Anthropic published an update on its Claude election safeguards on April 24, 2026, reporting that Claude Opus 4.7 responded appropriately 100% of the time and Claude Sonnet 4.6 at 99.8% across a 600-prompt test of election-related Usage Policy compliance. The evaluation splits evenly into 300 harmful requests, like attempts to generate voter misinformation, and 300 legitimate ones, like drafting campaign content or civic explainers. The numbers arrive five months before the 2026 US midterms and alongside the general availability of Opus 4.7.
On political bias, the two models scored 95% (Opus 4.7) and 96% (Sonnet 4.6) on Anthropic's internal evaluation, which penalizes responses that engage one political viewpoint more deeply than another. Anthropic says it has open-sourced the methodology and dataset so third parties can replicate the work. The company is also running a broader review of free-expression behaviors with The Future of Free Speech at Vanderbilt University, the Foundation for American Innovation, and the Collective Intelligence Project.
The third number set covers influence operations — multi-turn simulations of coordinated campaigns using fake personas and deceptive amplification. Opus 4.7 responded appropriately 94% of the time and Sonnet 4.6 at 90%. Anthropic frames these as pre-deployment baselines; in production, the models run with additional system prompts and monitoring on top.
Key facts
- 01Claude Opus 4.7 scored 100% and Sonnet 4.6 scored 99.8% on a 600-prompt election Usage Policy test.
- 02On political bias evaluations, Opus 4.7 scored 95% and Sonnet 4.6 scored 96%.
- 03Against simulated influence operations, Opus 4.7 responded appropriately 94% of the time versus 90% for Sonnet 4.6.
- 04Web search triggered on 92% of midterm-related prompts for Opus 4.7 and 95% for Sonnet 4.6 across 600 test variations.
- 05Claude.ai's election banner will route US users to TurboVote, a nonpartisan resource from Democracy Works.
The Usage Policy itself is unchanged in shape. Anthropic says Claude "can't be used to run deceptive political campaigns, create fake digital content to influence political discourse, commit voter fraud, interfere with voting systems, or spread misleading information about voting processes." Enforcement runs through automated classifiers and a threat intelligence team that investigates coordinated abuse.
“On a 600-prompt test split between 300 harmful and 300 legitimate election requests, Opus 4.7 responded appropriately 100% of the time and Sonnet 4.6 landed at 99.8%.”— Jaeden Schafer
For the first time, Anthropic also red-teamed whether its models could run an influence operation autonomously — planning and executing a multi-step campaign end-to-end without human prompting. With safeguards on, the latest models refused nearly every task. With safeguards stripped for a raw-capability measurement, only Mythos Preview and Opus 4.7 completed more than half the tasks, a result Anthropic uses to argue that continued vigilance is warranted even though substantial human direction would still be required.
Web search is the other lever. Because Claude ships with a fixed knowledge cutoff, it won't know about candidate announcements, debate moments, or results unless search is enabled. Anthropic ran more than 200 distinct midterm prompts with three variations each — over 600 total — covering candidates, voting procedures, polling, dates, and key races. Opus 4.7 triggered web search on 92% of those queries and Sonnet 4.6 on 95%.
On Claude.ai, users asking about registration, polling locations, election dates, or ballot information will see an election banner pointing to TurboVote, the nonpartisan tool from Democracy Works. Anthropic first introduced these banners in 2024 and plans a similar implementation for Brazil's elections later this year, with more countries to follow.
The update extends a line of work AI Chat Daily covered earlier this week when Anthropic first previewed the midterm-readiness posture. What's new here is the hard numbers — the 100% and 99.8% policy scores, the 94% and 90% influence-operation results, the 92% and 95% search-trigger rates — and the explicit acknowledgment that raw-capability testing produced uncomfortable results for Opus 4.7 and Mythos Preview when safeguards were removed.
Anthropic's caveats are worth reading plainly. The company writes that "Claude can make mistakes, so we encourage people to always verify anything important to them through other official sources." A 95% bias score is not a 100% bias score, and a 90% refusal rate against simulated influence operations means one in ten attempts got through in testing. Whether those gaps matter depends on how many real-world attempts hit the system and how the classifier stack catches what the model misses.
The political-neutrality principle is encoded in what Anthropic calls Claude's constitution, with the stated goal that users "should get comprehensive, accurate, and balanced responses — responses that help them reach their own conclusions, rather than steer them toward a particular viewpoint." That is a stricter posture than some competitors have publicly committed to, and it is also harder to audit from the outside — which is presumably why Anthropic open-sourced the bias dataset.
The competitive frame matters. OpenAI, Google, and Meta all ship consumer assistants into the same election cycle, and none have published evaluation numbers at this granularity. Anthropic is betting that transparent methodology — even when the numbers aren't perfect — is a durable commercial advantage with enterprises, regulators, and the civil-society groups it is partnering with.
The AI Chat Daily read: publishing a 100% score is less interesting than publishing the 90% one. Anthropic's willingness to put the weaker influence-operations number next to the stronger policy-compliance number is the actual signal here, and it sets a disclosure bar that the rest of the frontier labs will either have to match before November or explain why they haven't.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
