DeepSeek's MIT-licensed update beats Opus 5 on four agent benchmarks, as Yelp, Delta and two US banks block AI agents
DeepSeek's free-weight update lands above Claude Opus 5 on four agent benchmarks, while Yelp, Delta, JPMorgan and Charles Schwab move to keep AI agents off their sites.
This brief covers the 24 hours from 08:15 on October 17 to 08:15 on October 18, Dhaka time. The day's frontier-scale release came from DeepSeek, as weights anyone can download, and its own table puts them beside Claude Opus 5.
DeepSeek's Flash update, and what its own table says
DeepSeek-V4.1-Flash-1017 went onto Hugging Face under the MIT licence on October 17, weights included 1. It is live in DeepSeek's app and on its developer platform 2. DeepSeek calls it a minor update to V4.1-Flash, and says the update brings stronger benchmark performance, faster task completion and shorter outputs.
The 1017 in the name is the date. The model takes images as well as text, works through a 1M-token context, and is a sparse mixture of experts, meaning only a slice of its weights runs for any given token. The repository lists 763 billion parameters across its weight files 1.
The MIT licence carries no revenue cap, no registration step and no limit on commercial use. Any competitor can run these weights, fine-tune them, or sell access to them.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|
| HLE, no tools | 42.6 (44.0†) | 36.8 (39.1†) | 43.5 | 42.0† | 44.5 | 56.3 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 17.5 | 19.0 | 23.0 | 37.0 |
| NL2Repo-Bench | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| Agents' Last Exam | 35.4 | 31.8 | 27.6 | 28.5 | 26.7 | 28.6 |
| Apex Bench v1.1 | 72.1 | 64.6 | 47.8 | 56.6 | 54.3 | 65.8 |
| Chartography, with tools | 83.0 | 78.9 | 68.1 | — | 79.9 | 84.0 |
| BabyVision, with tools | 92.5 | 89.6 | 85.7 | — | 88.9 | 94.1 |
| ZeroBench-main, with tools | 51.0 | 49.0 | 41.0 | — | 53.0 | 52.0 |
These are DeepSeek's own figures, published on the model card and unverified by anyone else 1. The † marks the text-only subset of HLE, Humanity's Last Exam, a set of thousands of expert-written questions across maths, science and the humanities. DeepSeek ran Terminal-Bench 4.0 inside Anthropic's Claude Code harness and DeepSWE v1.1 in Mini-SWE-V2, and says so in a footnote.
Read row by row, the table is uneven. The new model is below Opus 5 on recall-heavy questions and on repairing existing code, with 42.6 against 56.3 on HLE and 29.1 against 37.0 on ProgramBench. It is above Opus 5 on four benchmarks that reward acting inside another system: DeepSWE, Automation-Bench, Agents' Last Exam and Apex.
Independent scores had yet to appear when this brief was written. Artificial Analysis, which publishes an assessment on the day of most frontier launches, had posted nothing about this release as of Sunday morning; its most recent model assessment was Reflection's Beam on October 16 3.
DeepSeek's own service is the other open question. The company suffered an outage of roughly five hours earlier in the week, and slowdowns have grown more frequent as developers pile onto its cheap models 2. Free weights cost nothing to download and a great deal to serve.
The doors on the other side
In the same window, The Verge's Emma Roth catalogued the sites that have stopped letting agents in 4. Users of Meta's Muse, xAI's Grok, OpenAI's Dots and the startup Instinct reported blocks on services that had worked weeks earlier.
The named refusals run from the mundane to the financial. Yelp told TechCrunch it admits only agents that pay for its data-licensing programme. Delta Air Lines stops agents booking tickets on a customer's behalf. JPMorgan Chase and Charles Schwab both confirmed they keep agents out of accounts.
Charles Schwab chief executive Rick Wurster gave the reason as liability, in an interview with Bloomberg 5.
We don't want someone just giving their password to Muse, and letting Muse come in and take action on that client's behalf. We can't control it, the client can't control it, something could go wrong and who's liable? For the protection of the client and for the protection of us, we're not allowing it.
Amazon is the furthest along, and the only one litigating. It barred Perplexity's shopping agent, was ordered by a judge to keep it out, then saw that order overturned on appeal in August, and has since filed an amended complaint. Reports now put Grok, Instinct and Dots on the same list. Elon Musk's response, on X, was that he "should start a competitor to Amazon."
The push runs the other way in places. Meta has signed retailers including Walmart, Gap, Sephora and Best Buy to let Muse shop, and Shopify has opened its catalogue to agents. Meta's argument is that "turning away a personal agent means turning away the customer behind it."
Then there is the quieter half, where nobody decided anything. Anti-bot defences built for scrapers also catch agents, because an agent browsing a checkout looks like a bot browsing a checkout. Chris Canal, chief executive of the safety evaluator EquiStamp, told The Verge that agents trip those same safeguards, and that the likely result is more prove-you-are-human prompts on ordinary shopping sites.
Canal also put the commercial question directly: "Does this impact the perceived value of the Anthropics and the OpenAIs? I would say absolutely."
The four benchmarks DeepSeek leads Opus 5 on are the ones that reward acting inside someone else's system. The systems in question are the ones now refusing that access.
Stored keys, and a bill someone else pays
OpenCode, an open-source coding agent built by Anomaly Innovations, emailed affected users on October 17 6. The company said a large-scale distillation attack had reached stored OpenAI, Anthropic and Google API keys belonging to 1.4% of OpenCode Console users, and that those keys were used to run up charges 7.
A distillation attack is the term for querying a model at volume to harvest its outputs as training data for a rival. It needs credentials, and a coding agent that stores your provider keys is a convenient place to find them.
OpenCode told users to revoke the keys and change their passwords, and said it is working with Google, OpenAI and Anthropic on the cleanup. It declined to give further technical detail while the incident is open, so the count of affected accounts, the size of the bills and the route the attacker took are all still undisclosed.
The exposure sits with the customer, which is the part that generalises. The charge lands on whichever account the key belongs to, and the agent tool in the middle is where the key was kept.
Meta takes the sentence out
Meta removed the disputed sentence from its commercial terms on October 17, after legal experts read it as permission to scrape publisher sites for training data 8. The deleted words said Meta may use any content from a business account's domains, web pages or URLs to "provide, develop, improve, research, train, and maintain the integrity of" its products.
Nieman Lab and Cornell law professor James Grimmelmann read that as an attempt to give Meta free training data and to override a site's robots.txt file 9. Meta vice-president of communications Andy Stone had called any suggestion of a change "false."
The company's new statement goes further than the denial. "The crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves," spokesperson Daniel Roberts told The Verge. "Content owners can use robots.txt to signal their preferences."
The terms still take effect on October 30 10. The sentence is gone, and the rest of the consolidated terms stay.
What to watch
An independent run of DeepSeek's table, and specifically whether the four agentic wins survive a neutral harness rather than Claude Code and Mini-SWE-V2.
Whether the quiet blocks become published policy. Right now Yelp and Delta state theirs, Amazon litigates its own, and Cars.com and Amtrak have said nothing at all.
Whether OpenCode publishes an incident report with numbers, and whether other agent tools that store provider keys tighten how they hold them.
Meta's final commercial terms before they take effect on October 30.
References
- 1deepseek-ai/DeepSeek-V4.1-Flash-1017
huggingface.co
- 2DeepSeek pushes out 'faster, leaner' updated AI model
straitstimes.com
- 3Artificial Analysis articles index
artificialanalysis.ai
- 4Websites are closing their doors to AI agents
theverge.com
- 5
- 6
- 7OpenCode breach exposes stored OpenAI and Anthropic keys
startupfortune.com
- 8
- 9
- 10
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.