This Week’s Rankings

#1 Claude Duet (Anthropic) — Watch

Anthropic’s biggest announcement of the week is not a model update but a new paradigm: instead of waiting for you to open a chat, Duet sends you a morning brief, flags what needs attention in your inbox and calendar, takes tasks by message or voice, and completes them while you’re away — checking back before any sensitive action such as sending a message or moving money. It runs on Claude Opus 5.5 and connects to more than 6,000 plugins, including calendar, email, and document tools. The research preview opened October 21 to US Claude Pro and Max subscribers on a waitlist; no EU, Enterprise, or Team date has been announced. Anthropic announcement | MacRumors

Anthropic published no benchmark scores for Duet’s agentic behaviour. What the company did publish: Duet “may misread a thread or miscalibrate what’s urgent” and “can miss something that arrives between checking in.” Given competing launches from OpenAI (ChatGPT Dots) and Meta (Muse), it is clear both labs are racing to own the proactive-agent layer. The verdict is Watch — not because the product is uninteresting, but because no independent evaluation exists yet and access remains gated.

#2 DeepSeek V4.1-Flash-1017 (DeepSeek) — Use It

Released October 17 with MIT weights on Hugging Face and ModelScope, this snapshot raises the bar for open-weight coding agents. DeepSWE v1.1 reaches 77.0% (up from 74.2% in September), Terminal-Bench 4.0 jumps to 42.3% (from 31.2%), and all five benchmark rows DeepSeek published show gains with no regressions — though every figure is vendor-run. Managed inference provider Sference independently measured 20–30% fewer reasoning tokens per task at unchanged correctness, which is the most useful independent signal this week. AIWeekly | MetricNexus | Our full breakdown

One operational note before you integrate: the Search and Fetch endpoints begin billing at $0.75 per 1,000 successful requests on October 31 at 16:00 UTC. If you use those endpoints today, update your cost model now. The API model name stays deepseek-flash — no code change required for the snapshot upgrade.

#3 Mercury 3 (Inception Labs) — Test It

Inception Labs launched Mercury 3 on October 19 as a diffusion language model (dLLM) that generates tokens in parallel rather than one at a time. The lead claim is 1,068 tokens per second on NVIDIA GPUs — Inception does not name the hardware, batch size, or effort setting, so the figure is illustrative, not a reproducible specification. Artificial Analysis had no Mercury 3 page as of October 20. Inception’s own benchmark table puts Mercury 3 at 51.1% on SWE-Bench Atlas versus 33.3% for GPT-6 Luna and 32.3% for Claude Haiku 5.5, but Haiku 5.5 leads on DeepSWE (51.6% vs. 46.3% for Mercury 3). CellCog overview | OrcaRouter analysis

The Test It verdict reflects the genuinely attractive launch pricing ($0.10/$0.35 per MTok until further notice) and the architectural novelty. Upgrade to Use It when an independent lab confirms the speed and quality claims on your target task. Plan using list prices ($0.20/$0.70) for budget projections — the 50% discount has no stated end date.

#4 ChatGPT Browser Actions (OpenAI) — Test It

Not a model release, but a meaningful step in the agent-capability race: OpenAI’s ChatGPT can now use your open Chrome tabs as context, fill forms, click buttons, and download files — with no extra charge for Plus ($20/mo) or Pro ($200/mo) subscribers. The feature is closer to computer-use than a browser extension integration, and its timing alongside Claude Duet’s launch makes clear that both Anthropic and OpenAI are racing to own the agentic-workflow layer before it commoditises. If you are already on ChatGPT Plus, this is worth a test on a real-world task this week. IBTimes

#5 MiniMax M3.1 Flash (MiniMax) — Watch

MiniMax released M3.1 Flash on October 19 with a 1M-token input context, 524K-token output limit, and pricing identical to DeepSeek V4.1-Flash-1017 at $0.15/$0.60 per million tokens. The vendor claims 3× faster throughput and 70% lower costs versus the previous version but does not name the baseline. No independent benchmark scores were published at launch. At the same price with no benchmark edge to show, M3.1 Flash needs a compelling independent eval result before it earns production consideration. Come back when Artificial Analysis posts numbers. LLM Stats

Cost Math: 100 Code Reviews a Day

Assumptions: each review uses 4,000 input tokens and 2,000 output tokens. Daily volume: 400K input, 200K output. Prices are from provider listings cited above. DeepSeek uses off-peak rate; Mercury 3 shows both launch and list price. Opus 5.5 price is from OrcaRouter, not Anthropic’s pricing page.

Mercury 3 at launch price wins on token cost — but budget at list price ($0.22/day) until the discount timeline is clear. At list, DeepSeek is cheaper and carries the only independently verified token-efficiency improvement of the week. Opus 5.5 is 25× more expensive than DeepSeek at off-peak — Duet is not the right call for high-volume automated code review.

For Swiss & EU teams

DeepSeek V4.1-Flash-1017’s MIT weights are the most relevant development for FADP-conscious teams this week. Self-hosting the model on Swiss or EU infrastructure means personal data stays on your network, with no data-sharing terms to negotiate with DeepSeek’s API. Claude Duet is US-only in the research preview; no EU availability date has been announced. Mercury 3 does not yet publish data-processing terms for its hosted API — treat the endpoint as outside EU jurisdiction until confirmed. See our guide on securing your AI agent pipeline for FADP and NIS2 for a self-hosting checklist.

What to Watch Next Week

Claude Duet’s waitlist opening will test whether Anthropic can scale a production-quality proactive agent — and independent task-completion benchmarks will tell us whether the design matches the demo. Mercury 3 needs a third-party evaluation: if Artificial Analysis confirms the 1,068 tokens/s figure and competitive quality scores, diffusion LLMs become a serious architectural option for latency-sensitive pipelines. DeepSeek’s Search and Fetch API billing starts October 31 at 16:00 UTC — $0.75 per 1,000 requests; review your call volume this week.

Further Reading

• DeepSeek V4.1-Flash-1017: MIT Weights and 77% on DeepSWE — our full breakdown including the Search API billing change and self-hosting notes
• AI Coding Agents in 2026: Benchmarks, Costs, and How to Choose — the pillar for any coding-agent procurement decision; covers 7 agents on Terminal-Bench 4.0 with cost-per-task math
• DeepSeek V4.1-Flash-1017: Benchmarks and Pricing (MetricNexus) — includes Sference’s independent token-efficiency finding and the context/output limits

Your turn: Is your team testing any of the proactive-agent platforms — Claude Duet, ChatGPT Dots, or Meta Muse — or are you waiting for independent benchmarks first? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.