Claude Opus 5 still wins the naked exam. DeepSeek's cheap Flash snapshot just took the agent floor anyway - and handed free search keys to Claude Code and Codex until Halloween.
On paper it is a minor update. On October 17, 2026, DeepSeek shipped DeepSeek-V4.1-Flash-1017, thirty-seven days after V4.1-Flash. The API name did not change. You keep calling deepseek-flash. Pricing did not change. Existing traffic silently moved to the new checkpoint.
Then the scoreboard split in two.
77.0 on the loop. 42.6 on the exam.
DeepSeek's own table puts Flash-1017 at 77.0 on DeepSWE v1.1, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0, up from 74.2 on the prior Flash. On Agents' Last Exam it lands 35.4 against Opus 5's 28.6. Automation-Bench hits 60.5 against 50.3.
Flip the page to Humanity's Last Exam without tools and the story reverses. Opus 5 sits at 56.3. Flash-1017 sits at 42.6. On Terminal-Bench 4.0 run through the Claude Code harness, Opus still leads 51.8 to 42.3.
That is not a small delta. It is a product category.
The model that still loses naked hard reasoning is the one winning software-agent loops - the jobs where tools exist, retries are free, and shorter answers cut the bill. DeepSeek said as much in its changelog: faster task completion, shorter outputs, a performance-cost Pareto move across multiple harnesses. The weights sit on Hugging Face under MIT.
PC Watch framed the Japanese read cleanly: 552B-class MoE backbone, Engram memory on the side, MIT commercial weights the same day, DeepSWE above Opus 5. The official English post from @deepseek_ai crossed four million impressions in hours.
The social chart said 1/100. The rate card said something quieter.
Japanese X filled with a cost slide that made the week feel like another DeepSeek shock. One widely shared post claimed about $0.2 per coding-agent task for Flash-1017 against roughly $7.5 for GPT-6 Astra and $22 for Claude Fable 5 - Astra at 1/40, Fable at 1/100. That framing moved fast; treat the third-party task averages as social reaction, not a lab ledger.
What DeepSeek actually published is more structural. Token prices for deepseek-flash stay on the existing peak/off-peak card: off-peak cache-miss input at $0.15 per million tokens, output at $0.60, with peak hours at double those rates. The company claim is that the new checkpoint finishes work with fewer tokens. Cost falls because the loop got quieter, not because a new sticker appeared on the menu.
One careful Japanese thread said the quiet part out loud: the shock is less "smarter than Opus" and more "this cheap." That is the right hierarchy.
Upgrade the harness first
The announcement is not only a model drop. DeepSeek says Flash-1017 is optimized for DeepSeek Harness (v0.2.1-rc.1 or later), Tencent CodeBuddy, WorkBuddy, and more. The model and the loop software ship as a pair.
Then the distribution move: Search and Fetch APIs open to everyone at $0.75 per 1,000 successful calls, powered by DeepSeek's in-house engine. Through server-side tools in DeepSeek Harness, Claude Code, and Codex, those calls stay free for API users until October 31, 2026, 16:00 UTC.
Read that again. A Chinese open-weight lab is subsidizing web search inside Anthropic's and OpenAI's coding agents so that the default path stays warm on deepseek-flash. The MIT dump is the press release. The free search window inside rival harnesses is the install base play.
It rhymes with older platform wars: give away the commodity layer where the loop already lives, then own the meter on the model that keeps winning that loop.
What this is not
It is not "frontier models are dead." Opus still owns HLE without tools and several repo/program benches on DeepSeek's own chart. Community reproductions will matter more than vendor DeepSWE in the next week.
It is not a surprise that open weights under MIT still need serious silicon to self-host the full FP8 shard set. The practical path for most teams remains the API identifier that auto-routes.
It is also not a one-off miracle. V4.1-Flash landed September 10. The 1017 suffix is a five-week cadence on the cheap tier: keep the name, keep the price, raise the Pareto line, pull the harness forward. Flash is becoming the agent SKU. Flagship stays the exam SKU.
If you budget agents by flagship tokens because "coding needs the best model," this week is the audit. The loop may already prefer the model that lost the naked test.