Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

DeepSeek's cheap coding model changed under the same name: retest Flash 1017 before you rely on it

Written October 31, 2026 · Edited November 11, 2026

Anyone calling deepseek-flash has been running different weights since 17 October 2026, without changing anything. The new checkpoint costs the same and, on DeepSeek's own figures, scores higher — but the September checkpoint is gone, so the only way to know what changed for your work is to retest your own tasks.

What changed

On 17 October 2026, DeepSeek's API changelog recorded the swap in one paragraph: DeepSeek V4.1 Flash had been upgraded to DeepSeek-V4.1-Flash-1017, the web and app models had been updated too, and the API calling method and pricing remained unchanged — simply use deepseek-flash to access the latest model. There was no migration step, no new model name to adopt, and no version string to pin. DeepSeek's pricing page confirms the position: deepseek-flash is served by DeepSeek-V4.1-Flash-1017, and the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but their models have been retired, with requests served by the 1017 checkpoint and billed at the Flash price. DeepSeek lists no model name for the September checkpoint, so there is no supported way to keep calling it.

The model card calls 1017 a minor update to V4.1 Flash, with gains across reasoning, coding, agentic, and visual benchmarks, and says it reduces task execution time and output length. The price did not move: $0.30 per million input tokens and $1.20 per million output tokens at peak, half that off-peak, with cache-hit input at $0.006 peak. The context window stays at one million tokens, with up to 384,000 output tokens. DeepSeek recommends DeepSeek Harness 0.2.1-rc.1 or later for the best experience with the new checkpoint.

What got better, on DeepSeek's own figures

Every figure in this section is DeepSeek's own, from its changelog and model card, measured on test setups DeepSeek chose. Terminal-Bench 4.0 rises from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, ProgramBench from 20.3 to 29.1, and Humanity's Last Exam without tools from 36.8 to 42.6. The changelog notes the test setup for two of them: Terminal-Bench 4.0 was tested with the Claude Code framework, and DeepSWE with the Mini-SWE-V2 framework. ZeroBench-main was evaluated with Pass@5, a more forgiving scoring rule than a single attempt, which matters when comparing its 51.0 with scores measured differently.

The same table keeps the gains in proportion. Claude Opus 5 still leads most rows in DeepSeek's own comparison: Humanity's Last Exam at 56.3 against 42.6, Terminal-Bench 4.0 at 51.8 against 42.3, and ProgramBench at 37.0 against 29.1. The 1017 checkpoint leads on DeepSWE v1.1, at 77.0 against Opus 5 at 74.0. The honest reading is narrow: on DeepSeek's chosen tests, the new checkpoint is better than its predecessor and competitive with frontier models on agent-style work, while remaining behind them on knowledge-heavy reasoning. Nothing in the table says the model got better at your task.

What did not change, and what it cannot do

The swap changed weights, not the shape of the model. The API name, the calling method, the price, the context window, and the output limit are all as they were. That stability is also the limit: nothing about the swap gives you a way back. If the new checkpoint handles one of your tasks worse than the September checkpoint did, you cannot pin the old one, and the fix is to change the prompt, the setup, or the model — not to wait for the old weights to return.

The vendor figures also carry a caveat that the table itself states. DeepSeek's card notes that the same September model scored 74.2 on DeepSWE inside one test setup and 69.8 inside Claude Code — a spread of more than four points from scaffolding alone, on identical weights. Test setup moves these scores. That is why a benchmark gain, even an eleven-point one, is a reason to retest, not a promise about your setup.

Retest your own tasks before trusting it

  1. Pick three to five tasks you already know the answer to. Real tasks from your own work, with known-good outputs: a bug fix with a test, a small feature with an acceptance check, a refactor where behavior must not change. A task without a known answer cannot tell you whether the model changed.
  2. Run them exactly as before. Same prompts, same tooling, same settings. Change nothing else, or the comparison measures your edits instead of the model.
  3. Record three things per task. Whether the result was right, how long it took, and how much output it produced. DeepSeek says the new checkpoint is faster with shorter outputs; your own log is the only place that claim becomes a fact about your work.
  4. Decide per task type, not per model. A task type that passes unchanged keeps running. A task type that regresses gets a new prompt, a different setting, or a different model. One clean task type says nothing about the next one.
  5. Keep the task set. The next silent swap will arrive the same way this one did — under the same name, with no action required. A kept set of known tasks turns that event from a surprise into a ten-minute check.

Limits of this evaluation

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Budget DeepSeek Search and Fetch calls · DeepSeek Search and Fetch projector · Pick the cheapest model for the job