大模型网摘V4.1-Flash-1017 One Week On: Reading the Official Benchmarks, and Five Open Questions
DeepSeek-V4.1-Flash-1017 shipped on October 17, and as of press time it has been one week. The official benchmark numbers look strong, but independent third-party evaluations have not appeared yet. This review answers two questions: what the official figures actually say, and what still needs independent verification.
The numbers
On the official DeepSWE v1.1 benchmark (mini-swe-v2 test harness), the new model scores Pass@1 77.0% at an average cost of about $0.20 per task; the low-usage tier hits 72% at roughly $0.10. The previous V4.1-Flash scored 74.2% at about $0.50. In the official cost chart, GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 land in the $2–10+ range. Full release notes in our launch roundup.
How to read them
First, within-framework gains are more trustworthy than cross-framework ones. The 77% vs 74.2% comparison is same-harness, generation over generation, which carries weight; cross-vendor comparisons need an apples-to-apples check. Second, "$0.20 per task" is an average that includes retries — it is not a per-call API price, and dividing it by your call count will misprice your bill. Third, "stronger, faster, shorter outputs" is the official Pareto framing — shorter outputs change the token mix, affecting both cache hit rates and real invoices, an underrated variable.
Five open questions
1. How large is the gap between mini-swe-v2 scores and real repository tasks? 2. How stable are long-horizon agent runs — after the near-five-hour outage on October 15, the community's sensitivity to stability is high. 3. How does native vision input perform on real documents such as scans and charts? 4. Does RLSF training introduce a perceptible style bias, for instance over-compression that drops context? 5. When will third-party replications arrive — official scores are necessary, not sufficient. For the mechanics of the training route, see the RLSF technical deep dive.
Verdict
Rather than waiting for a leaderboard ranking, run a small-sample A/B on your own workload: the same batch of real tasks through old and new models, comparing completion rate, retries and amortized cost. For the previous generation's longer-term record, see the V4.1-Flash five-week review; for cost analysis against GPT-Sol, Claude Opus 5, GLM-5.3 and Kimi K3, see the agentic coding cost comparison.