loading
loading
Cursor Adds DeepSeek V4.1 Flash Hitting 47.7% for Under $1 per Task DeepSeek V4.1 Flash lands in Cursor with four effort tiers that sit on the CursorBench cost frontier from $0.60 to $0.98 per task. - DeepSeek V4.1 Flash is now selectable in Cursor across four effort tiers. - Costs range from $0.60 to $0.98 per CursorBench task, all on the cost frontier. - Max tier scores 47.7% on CursorBench 4.0, Low scores 44.9%. - Underlying model hits 74.2% on Deep-SWE v1.1, beating Opus-5 and GPT-5.6 Sol. - Features a 1M token context window and roughly 210 output tokens per second. - Enable via Cursor Settings then Models; no BYOK proxy needed. Cursor has added DeepSeek V4.1 Flash to its standard model picker, with four reasoning-effort settings and a specific benchmark claim: every setting lies on the CursorBench 4.0 cost frontier. For developers running long, multi-file agents, the model promises more attempts per budget if Cursor’s internal results transfer to their repositories. The model is available under Cursor Settings > Models. Developers can select Low, Medium, High, or Max effort for each agent run. Four settings put cost on display CursorBench 4.0 evaluates coding agents on ambiguous, multi-file tasks drawn from real Cursor sessions. Its cases cover editing, refactoring, investigation, intent interpretation, job management, and adherence to design constraints. The results reflect the model working inside Cursor’s agent harness, including its tools and prompts. The published figures place Medium 0.7 percentage points above High while using 13,000 more tokens and costing three cents more. Max adds 2.8 points over Low for another $0.38 per evaluated task. Average tokens represents benchmark consumption across an agent run; a single prompt can contain fewer tokens. A model reaches the cost frontier when no tested alternative achieves a higher score at the same or lower estimated cost. Cursor’s claim therefore concerns efficiency within its test set and agent configuration. Premium proprietary models still record higher absolute scores. Agent scores drive the pitch DeepSeek positions V4.1 Flash as a low-latency model for coding and tool use. In the company’s technical report, it receives a Codeforces rating of 3471, ahead of V4-Flash at 3289 and V4-Pro at 3348. Codeforces measures competitive-programming performance, including algorithm selection and implementation under contest conditions. Deep-SWE v1.1 maps more directly to repository work. DeepSeek reports a 74.2% pass rate for V4.1 Flash, compared with 54.4% for V4-Flash. Its comparison table lists Opus-5 at 74.0% and GPT-5.6 Sol at 73.0%. Pass@1 and single-pass rates measure whether the first generated attempt succeeds. - Terminal-Bench 2.1: 90.6 on tasks that require operating through a terminal. - GPQA Diamond: 90.9% on graduate-level science questions. - MathArena Apex: 65.6% Pass@1 on advanced mathematics problems. - Reported context window: 1 million tokens. - Reported output speed: about 210 tokens per second under DeepSeek’s test conditions. Vendor benchmarks use different harnesses, prompts, tool permissions, and scoring rules. Cross-model comparisons carry the most weight when the same evaluator runs every model under identical conditions, as CursorBench attempts to do for models available in Cursor. Cheap tokens, separate bills DeepSeek lists raw API rates of roughly $0.14 per million input tokens and $0.28 per million output tokens. Those rates are substantially below premium proprietary APIs and support high-volume uses such as repeated repository searches, test-fix loops, and parallel agent runs. Cursor’s $0.60 to $0.98 cost-per-task figures use benchmark accounting and may differ from a customer’s charge. The published summary does not reconcile those estimates with DeepSeek’s raw API rates, token mix, caching, agent overhead, Cursor plan allowances, or usage-based billing. Teams should use their Cursor dashboard and plan terms for budget forecasts. Choose by verification cost Repository-level evaluation should determine the default setting. Useful measures include accepted-diff rate, test-pass rate, review time, retries per completed task, latency, and total cost. A small suite of representative issues will reveal more about local performance than a leaderboard rank alone. What the leaderboard leaves open - Success remains uneven: Max reaches 47.7% on CursorBench, leaving substantial room for failed attempts. Production use still requires tests, code review, and reversible changes. - Long context has limits: A 1-million-token window describes capacity. Retrieval quality, attention across large inputs, latency, and Cursor’s own limits determine how much context is useful in practice. - Reasoning fields can affect custom integrations: DeepSeek thinking variants may return a reasoning_content field. Direct API clients and BYOK proxies should preserve or handle that field correctly. Cursor’s native model selection removes the need for a custom adapter. - Deployment details require review: The benchmark announcement does not establish data residency, retention, regional availability, or enterprise privacy terms. Teams handling sensitive code should confirm those policies before rollout. - The sources are vendor-run: Cursor produced the agent benchmark, while DeepSeek reported the underlying model results. Independent repository evaluations would provide stronger confirmation. Teams with high agent volume and reliable automated verification have the clearest economic case for V4.1 Flash. Low offers inexpensive retries, Medium and High cover longer routine runs, and Max keeps Cursor’s estimated benchmark cost below one dollar per task. A representative repository trial should precede any default-model change.
document fetched 2026-10-27 02:55