Claude Opus 4.8 vs DeepSeek V4 Pro (0813)

Claude Opus 4.8 vs DeepSeek V4 Pro (0813)

Independent runs have rewritten this matchup, mostly into ties. Claude Opus 4.8's one clean, confirmed win is raw reasoning without tools (HLE 48.7 vs 41.0, both Artificial Analysis runs). DeepSeek V4 Pro posts a bigger SWE-bench Verified number (96.4 vs 88.6, both vals.ai) — but that benchmark is saturated (grade D here), and by our own methodology rankings on it are unreliable regardless of gap size. LiveCodeBench (87.82 vs 87.53) and GPQA Diamond (92.42 apiece) no longer count as ties either: both are now graded saturated (grade D, same reason as SWE-bench Verified above; LiveCodeBench joined on 2026-09-29 after vals.ai stopped running it) — near-identical scores there say the test has run out of headroom, not that the two models are evenly matched. Terminal-Bench looks close too (78.9 vs 78.7) but the numbers come from different evaluation harnesses (tbench.ai's official board vs Artificial Analysis's own runs) — not a fair comparison even though both sides are independently run. vals.ai, which ran both on one Terminus 2 harness, disagrees with "close": 71.91 for Opus 4.8 against 54.68 for DeepSeek V4 Pro 0813, a 17.2-point gap past the 10.6 band, though neither figure is this site's tracked row. DeepSeek's own numbers for multi-tool chores and professional tasks used to sit close behind Opus's independently-run scores on those two (74.1 vs 76.2, 25.7 vs 27.0) — but as of 2026-10-03 the multi-tool one is measured, not claimed: Toolathlon's own board run puts DeepSeek V4 Pro 0813 at 74.4 against Opus's 76.2 on the same harness, a 1.8-point gap inside the 9.7 band, so that line is now a tie both sides independently ran. Long-horizon coding (62.7 vs 59.0) and professional tasks (25.7 vs 27.0) are still DeepSeek's own numbers against Opus's independent ones — since only Opus's side of those two has been independently checked, call both unverified rather than tied or real — DeepSeek hasn't confirmed it's keeping pace, or ahead, just claimed it. Three more newly-tracked benchmarks don't change the picture: LiveBench is another tie (76.2 vs 77.4, both independently run on the same harness); Terminal-Bench 4.0, added 2026-10-08, is another tie (23.23 vs 14.14, both vals.ai mini-swe-agent runs at max effort, the 9.1-point gap sitting inside that column's 12.4-point band); and ARC-AGI-2 doesn't add a comparable line — the two models' best publicly reported scores come from different reasoning-effort tiers (Opus 4.8's High-tier 72.1 vs DeepSeek's Max-tier 61.3), so there's no matched-tier pair to call. OSWorld 2.0, added the same day, is tracked for Opus 4.8 only and adds no line here. The launch chart's drama mostly dissolves under independent measurement; what survives is a price gap: at list rates DeepSeek costs about a sixth of Opus per output token ($3.96 vs $25.00).

Claude Opus 4.8 vs DeepSeek V4 Pro (0813): benchmark by benchmark

Across 16 rows comparing Claude Opus 4.8 and DeepSeek V4 Pro (0813): 1 row carries a confirmed real gap, 3 rows land inside the noise band, and 12 rows carry other caveats — unverified, tainted, or run on mismatched tool setups. Read each row’s Signal label before ranking Claude Opus 4.8 against DeepSeek V4 Pro (0813).

Who should pick Claude Opus 4.8, and who should pick DeepSeek V4 Pro (0813)

  • High-volume or budget-constrained workloads → DeepSeek V4 Pro (0813) (measured parity at roughly a sixth of the output-token price — the ties are the story: you rarely give up confirmed capability for the discount)
  • Single-shot reasoning without tool access → Claude Opus 4.8 (the matchup's only clean independently-confirmed real gap: HLE no-tools 48.7 vs 41.0 (Artificial Analysis ran both))
  • Repository-scale bug fixing → DeepSeek V4 Pro (0813) (96.4 vs 88.6 on vals.ai's identical harness — but flagged honestly: SWE-bench Verified is saturated (grade D), so treat this as weak evidence, and the price still favors DeepSeek anyway)

Claude Opus 4.8 vs DeepSeek V4 Pro (0813) pricing

Input / 1M tokens$5.00$1.32
Output / 1M tokens$25.00$3.96

Claude Opus 4.8 vs DeepSeek V4 Pro (0813) FAQ

Is Claude Opus 4.8 better than DeepSeek V4 Pro (0813)?

Independent runs have rewritten this matchup, mostly into ties. Claude Opus 4.8's one clean, confirmed win is raw reasoning without tools (HLE 48.7 vs 41.0, both Artificial Analysis runs). DeepSeek V4 Pro posts a bigger SWE-bench Verified number (96.4 vs 88.6, both vals.ai) — but that benchmark is saturated (grade D here), and by our own methodology rankings on it are unreliable regardless of gap size. LiveCodeBench (87.82 vs 87.53) and GPQA Diamond (92.42 apiece) no longer count as ties either: both are now graded saturated (grade D, same reason as SWE-bench Verified above; LiveCodeBench joined on 2026-09-29 after vals.ai stopped running it) — near-identical scores there say the test has run out of headroom, not that the two models are evenly matched. Terminal-Bench looks close too (78.9 vs 78.7) but the numbers come from different evaluation harnesses (tbench.ai's official board vs Artificial Analysis's own runs) — not a fair comparison even though both sides are independently run. vals.ai, which ran both on one Terminus 2 harness, disagrees with "close": 71.91 for Opus 4.8 against 54.68 for DeepSeek V4 Pro 0813, a 17.2-point gap past the 10.6 band, though neither figure is this site's tracked row. DeepSeek's own numbers for multi-tool chores and professional tasks used to sit close behind Opus's independently-run scores on those two (74.1 vs 76.2, 25.7 vs 27.0) — but as of 2026-10-03 the multi-tool one is measured, not claimed: Toolathlon's own board run puts DeepSeek V4 Pro 0813 at 74.4 against Opus's 76.2 on the same harness, a 1.8-point gap inside the 9.7 band, so that line is now a tie both sides independently ran. Long-horizon coding (62.7 vs 59.0) and professional tasks (25.7 vs 27.0) are still DeepSeek's own numbers against Opus's independent ones — since only Opus's side of those two has been independently checked, call both unverified rather than tied or real — DeepSeek hasn't confirmed it's keeping pace, or ahead, just claimed it. Three more newly-tracked benchmarks don't change the picture: LiveBench is another tie (76.2 vs 77.4, both independently run on the same harness); Terminal-Bench 4.0, added 2026-10-08, is another tie (23.23 vs 14.14, both vals.ai mini-swe-agent runs at max effort, the 9.1-point gap sitting inside that column's 12.4-point band); and ARC-AGI-2 doesn't add a comparable line — the two models' best publicly reported scores come from different reasoning-effort tiers (Opus 4.8's High-tier 72.1 vs DeepSeek's Max-tier 61.3), so there's no matched-tier pair to call. OSWorld 2.0, added the same day, is tracked for Opus 4.8 only and adds no line here. The launch chart's drama mostly dissolves under independent measurement; what survives is a price gap: at list rates DeepSeek costs about a sixth of Opus per output token ($3.96 vs $25.00).

Which is cheaper, Claude Opus 4.8 or DeepSeek V4 Pro (0813)?

DeepSeek V4 Pro (0813) costs less per output token ($3.96 vs $25.00 per 1M tokens, official listed rates).

Is either of these models open weight?

DeepSeek V4 Pro publishes open weights (MIT license); Claude Opus 4.8 is proprietary. The benchmark record between them is mostly ties — Opus 4.8's one confirmed win is HLE without tools (48.7 vs 41.0, both Artificial Analysis runs), while DeepSeek V4 Pro posts the larger SWE-bench Verified number (96.4 vs 88.6, both vals.ai) on a board this site grades saturated.

Which is better at coding?

The independent coding evidence calls this matchup a tie. DeepSeek V4 Pro scores higher on SWE-bench Verified (96.4 vs 88.6, both vals.ai runs), but that benchmark is saturated here (trust grade D), and LiveCodeBench reads 87.82 vs 87.53 — a statistical tie on a board this site no longer ranks.

Is there a newer version of either model?

Claude Opus 4.8 is computed as a previous version on this site, with Claude Opus 5 as its successor. DeepSeek V4 Pro has no computed successor: DeepSeek's V4.1 Flash replaced the smaller V4 Flash, not the Pro tier.

How big does a score gap need to be to count here?

More than the benchmark's meaningful gap. On HLE without tools, 48.7 vs 41.0 clears the band and reads as a real gap. On GPQA Diamond (92.42 apiece) and LiveCodeBench (87.82 vs 87.53), no gap counts at all: both boards are graded saturated, so every pair on them is labeled Tainted regardless of the arithmetic.