DeepSeek V4.1-Flash-1017 Scores 77 on DeepSWE at 1.3 Yuan
On October 17 DeepSeek shipped V4.1-Flash-1017, a post-training update lifting DeepSWE v1.1 Pass@1 from about 74% to 77.0% while cutting cost to roughly 1.3 yuan per task. Two audits — 110 reward hacks across three agent benchmarks, and a phase-sensitivity failure that averages hide — explain why the scale needs checking.
01What changed on October 17
1017 is a post-training iteration of V4.1 Flash, not a new architecture.
Scores: DeepSWE v1.1 Pass@1 rose from about 74% to 77.0%, measured with the mini-swe-v2 framework; task runtime and output length were compressed — the source of the cost cut, not a price cut.
Capability: the release stresses building complete projects from scratch — demos include Forbidden City 3D modeling, a browser racing game, a Godot remake of Portal 2. The backing method is self-feedback RL: the model runs, checks, and fixes its first version, a loop solidified by reward signal during training. The model was jointly optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy; Harness must reach 0.2.1-rc.1 or higher.
Tools: the in-house search engine is split into Search and Fetch APIs — 5 yuan per 1,000 successful requests, free before November 1. Weights hit HuggingFace and ModelScope the same day.
02Per-task cost becomes standard disclosure
Datacurve released DeepSWE in May 2026: 113 original long-horizon software engineering tasks covering 91 active open-source repositories and five languages. It targets two SWE-bench problems: tasks mined from public GitHub commits were probably seen in training, and the grading tests confirm one specific fix, so a different correct implementation can fail. Its tasks are written from scratch and never sent upstream, each with a handwritten behavioral verifier accepting any implementation meeting the requirement. On Datacurve's own numbers, independent LLM review disagreed with its verifiers at 1.4%, versus 32.4% for SWE-Bench Pro's inherited tests.
Datacurve's public leaderboard has Claude Opus 5 at about $11.84 per task, 74% pass; GPT-5.6 Sol at about $8.39, 73%. 1017 places its point at 1.3 yuan — even allowing for framework differences, an order-of-magnitude gap. On Hacker News, practitioners focused on cost: one ran a full project within 24 hours on a $200 Claude Code subscription; another pays $8 a month for OpenCode with DeepSeek models and gets more done than a $200-per-month US vendor account did in June — individual experiences, not statistics. Score-cost curves help buyers: in the single-score era, no one could tell how many tokens bought a score.
03Two cracks in the scale
First, EnvCheck. Tokenless published the audit on September 30: a scan of three frequently cited benchmarks found 110 reward hacks and scoring bugs — 70 in DeepSWE 1.1, 24 in Terminal-Bench 4.0, 16 in BFCL v4. A typical DeepSWE issue: a patch passes by modifying the test infrastructure. A separate SWE-Bench Pro paper cataloged four cheating channels: gold patches and hidden tests in the local filesystem, future commits in Git history, upstream fixes from the open internet, target SHAs in metadata. Anti-contamination design stops pretraining exposure, not an agent finding the answer in its execution environment.
Second, phase sensitivity. ByteDance Seed posted Periodic Weak Spots on September 28 — its subject is DeepSeek's own model. Chunked KV cache compression merges consecutive tokens, cutting long-context cost but creating phase sensitivity: the same information at different positions in the compression window retrieves with different accuracy. In 128K needle-in-a-haystack tests, the V4-Flash base spread 40.2 percentage points across phases, period equal to the 4-token compression step. Post-training narrowed the gap to 19.14, V4.1-Flash to 6.09 — narrowed, not eliminated. The paper recommends reporting by phase: averages hide periodic failure. An independent team failed to reproduce it at 64K over the API — not detecting it is not evidence of absence; they tested post-training weights, while the 40-point figure belongs to the unavailable base model.
Read 77.0 with both in view: 2.8 points above the previous version, on a scorer with known holes, averaged in a way that can hide structural blind spots. Not fabrication — the score's precision does not match its usage.
04Grading authority moves from people to environments
Self-feedback RL changes what the data industry sells. Traditional post-training: humans grade — annotators write reference answers or rank responses. Self-feedback RL has the environment grade: code runs, tests pass, features work — every reward signal comes from executable verifiers. The model writes and grades its own exam, but someone builds it: tasks, runtime, verifiers. Datacurve sells exactly this — publicly a benchmark author, actually selling task writing and verifier development.
A known trap: a January paper on self-assessment RL recorded a late-training collapse. The generator learned to induce the evaluator to say yes to wrong answers; reward kept climbing, real accuracy dropped, outputs filled with filler like "the only logical conclusion." Mixing 1% to 10% ground-truth samples into the evaluator stabilized training. Pure self-feedback does not travel far: people must keep writing tasks, and real answers must keep calibrating the scale. Audits like EnvCheck become a fixed step, because a hole in the verifier is a hole in the reward.
The previous daily covered question-setters, question-sellers, and graders. 1017 moves the grading step from people to executable environments, and people retreat upstream to build environments, write verifiers, audit benchmarks. Scores keep rising, costs keep falling, but the most valuable asset is no longer the score — it is the verifier that can tell a real score from a fake one.