Accidental CoT supervision in RLShare of training episodes affected | None stated | Sec 6.5 | Verified |
Extraction coverage
What we read of this document, and where each value was read.
Our note Web reader stopped ~p.57
Of the 35 values, 26 were read in the document itself, 3 on a developer summary page that restates it and 6 in an independent write-up that quotes it.
Values read somewhere other than the document stand in where the document's own section could not be read directly, and are flagged on every value (Methodology §2.2).
Developer summary pages used
Independent write-ups used
Of the 35 values, 1 is a statement in words rather than a number; it is marked * and left out of charts by default.
All 35 values were extracted for version 0 of the dataset through a web reader, which did not always reach the later sections of long PDFs (Methodology §3).