ClaimsDario Amodei

“we find that Claude 3.5 Sonnet in particular remains much better on some other key tasks, such as real-world coding”

Dario Amodei, Anthropic CEO, essay author 13 October 2022 On DeepSeek and Export Controls (darioamodei.com)

Factual AI capabilities

Anthropic finds Claude 3.5 Sonnet remains much better than DeepSeek-V3 on some key tasks such as real-world coding.

Unresolvable The public record could not settle it.

Factual: Statements about the present or the past, checked against the record: did they hold up?

Said for an organisation they lead or work for; it counts on both records.

What would count as true or false

The recorded date (2022-10-13) comes before both models existed. The quote matches Amodei's essay 'On DeepSeek and Export Controls' (January 2025), so judge it as of when the statement was actually published. The claim is about what Anthropic found: that Claude 3.5 Sonnet was much better than DeepSeek-V3 on some key tasks, including real-world coding. True: Anthropic's internal or published evaluations, or widely cited real-world coding benchmarks of that time (e.g. SWE-bench Verified, Aider, LMSYS coding arena), show Claude 3.5 Sonnet clearly outperforming DeepSeek-V3 on real-world coding. Mostly true: Claude 3.5 Sonnet leads on real-world coding, but the margin is modest rather than 'much better', or the support for 'other key tasks' is thin. Mixed: the benchmarks are split, with Claude ahead on some real-world coding measures and DeepSeek-V3 ahead on others. Mostly false: DeepSeek-V3 roughly matched or beat Claude 3.5 Sonnet on most real-world coding measures, with only a marginal Claude edge on some. False: DeepSeek-V3 clearly outperformed Claude 3.5 Sonnet on real-world coding.

Set before the research began.

Evidence (6)

  1. E1 Computer use, Claude 3.5 Sonnet, and Claude 3.5 Haiku Anthropic · 22 October 2024 Own record

    “On coding, it improves performance on SWE-bench Verified from 33.4% to 49.0%, scoring higher than all publicly available models—including reasoning models like OpenAI o1-preview and specialized systems designed for agentic coding.”

  2. E2 Claude SWE-Bench performance Anthropic · 6 January 2025 Own record

    “No model has yet crossed 50% completion on SWE-bench Verified (though the updated Claude 3.5 Sonnet is, at the time of writing, at 49%)”

  3. E3 Claude 3.5 Sonnet vs DeepSeek V3: Benchmarks & Cost BenchLM · 4 November 2026

    “Claude 3.5 Sonnet is not ranked on the public lane for coding, so no winner is named for coding.”

  4. E4 DeepSeek-V3/README.md at main · deepseek-ai/DeepSeek-V3 GitHub (deepseek-ai) · 26 December 2024

    “| Claude-Sonnet-3.5-1022 | 85.2 | 52.0 | | DeepSeek-V3 | 85.5 | 70.0 |”

  5. E5 DeepSeek-V3/README.md at main · deepseek-ai/DeepSeek-V3 GitHub (deepseek-ai) · 26 December 2024

    “DeepSeek-V3 achieves the best performance on most benchmarks, especially on math and code tasks.”

  6. E6 Model Card Addendum: Claude 3.5 Haiku and Upgraded Claude 3.5 Sonnet Anthropic Own record

    “upgraded Claude 3.5 Sonnet solving 78% and Claude 3.5 Haiku solving 74% of problems, compared to 64% for Claude 3.5 Sonnet and 38% for Claude 3 Opus”

Reasoning

E1 and E2 are Anthropic's own sources. They report that Claude 3.5 Sonnet scored 49% on SWE-bench Verified, but they give no DeepSeek-V3 score to compare it with. E6 compares only Claude-family models. E4 is an unlabelled table row with no column headings in the quote, and E5 is a general statement by DeepSeek that does not name Claude. E3 explicitly names no coding winner. None of the evidence directly compares Claude 3.5 Sonnet with DeepSeek-V3 on real-world coding, so the claim cannot be resolved.

Cite this claim

Dario Amodei, “Anthropic finds Claude 3.5 Sonnet remains much better than DeepSeek-V3 on some key tasks such as real-world coding.” (said 13 October 2022). Verdict: unresolvable. Reliability API, claim c_958e4f582157, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_958e4f582157

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Dario Amodei, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Dario Amodei

  1. Mixed Factual 23 September 2026
    “It’s easy to dismiss this as a one-off or curiosity, but we’ve repeatedly seen a pattern where AI performance in new intellectual domains goes from weak to superhuman in a matter of a few years.”

    AI performance in new intellectual domains has repeatedly gone from weak to superhuman within a few years.

  2. Mixed Factual 10 March 2025
    “we were measuring these trends back when models cost $1,000 or $10,000 to train”

    Amodei's group measured AI scaling trends when models cost $1,000 or $10,000 to train.

  3. Mixed Factual 28 February 2025
    “a lot of the other reasoning models in the market are trained primarily on math and competition coding,”

    Many other companies' reasoning models on the market are trained primarily on math and competition coding.

  4. Mostly true Factual 28 February 2025
    “I do think AI coding is moving the fastest of all the other areas.”

    Amodei said he thinks AI coding is moving the fastest of all the areas of AI.

  5. Mostly true Factual 28 February 2025
    “Because when it knows that it doesn't get any gain out of thinking further, it doesn't think for longer.”

    The model sometimes gives a very short response because, when it knows it gains nothing from thinking further, it does not think for longer; a bound on how long it thinks is still described as valuable.

All of them

Verdicts are made automatically; how verdicts are made.