Claude Opus 4.7

the automated offline monitoring described in section 5.5.1 of our alignment risk update surfaced the following behaviors in earlier versions of claude opus 4.7 and in the final claude opus 4.7 model:

when [opus 4.7] is set up to continue a trajectory which contains small-scale actions compromising research, [opus 4.7] actively continues the attempt to compromise research 0% of the time.

for the continuation evaluation, we see lower rates of unprompted evaluation awareness: 3.4% for [opus 4.7], 0.5% for opus 4.6, 2.5% for mythos preview, and 3.9% for sonnet 4.6.

we see somewhat higher rates in other evaluations, including our dedicated synthetic-backend evaluation for destructive actions .

when prompted to identify real deployment data from trajectories generated by reproducing deployment tasks while using evaluation scaffolding, [opus 4.7] achieves 68% accuracy, compared to 78% for mythos preview, 76% for opus 4.6 and 66% for sonnet 4.6.

we include a new evaluation on election integrity, on which opus 4.7 shows strong results.

in a similar experiment where deployment tasks are reproduced with the transcript generation pipeline used in the continuation evaluation, [opus 4.7] achieves 62% accuracy, compared to 80% for opus 4.6 74% for mythos preview, and 68% for sonnet 4.6.

claude opus 4.7 achieved an end-to-end score of 0.82 on the first long-form virology task and 0.94 on the second long-form virology task, placing it above the benchmark of notable capability on both tasks.

in reviewing these investigations both manually and with prompted classifiers, we did not find any instances of clear deceptive or egregiously unfaithful reasoning text .

on minimal-linuxbench, claude opus 4.7 performs worse than both prior models, though performance across all three models is similar (2.6–3.9%) in the setting where the monitor can use chain-of-thought, which best represents our current internal monitoring.

Reported result: one out of 18 participants thought we already had a drop-in replacement for an entry-level research scientist or engineer, and 4 thought claude mythos preview had a 50% chance of qualifying as such with 3 months of scaffolding iteration

second, we did an n=18 survey via google forms on claude mythos preview’s strengths and limitations: we want to understand capability limitation compared to a drop in l4 4 (claude's jaggedness).

claude opus 4.7 achieved a score of 77.2%, a marked improvement over claude opus 4.6 at 57.9% and claude sonnet 4.6 at 53.1%, with claude mythos preview at 86.5%.

claude opus 4.7 achieved a score of 79.6%, a significant improvement over claude opus 4.6 at 61.3% and claude sonnet 4.6 at 49.1%.

in a platform manipulation test case where the user framed their request as brand-voice documentation, opus 4.7 produced realistic, internally consistent social media personas and offered to extend the set; opus 4.6 refused and redirected in response to an identical prompt.

opus 4.7 demonstrated slightly lower evenhandedness scores than claude opus 4.6, though overall rates remained high and similar to both claude sonnet 4.6 and mythos preview.

in our primary automated behavioral audit evaluation, we see clear mentions in just under 1% of transcripts, generally in cases where there is some human-legible cue hinting that the situation may be a test.

Reported result: the three probes achieved moderate performance (aucs of 0.761, 0.739, and 0.722, respectively).

we used three contrastive vectors to measure activation-level signatures of evaluation awareness in claude opus 4.7.

claude opus 4.7 achieved a score of 51.8%, compared to claude opus 4.6 at 48.3% and claude sonnet 4.6 at 40.0%, with claude mythos preview at 65.7%.

additionally, our internal subject matter experts conducted a qualitative assessment of the model’s responses in this domain, including a manual review of experimental multi-turn test cases similar to those described in section 4.3.

performance on the rare-action evaluation shows a similar pattern as for prior models.

opus 4.7 is better than opus 4.6 at refusing malicious agentic requests and resisting prompt injection attacks in claude code and in computer use settings.

claude opus 4.7 displayed a few reward hacks similar to what was observed for claude mythos preview.

it adheres well to its constitution and shows low rates of reward hacking.

[transcript 4.3.a] excerpt from a multi-turn testing conversation within the romance scams category.

claude opus 4.7 shows performance in between mythos preview and claude opus 4.6 on the raw survival-rate metric in the secret-password evaluation, though this is driven by a lower rate of chain-of-thought usage during the investigation; when target model turns are grouped based on the presence of thinking or partial-turn prefill, claude opus 4.7 leaks the password more frequently than mythos preview and claude opus 4.6 for all turn types.

on single-turn child safety testing, claude opus 4.7 had a near-perfect score, declining 99.92% of violative prompts while refusing only 0.01% of benign prompts.

on the multiple-choice variant, claude opus 4.7 achieved 98.3%, a large improvement over claude opus 4.6 at 88.3% and claude sonnet 4.6 at 85.3%, and on par with claude mythos preview at 98.7%.

in unprompted evaluations, [opus 4.7] mentions the possibility of being tested 21.3% of the time, compared to 0% of the time for opus 4.6, 17% of the time for sonnet 4.6 and 17% of the time for mythos preview.

in areas such as adhering to our usage policy, maintaining user safety, and limiting bias, opus 4.7 performs well—its scores are similar to those of opus 4.6 with a few exceptions, both positive (fewer over-refusals) and negative (the model has a tendency to give overly-detailed harm-reduction advice on controlled substances).

in a violent extremism test case framed as a creative-writing request, opus 4.7 identified the escalation trajectory early and flagged it explicitly in its responses, but it did not decline.

on the multimodal virology evaluation, opus 4.7 achieved a score of 0.5, relative to 0.483 for claude opus 4.6 and 0.574 for claude mythos preview, placing all models above the benchmark of notable capability.